18.bcc

{"font_size":0.4,"font_color":"#FFFFFF","background_alpha":0.5,"background_color":"#9C27B0","Stroke":"none","body":[{"from":5.11,"to":10.29,"location":2,"content":"okay hi everyone let's get"},{"from":10.29,"to":14.62,"location":2,"content":"okay so so for today's lecture what"},{"from":14.62,"to":17.26,"location":2,"content":"we're gonna do is look at the topic of"},{"from":17.26,"to":19.93,"location":2,"content":"having tree recursive neural networks I"},{"from":19.93,"to":23.29,"location":2,"content":"mean this is actually a topic which I"},{"from":23.29,"to":26.05,"location":2,"content":"feel especially fond of and attached to"},{"from":26.05,"to":28.84,"location":2,"content":"because actually when we started doing"},{"from":28.84,"to":31.41,"location":2,"content":"deep learning friend or P at Stanford in"},{"from":31.41,"to":34.65,"location":2,"content":"2010 really for the sort of period from"},{"from":34.65,"to":39.43,"location":2,"content":"2010 to 2015 the dominant set of ideas"},{"from":39.43,"to":41.8,"location":2,"content":"that we were working on was this topic"},{"from":41.8,"to":44.56,"location":2,"content":"of how you could build a recursive tree"},{"from":44.56,"to":47.35,"location":2,"content":"structure into neural networks so in a"},{"from":47.35,"to":49.09,"location":2,"content":"way it's kind of funny that I'm only"},{"from":49.09,"to":51.73,"location":2,"content":"getting to it now I mean there are sort"},{"from":51.73,"to":54.34,"location":2,"content":"of reasons for that that I think there"},{"from":54.34,"to":56.98,"location":2,"content":"are a bunch of interesting ideas here"},{"from":56.98,"to":59.53,"location":2,"content":"which relate closely to linguistic"},{"from":59.53,"to":61.87,"location":2,"content":"structure and so it's good stuff to have"},{"from":61.87,"to":65.77,"location":2,"content":"seen but in practice these ideas have"},{"from":65.77,"to":68.53,"location":2,"content":"proven kind of hard to scale and not"},{"from":68.53,"to":71.14,"location":2,"content":"necessarily to work better in practice"},{"from":71.14,"to":73.9,"location":2,"content":"than the kind of things that we've spent"},{"from":73.9,"to":76.75,"location":2,"content":"more time on meaning things like looking"},{"from":76.75,"to":78.97,"location":2,"content":"at LST M's and looking at transformers"},{"from":78.97,"to":81.49,"location":2,"content":"and things like that and so that's kind"},{"from":81.49,"to":84.1,"location":2,"content":"of why we sort of shunted them towards"},{"from":84.1,"to":86.41,"location":2,"content":"the end of the curriculum but I want to"},{"from":86.41,"to":88.12,"location":2,"content":"sort of say something about the"},{"from":88.12,"to":90.28,"location":2,"content":"motivations and the way you can build"},{"from":90.28,"to":92.11,"location":2,"content":"tree structures and neural networks and"},{"from":92.11,"to":94.6,"location":2,"content":"look at some of the possibilities we"},{"from":94.6,"to":99.31,"location":2,"content":"explored in during this class another"},{"from":99.31,"to":101.77,"location":2,"content":"fact about this class is actually this"},{"from":101.77,"to":105.1,"location":2,"content":"is the last class I'm going to give so"},{"from":105.1,"to":107.14,"location":2,"content":"two more classes in the next week don't"},{"from":107.14,"to":110.47,"location":2,"content":"forget about next week's cs2 24 in"},{"from":110.47,"to":115.18,"location":2,"content":"classes but on Tuesday we've gone in the"},{"from":115.18,"to":117.16,"location":2,"content":"final invited speaker make Mitchell"},{"from":117.16,"to":119.41,"location":2,"content":"who's a great speaker and has tons of"},{"from":119.41,"to":122.08,"location":2,"content":"interesting stuff to say about fairness"},{"from":122.08,"to":125.53,"location":2,"content":"and ethics in NOP nai and then for the"},{"from":125.53,"to":127.87,"location":2,"content":"final lectures one another my PhD"},{"from":127.87,"to":130,"location":2,"content":"students Kevin Clark is going to give"},{"from":130,"to":132.91,"location":2,"content":"that and talk about some of the recent"},{"from":132.91,"to":134.62,"location":2,"content":"what's been happening in deep learning"},{"from":134.62,"to":138.07,"location":2,"content":"in 2018-19 or some of the sort of recent"},{"from":138.07,"to":141.19,"location":2,"content":"developments in NLP and deep learning so"},{"from":141.19,"to":142.58,"location":2,"content":"as"},{"from":142.58,"to":144.29,"location":2,"content":"I'll say my farewells at the end of this"},{"from":144.29,"to":148.03,"location":2,"content":"one so hopefully everyone has submitted"},{"from":148.03,"to":152.78,"location":2,"content":"their a milestone for their final"},{"from":152.78,"to":154.52,"location":2,"content":"project if you haven't you should really"},{"from":154.52,"to":157.67,"location":2,"content":"be getting your milestone in you know"},{"from":157.67,"to":159.77,"location":2,"content":"it's inevitable that somewhere around"},{"from":159.77,"to":163.55,"location":2,"content":"here there start to be problems that"},{"from":163.55,"to":165.86,"location":2,"content":"people have a situation that nothing"},{"from":165.86,"to":167.81,"location":2,"content":"works and everything is too slow and you"},{"from":167.81,"to":173.09,"location":2,"content":"panic and this happens I wish you luck"},{"from":173.09,"to":175.31,"location":2,"content":"of course I mean what can you do about"},{"from":175.31,"to":178.01,"location":2,"content":"it I mean it can be really hard when you"},{"from":178.01,"to":180.29,"location":2,"content":"have things that don't work as to work"},{"from":180.29,"to":183.17,"location":2,"content":"out why they don't work and how to fix"},{"from":183.17,"to":186.26,"location":2,"content":"them I mean I think often the best thing"},{"from":186.26,"to":188.63,"location":2,"content":"to do is really to go back to something"},{"from":188.63,"to":191.81,"location":2,"content":"simple that you can get working and to"},{"from":191.81,"to":195.11,"location":2,"content":"work forward from there again it also"},{"from":195.11,"to":198.41,"location":2,"content":"really helps to have really small data"},{"from":198.41,"to":201.2,"location":2,"content":"sets I really recommend the strategy of"},{"from":201.2,"to":204.53,"location":2,"content":"sort of having a ten item or 20 item"},{"from":204.53,"to":207.11,"location":2,"content":"data set and checking that your model"},{"from":207.11,"to":209.45,"location":2,"content":"works perfectly over trains to 100%"},{"from":209.45,"to":212.12,"location":2,"content":"accuracy on that kind of data set saves"},{"from":212.12,"to":215.09,"location":2,"content":"you huge amounts of time and it's sort"},{"from":215.09,"to":217.07,"location":2,"content":"of after you've gone something simple"},{"from":217.07,"to":219.35,"location":2,"content":"working on a small amount of data that's"},{"from":219.35,"to":223.13,"location":2,"content":"the right time to serve then expand"},{"from":223.13,"to":224.84,"location":2,"content":"forward again"},{"from":224.84,"to":227.03,"location":2,"content":"you should definitely always make sure"},{"from":227.03,"to":228.89,"location":2,"content":"that you can completely over fit on your"},{"from":228.89,"to":231.56,"location":2,"content":"training data set that sort of and not"},{"from":231.56,"to":233.33,"location":2,"content":"quite a proof but as at least the first"},{"from":233.33,"to":235.4,"location":2,"content":"good requirement for your model being"},{"from":235.4,"to":240.2,"location":2,"content":"implemented properly you you know part"},{"from":240.2,"to":241.97,"location":2,"content":"of the trick of being a successful deep"},{"from":241.97,"to":244.49,"location":2,"content":"learning researcher is actually managing"},{"from":244.49,"to":246.56,"location":2,"content":"to get things done and not wasting a ton"},{"from":246.56,"to":248.78,"location":2,"content":"of time and so it definitely always"},{"from":248.78,"to":251.15,"location":2,"content":"helps just to be you know plotting as"},{"from":251.15,"to":253.13,"location":2,"content":"you go along your training and dev"},{"from":253.13,"to":254.99,"location":2,"content":"errors so that you can sort of tell if"},{"from":254.99,"to":256.94,"location":2,"content":"things are working or if things aren't"},{"from":256.94,"to":259.52,"location":2,"content":"working and you should abandon and start"},{"from":259.52,"to":261.26,"location":2,"content":"again with a new experiment that just"},{"from":261.26,"to":263.66,"location":2,"content":"things like that save you hours and get"},{"from":263.66,"to":266.72,"location":2,"content":"you more done and so then once things"},{"from":266.72,"to":268.25,"location":2,"content":"are working this sort of a whole bunch"},{"from":268.25,"to":269.93,"location":2,"content":"of things to make it work better"},{"from":269.93,"to":272.93,"location":2,"content":"there's regularization with l2 and drop"},{"from":272.93,"to":275.75,"location":2,"content":"out there's time to do hyper parameter"},{"from":275.75,"to":276.88,"location":2,"content":"search"},{"from":276.88,"to":279.98,"location":2,"content":"and you know often doing these things"},{"from":279.98,"to":282.23,"location":2,"content":"and make quite a lot of difference to"},{"from":282.23,"to":284.9,"location":2,"content":"what your final results are and so it's"},{"from":284.9,"to":286.79,"location":2,"content":"good to have time to do those things but"},{"from":286.79,"to":289.37,"location":2,"content":"qu you want to get things working first"},{"from":289.37,"to":292.67,"location":2,"content":"before you go on to that and sort of"},{"from":292.67,"to":294.86,"location":2,"content":"really encourage people to still stop by"},{"from":294.86,"to":296.48,"location":2,"content":"in office hours if you've got any"},{"from":296.48,"to":299.03,"location":2,"content":"problems and we'll try our best to help"},{"from":299.03,"to":301.34,"location":2,"content":"though here within the limitations of"},{"from":301.34,"to":303.41,"location":2,"content":"what we can do from just being hit cold"},{"from":303.41,"to":309.32,"location":2,"content":"with problems okay yeah so I wanted to"},{"from":309.32,"to":311.71,"location":2,"content":"sort of just say some general remarks"},{"from":311.71,"to":317.62,"location":2,"content":"about language and theories of language"},{"from":317.62,"to":320.6,"location":2,"content":"that in the context that motivate this"},{"from":320.6,"to":324.23,"location":2,"content":"tree recursive networks so this is an"},{"from":324.23,"to":326.45,"location":2,"content":"art installation at Carnegie Mellon"},{"from":326.45,"to":329.75,"location":2,"content":"University and as an NLP person I really"},{"from":329.75,"to":332.6,"location":2,"content":"love this art installation so we need"},{"from":332.6,"to":334.64,"location":2,"content":"better art installations around the"},{"from":334.64,"to":337.52,"location":2,"content":"Stanford School of Engineering so this"},{"from":337.52,"to":340.28,"location":2,"content":"is the bag of words art installation"},{"from":340.28,"to":342.29,"location":2,"content":"there's the bag with a lot of words in"},{"from":342.29,"to":344.48,"location":2,"content":"it and you see down here there are the"},{"from":344.48,"to":347.48,"location":2,"content":"stop words the dart and the us that have"},{"from":347.48,"to":350.12,"location":2,"content":"fallen out of the bag and a represented"},{"from":350.12,"to":351.91,"location":2,"content":"on the ground as the stop words"},{"from":351.91,"to":356.59,"location":2,"content":"beautiful artwork right so one of the"},{"from":356.59,"to":359.69,"location":2,"content":"interesting things that has been found"},{"from":359.69,"to":362.9,"location":2,"content":"about NLP models of language and I think"},{"from":362.9,"to":365.36,"location":2,"content":"this is even more true in the deep"},{"from":365.36,"to":366.74,"location":2,"content":"learning world than it used to be"},{"from":366.74,"to":369.8,"location":2,"content":"previously is boy you can do a lot with"},{"from":369.8,"to":372.11,"location":2,"content":"bag of words models right that you can"},{"from":372.11,"to":375.77,"location":2,"content":"just often get a lot of power by saying"},{"from":375.77,"to":378.2,"location":2,"content":"well let's get our new word vectors"},{"from":378.2,"to":380.39,"location":2,"content":"we're gonna average them or max pull"},{"from":380.39,"to":382.46,"location":2,"content":"them or something like this and do"},{"from":382.46,"to":384.44,"location":2,"content":"nothing more and that gives me a pretty"},{"from":384.44,"to":386.72,"location":2,"content":"good sentence representation or document"},{"from":386.72,"to":388.43,"location":2,"content":"representation that I could use in a"},{"from":388.43,"to":390.83,"location":2,"content":"classifier or something and sometimes"},{"from":390.83,"to":392.96,"location":2,"content":"you can do not much more than that and"},{"from":392.96,"to":394.43,"location":2,"content":"get even better so people have done"},{"from":394.43,"to":396.65,"location":2,"content":"things like deep averaging networks"},{"from":396.65,"to":399.2,"location":2,"content":"where you're taking the output of a bag"},{"from":399.2,"to":400.76,"location":2,"content":"of words model and sort of feeding it"},{"from":400.76,"to":401.87,"location":2,"content":"through a couple more layers and"},{"from":401.87,"to":405.73,"location":2,"content":"improving things so that is in complete"},{"from":405.73,"to":408.83,"location":2,"content":"distinction to what's been dominant in"},{"from":408.83,"to":409.62,"location":2,"content":"linguistic"},{"from":409.62,"to":412.13,"location":2,"content":"of looking at language structure that"},{"from":412.13,"to":415.74,"location":2,"content":"typically in linguistics the emphasis"},{"from":415.74,"to":418.83,"location":2,"content":"has been on identifying kind of huge"},{"from":418.83,"to":422.04,"location":2,"content":"amounts of structure of linguistics"},{"from":422.04,"to":424.53,"location":2,"content":"utterances through very complex"},{"from":424.53,"to":426.51,"location":2,"content":"formalisms I guess this is sort of a bit"},{"from":426.51,"to":428.94,"location":2,"content":"of a picture of a Chomsky in minimalism"},{"from":428.94,"to":431.82,"location":2,"content":"syntactic tree and the one at the top is"},{"from":431.82,"to":434.37,"location":2,"content":"a bit of a picture of had driven phrase"},{"from":434.37,"to":436.32,"location":2,"content":"structure grammar theory there was"},{"from":436.32,"to":440.04,"location":2,"content":"predominantly developed at Stanford in"},{"from":440.04,"to":444.18,"location":2,"content":"the in the 90s but so very complex data"},{"from":444.18,"to":446.31,"location":2,"content":"structures and articulator structures"},{"from":446.31,"to":449.28,"location":2,"content":"used to describe linguistics and there's"},{"from":449.28,"to":452.37,"location":2,"content":"a huge gap between these two things"},{"from":452.37,"to":456.18,"location":2,"content":"and you might think that you know surely"},{"from":456.18,"to":457.86,"location":2,"content":"there's some good points in the middle"},{"from":457.86,"to":460.2,"location":2,"content":"where we have a certain amount of"},{"from":460.2,"to":462.36,"location":2,"content":"structure and that's going to help us do"},{"from":462.36,"to":466.77,"location":2,"content":"what we want and so in particular that"},{"from":466.77,"to":468.87,"location":2,"content":"if we wanted to semantically interpret"},{"from":468.87,"to":471.42,"location":2,"content":"language it seems like we don't just"},{"from":471.42,"to":473.79,"location":2,"content":"want to have word vectors we want to"},{"from":473.79,"to":476.28,"location":2,"content":"have meanings of bigger phrases so"},{"from":476.28,"to":478.86,"location":2,"content":"here's the snowboarder at leaping over a"},{"from":478.86,"to":482.04,"location":2,"content":"mogul and a person on a snowboard jumps"},{"from":482.04,"to":484.38,"location":2,"content":"into the air and what we'd like to be"},{"from":484.38,"to":486.36,"location":2,"content":"able to say is that the snowboarder"},{"from":486.36,"to":489.09,"location":2,"content":"means basically the same thing as a"},{"from":489.09,"to":491.52,"location":2,"content":"person on a snowboard so we want to have"},{"from":491.52,"to":494.37,"location":2,"content":"these chunks of language which in"},{"from":494.37,"to":496.8,"location":2,"content":"linguistics of be constituents phrases"},{"from":496.8,"to":499.32,"location":2,"content":"and say that they have a meaning and"},{"from":499.32,"to":501.18,"location":2,"content":"we'd like to be able to compare their"},{"from":501.18,"to":504.06,"location":2,"content":"meaning now we've looked at at least one"},{"from":504.06,"to":506.01,"location":2,"content":"tool that allows us to have chunks of"},{"from":506.01,"to":507.63,"location":2,"content":"language right because we looked at"},{"from":507.63,"to":509.4,"location":2,"content":"convolutional neural networks where you"},{"from":509.4,"to":511.71,"location":2,"content":"could take three words and make a"},{"from":511.71,"to":513.54,"location":2,"content":"representation of the convolutional"},{"from":513.54,"to":516.15,"location":2,"content":"neural network but the fundamental"},{"from":516.15,"to":519.06,"location":2,"content":"difference is that in human languages"},{"from":519.06,"to":520.98,"location":2,"content":"you have these chunks that have meaning"},{"from":520.98,"to":523.56,"location":2,"content":"there are different sizes so we'd like"},{"from":523.56,"to":527.13,"location":2,"content":"to say the snowboarder is pretty much"},{"from":527.13,"to":529.08,"location":2,"content":"semantically equivalent to a person on a"},{"from":529.08,"to":532.17,"location":2,"content":"snowboard but the top one is two words"},{"from":532.17,"to":534.45,"location":2,"content":"long and the bottom one is five words"},{"from":534.45,"to":537,"location":2,"content":"long and so if we're going to be able to"},{"from":537,"to":540.9,"location":2,"content":"do that we somehow want to have these"},{"from":540.9,"to":542.98,"location":2,"content":"sort of constituent chunks"},{"from":542.98,"to":544.78,"location":2,"content":"and be able to work with and represent"},{"from":544.78,"to":547.54,"location":2,"content":"them in neural networks and that's sort"},{"from":547.54,"to":552.46,"location":2,"content":"of the central idea of what motivated"},{"from":552.46,"to":554.14,"location":2,"content":"some of the sort of tree structured"},{"from":554.14,"to":555.73,"location":2,"content":"neural networks are about to show you"},{"from":555.73,"to":559.03,"location":2,"content":"there's another related thing that you"},{"from":559.03,"to":561.28,"location":2,"content":"might want to think about is you know a"},{"from":561.28,"to":564.46,"location":2,"content":"person on a snowboard how do human"},{"from":564.46,"to":567.01,"location":2,"content":"beings manage to understand what that"},{"from":567.01,"to":569.59,"location":2,"content":"means and then a person on a snowboard"},{"from":569.59,"to":570.97,"location":2,"content":"jumps into the air"},{"from":570.97,"to":573.73,"location":2,"content":"how does people manage to understand"},{"from":573.73,"to":577.33,"location":2,"content":"what that means and it sort of seems"},{"from":577.33,"to":581.67,"location":2,"content":"like the only possible answer to this is"},{"from":581.67,"to":584.2,"location":2,"content":"what's normally referred to as the"},{"from":584.2,"to":586.81,"location":2,"content":"principle of compositionality that"},{"from":586.81,"to":589.3,"location":2,"content":"people know the word person they know"},{"from":589.3,"to":591.09,"location":2,"content":"the word on they know the words"},{"from":591.09,"to":593.44,"location":2,"content":"snowboard therefore they can work out"},{"from":593.44,"to":596.53,"location":2,"content":"what on a snowboard means and they can"},{"from":596.53,"to":598.3,"location":2,"content":"work out what person on a snowboard"},{"from":598.3,"to":601.39,"location":2,"content":"means by knowing the meanings of"},{"from":601.39,"to":603.1,"location":2,"content":"components and putting them together"},{"from":603.1,"to":607.24,"location":2,"content":"into bigger pieces there's a there's a"},{"from":607.24,"to":610.81,"location":2,"content":"famous applied mathematician"},{"from":610.81,"to":613.78,"location":2,"content":"statistician at Brown University study"},{"from":613.78,"to":616.21,"location":2,"content":"Minh and I guess the way he summarized"},{"from":616.21,"to":618.37,"location":2,"content":"this is either the principle of"},{"from":618.37,"to":621.82,"location":2,"content":"compositionality is true or God exists"},{"from":621.82,"to":625.51,"location":2,"content":"very shiva's where you take that as you"},{"from":625.51,"to":628.39,"location":2,"content":"as you want but you know I think what he"},{"from":628.39,"to":631.27,"location":2,"content":"meant was well you know you can just"},{"from":631.27,"to":633.28,"location":2,"content":"make these infinite number of infinitely"},{"from":633.28,"to":635.02,"location":2,"content":"long sentences and human beings"},{"from":635.02,"to":637.42,"location":2,"content":"understand them then it just has to be"},{"from":637.42,"to":640.03,"location":2,"content":"that people can know about words and"},{"from":640.03,"to":642.52,"location":2,"content":"waits to combine meanings and and make"},{"from":642.52,"to":644.23,"location":2,"content":"bigger meanings cause you know how else"},{"from":644.23,"to":646.39,"location":2,"content":"could it possibly work that people could"},{"from":646.39,"to":648.94,"location":2,"content":"understand sentences and so we want to"},{"from":648.94,"to":651.04,"location":2,"content":"be able to do that we want to be able to"},{"from":651.04,"to":653.32,"location":2,"content":"work out semantic compositions of"},{"from":653.32,"to":655.45,"location":2,"content":"smaller elements to work out the"},{"from":655.45,"to":658.09,"location":2,"content":"meanings of bigger pieces and that this"},{"from":658.09,"to":661,"location":2,"content":"obviously isn't only a linguistic thing"},{"from":661,"to":663.88,"location":2,"content":"compositionality appears in other places"},{"from":663.88,"to":666.57,"location":2,"content":"as well right so if you want to"},{"from":666.57,"to":669.16,"location":2,"content":"understand how some piece of machinery"},{"from":669.16,"to":672.01,"location":2,"content":"works what you kind of want to know is"},{"from":672.01,"to":674.47,"location":2,"content":"it has different sub components and if"},{"from":674.47,"to":676.62,"location":2,"content":"you can understand how the differ"},{"from":676.62,"to":678.48,"location":2,"content":"subcomponents work and how they're"},{"from":678.48,"to":681.24,"location":2,"content":"fitted together then you might have some"},{"from":681.24,"to":683.34,"location":2,"content":"understanding of how the whole scene"},{"from":683.34,"to":689.01,"location":2,"content":"works and compositionality seems to be"},{"from":689.01,"to":691.92,"location":2,"content":"worked at work in vision as well so here"},{"from":691.92,"to":694.47,"location":2,"content":"is a scene and again it seems like this"},{"from":694.47,"to":696.48,"location":2,"content":"scene has parts so there are little"},{"from":696.48,"to":698.94,"location":2,"content":"parts that go together right so there"},{"from":698.94,"to":700.62,"location":2,"content":"are people that go together into a crowd"},{"from":700.62,"to":703.11,"location":2,"content":"of people and there's a roofer and a"},{"from":703.11,"to":704.91,"location":2,"content":"second floor and another bit of roof and"},{"from":704.91,"to":707.58,"location":2,"content":"the first floor that go together into a"},{"from":707.58,"to":709.65,"location":2,"content":"picture of this church and so this is"},{"from":709.65,"to":712.02,"location":2,"content":"also kind of a compositional scene in"},{"from":712.02,"to":715.53,"location":2,"content":"which pieces go together so it sort of"},{"from":715.53,"to":717.42,"location":2,"content":"seems like certainly for language"},{"from":717.42,"to":719.91,"location":2,"content":"understanding and then really for a lot"},{"from":719.91,"to":721.74,"location":2,"content":"of the other things that we use for"},{"from":721.74,"to":724.08,"location":2,"content":"intelligence then we somehow need to be"},{"from":724.08,"to":726.87,"location":2,"content":"able to understand bigger things from"},{"from":726.87,"to":731.21,"location":2,"content":"knowing about smaller parts yeah so"},{"from":731.21,"to":733.56,"location":2,"content":"computational so the most fun I"},{"from":733.56,"to":735.63,"location":2,"content":"mentioned this earlier sometime the most"},{"from":735.63,"to":739.11,"location":2,"content":"famous linguist is Noam Chomsky at MIT"},{"from":739.11,"to":743.64,"location":2,"content":"and you know really computational"},{"from":743.64,"to":745.86,"location":2,"content":"linguists a lot of the time haven't been"},{"from":745.86,"to":749.16,"location":2,"content":"that friendly to linguistics linguists"},{"from":749.16,"to":751.97,"location":2,"content":"and in particular some of Noam Chomsky's"},{"from":751.97,"to":755.19,"location":2,"content":"theories of language because really he's"},{"from":755.19,"to":758.33,"location":2,"content":"never been sympathetic to the idea of"},{"from":758.33,"to":760.83,"location":2,"content":"machine learning or in general just sort"},{"from":760.83,"to":763.26,"location":2,"content":"of the empirical ability to learn from"},{"from":763.26,"to":766.1,"location":2,"content":"data he's sort of has always been"},{"from":766.1,"to":768.75,"location":2,"content":"wanting to refuse that that exists but"},{"from":768.75,"to":771.03,"location":2,"content":"if we nevertheless look for a little bit"},{"from":771.03,"to":774.72,"location":2,"content":"of insight on that you know this is a"},{"from":774.72,"to":777.15,"location":2,"content":"recent paper of Chomsky's with authors"},{"from":777.15,"to":779.25,"location":2,"content":"and that they're sort of trying to give"},{"from":779.25,"to":782.28,"location":2,"content":"a version of what is unique about human"},{"from":782.28,"to":785.76,"location":2,"content":"language and essentially what they zero"},{"from":785.76,"to":788.73,"location":2,"content":"in on is that well if you're sort of"},{"from":788.73,"to":791.16,"location":2,"content":"looking at you know humans versus other"},{"from":791.16,"to":793.89,"location":2,"content":"fairly intelligent creatures they"},{"from":793.89,"to":795.93,"location":2,"content":"suggest that the defining difference of"},{"from":795.93,"to":799.02,"location":2,"content":"human beings is that they have this"},{"from":799.02,"to":803.13,"location":2,"content":"ability to model recursion and so the"},{"from":803.13,"to":805.94,"location":2,"content":"this paper argues that the the singular"},{"from":805.94,"to":808.08,"location":2,"content":"distinction that allowed language to"},{"from":808.08,"to":810.09,"location":2,"content":"develop in human beings"},{"from":810.09,"to":812.07,"location":2,"content":"that we could put together smaller parts"},{"from":812.07,"to":814.38,"location":2,"content":"to make bigger things in a recursive"},{"from":814.38,"to":816.45,"location":2,"content":"process and that that was the sort of"},{"from":816.45,"to":820.02,"location":2,"content":"defining new ability not sure not sure I"},{"from":820.02,"to":823.14,"location":2,"content":"believe that or not you can decide what"},{"from":823.14,"to":827.4,"location":2,"content":"you think but what I think is certainly"},{"from":827.4,"to":830,"location":2,"content":"the case is if there's just"},{"from":830,"to":832.26,"location":2,"content":"incontrovertible that the structure of"},{"from":832.26,"to":835.38,"location":2,"content":"human language sentences have these"},{"from":835.38,"to":839.64,"location":2,"content":"pieces of constituents that then form"},{"from":839.64,"to":841.83,"location":2,"content":"together hierarchically or recursively"},{"from":841.83,"to":844.86,"location":2,"content":"into bigger pieces as you go up in the"},{"from":844.86,"to":846.96,"location":2,"content":"tree and then particularly you get this"},{"from":846.96,"to":849.81,"location":2,"content":"recursion where you get a little noun"},{"from":849.81,"to":852.72,"location":2,"content":"phrase meat which then appears in a"},{"from":852.72,"to":854.67,"location":2,"content":"bigger noun phrase like spaghetti with"},{"from":854.67,"to":856.74,"location":2,"content":"meat and you can repeat that several"},{"from":856.74,"to":859.23,"location":2,"content":"times giving you a curse of structure"},{"from":859.23,"to":861.93,"location":2,"content":"and I have an example of that in blue up"},{"from":861.93,"to":864.57,"location":2,"content":"the top so the person standing next to"},{"from":864.57,"to":866.16,"location":2,"content":"the man from the company that purchased"},{"from":866.16,"to":868.62,"location":2,"content":"the firm that's used to work at that"},{"from":868.62,"to":873.51,"location":2,"content":"whole thing is big noun phrase but"},{"from":873.51,"to":876.42,"location":2,"content":"inside that there's a noun phrase the"},{"from":876.42,"to":878.07,"location":2,"content":"man from the company that purchased the"},{"from":878.07,"to":880.2,"location":2,"content":"firm that used to work ad which is"},{"from":880.2,"to":882.54,"location":2,"content":"another big noun phrase and well inside"},{"from":882.54,"to":887.19,"location":2,"content":"that there are smaller noun phrase like"},{"from":887.19,"to":888.84,"location":2,"content":"the company that purchased the firm used"},{"from":888.84,"to":891,"location":2,"content":"to work at but you know it's still got"},{"from":891,"to":893.52,"location":2,"content":"inside that noun phrases like the firm"},{"from":893.52,"to":895.68,"location":2,"content":"that you used to work at and actually"},{"from":895.68,"to":898.23,"location":2,"content":"even that's gotten inside at a smaller"},{"from":898.23,"to":901.59,"location":2,"content":"noun phrase which is just the word you"},{"from":901.59,"to":905.49,"location":2,"content":"so a individual pronoun is also a noun"},{"from":905.49,"to":910.61,"location":2,"content":"phrase so this kind of structuring of"},{"from":910.61,"to":913.56,"location":2,"content":"language where you get this sort of"},{"from":913.56,"to":915.72,"location":2,"content":"hierarchical structure at the same kind"},{"from":915.72,"to":917.4,"location":2,"content":"of things inside them I think that's"},{"from":917.4,"to":921.39,"location":2,"content":"just sort of totally totally correct the"},{"from":921.39,"to":923.55,"location":2,"content":"the claim then that you know our"},{"from":923.55,"to":927.18,"location":2,"content":"language is recursive I mean in a formal"},{"from":927.18,"to":929.7,"location":2,"content":"sense it's not quite clear that that's a"},{"from":929.7,"to":934.35,"location":2,"content":"it's a clear thing and that's the reason"},{"from":934.35,"to":937.23,"location":2,"content":"to say something this recursive it has"},{"from":937.23,"to":938.76,"location":2,"content":"to repeat out to infinity"},{"from":938.76,"to":942.24,"location":2,"content":"right so as soon as you put any bound on"},{"from":942.24,"to":943.59,"location":2,"content":"something and you"},{"from":943.59,"to":945.87,"location":2,"content":"say look that's it now I'm afraid you"},{"from":945.87,"to":948.54,"location":2,"content":"just gave me with five levels of nesting"},{"from":948.54,"to":950.94,"location":2,"content":"that's pretty implausible that someone"},{"from":950.94,"to":953.58,"location":2,"content":"is going to say that and so as soon as"},{"from":953.58,"to":955.83,"location":2,"content":"you sort of want to make an argument"},{"from":955.83,"to":958.32,"location":2,"content":"like okay even if they said that no one"},{"from":958.32,"to":959.85,"location":2,"content":"is going to say a noun phrase with ten"},{"from":959.85,"to":961.89,"location":2,"content":"levels of nesting and if you put some"},{"from":961.89,"to":965.28,"location":2,"content":"hard limit on it like that then in some"},{"from":965.28,"to":967.05,"location":2,"content":"sense it's not truly recursive because"},{"from":967.05,"to":970.08,"location":2,"content":"it doesn't go out to infinity but you"},{"from":970.08,"to":971.67,"location":2,"content":"know regardless what you think about"},{"from":971.67,"to":973.92,"location":2,"content":"that that doesn't negate the basic"},{"from":973.92,"to":975.9,"location":2,"content":"argument that you get this hierarchical"},{"from":975.9,"to":978.33,"location":2,"content":"structuring with the same kinds of"},{"from":978.33,"to":980.64,"location":2,"content":"things like noun phrases sentences verb"},{"from":980.64,"to":983.91,"location":2,"content":"phrases appearing inside each other in a"},{"from":983.91,"to":987.63,"location":2,"content":"way that has no clear bound the extent"},{"from":987.63,"to":990.42,"location":2,"content":"that I show you a complex sentence you"},{"from":990.42,"to":992.31,"location":2,"content":"can say I can make them an even bigger"},{"from":992.31,"to":994.83,"location":2,"content":"more complex sentence by putting it"},{"from":994.83,"to":997.38,"location":2,"content":"inside you said to me that and then"},{"from":997.38,"to":1000.68,"location":2,"content":"saying my sentence right so that's the"},{"from":1000.68,"to":1002.63,"location":2,"content":"sense in which it does appear to be a"},{"from":1002.63,"to":1005.3,"location":2,"content":"recursive generative process even though"},{"from":1005.3,"to":1008.21,"location":2,"content":"practically there are limits to how"},{"from":1008.21,"to":1011,"location":2,"content":"complex sentences people say and so"},{"from":1011,"to":1013.37,"location":2,"content":"that's the kind of structure that gets"},{"from":1013.37,"to":1017,"location":2,"content":"captured in these constituency structure"},{"from":1017,"to":1020.09,"location":2,"content":"trees so before the early time when we"},{"from":1020.09,"to":1021.98,"location":2,"content":"talked about parsing and you guys did"},{"from":1021.98,"to":1024.29,"location":2,"content":"some of it I emphasized dependency"},{"from":1024.29,"to":1027.32,"location":2,"content":"parsing but the other kind of parsing"},{"from":1027.32,"to":1029.21,"location":2,"content":"which is actually the kind that the"},{"from":1029.21,"to":1030.83,"location":2,"content":"models I'm going to talk about today I"},{"from":1030.83,"to":1034.25,"location":2,"content":"was using was this idea of what's often"},{"from":1034.25,"to":1037.22,"location":2,"content":"called constituency parsing or linguists"},{"from":1037.22,"to":1038.69,"location":2,"content":"often call at phrase structure grammar"},{"from":1038.69,"to":1043.52,"location":2,"content":"x' or in sort of computer science formal"},{"from":1043.52,"to":1045.65,"location":2,"content":"language theory these are context-free"},{"from":1045.65,"to":1050.09,"location":2,"content":"grammars where we are having these non"},{"from":1050.09,"to":1052.43,"location":2,"content":"terminals like noun phrase and verb"},{"from":1052.43,"to":1054.41,"location":2,"content":"phrase and that's inside anon the noun"},{"from":1054.41,"to":1056.24,"location":2,"content":"phrase that's inside another verb phrase"},{"from":1056.24,"to":1058.37,"location":2,"content":"which is inside more verb phrases"},{"from":1058.37,"to":1061.7,"location":2,"content":"heading up the sentence and so these are"},{"from":1061.7,"to":1064.85,"location":2,"content":"our constituency grammars and when we"},{"from":1064.85,"to":1066.92,"location":2,"content":"occasionally mentioned the pin treebank"},{"from":1066.92,"to":1069.41,"location":2,"content":"tree this was kind of an original pin"},{"from":1069.41,"to":1072.86,"location":2,"content":"treebank tree which is basically for a"},{"from":1072.86,"to":1074.63,"location":2,"content":"structure grammar like this with sort of"},{"from":1074.63,"to":1076.89,"location":2,"content":"various extra annotations"},{"from":1076.89,"to":1081.21,"location":2,"content":"put on the nodes okay so what it's seen"},{"from":1081.21,"to":1084.06,"location":2,"content":"what what you to capture some of these"},{"from":1084.06,"to":1085.98,"location":2,"content":"properties it seems like we'd like to"},{"from":1085.98,"to":1088.71,"location":2,"content":"have a neural model that can make use of"},{"from":1088.71,"to":1090.78,"location":2,"content":"some of this same kind of tree structure"},{"from":1090.78,"to":1094.41,"location":2,"content":"and so what we'd like to do for working"},{"from":1094.41,"to":1097.44,"location":2,"content":"out semantic similarity of constituents"},{"from":1097.44,"to":1100.77,"location":2,"content":"is we want to not only have a word"},{"from":1100.77,"to":1102.9,"location":2,"content":"vector space like we start off with"},{"from":1102.9,"to":1104.79,"location":2,"content":"right at the beginning of the quarter"},{"from":1104.79,"to":1107.72,"location":2,"content":"but we'd like to be able to take bigger"},{"from":1107.72,"to":1110.13,"location":2,"content":"constituents like noun phrases the"},{"from":1110.13,"to":1112.35,"location":2,"content":"country of my birth and the place where"},{"from":1112.35,"to":1115.17,"location":2,"content":"I was born and also give them a meaning"},{"from":1115.17,"to":1117.63,"location":2,"content":"and so it seems like what we'd like to"},{"from":1117.63,"to":1120.09,"location":2,"content":"do is have a method of computing the"},{"from":1120.09,"to":1123.48,"location":2,"content":"meaning of any phrase in a compositional"},{"from":1123.48,"to":1126.3,"location":2,"content":"manner such that the end result is also"},{"from":1126.3,"to":1129.92,"location":2,"content":"that these phrases could be stuck inside"},{"from":1129.92,"to":1132.99,"location":2,"content":"our vector space models so we're still"},{"from":1132.99,"to":1134.37,"location":2,"content":"going to stick with a vector space"},{"from":1134.37,"to":1136.8,"location":2,"content":"semantics of phrases and we want to"},{"from":1136.8,"to":1140.1,"location":2,"content":"compute the meanings of phrases and so"},{"from":1140.1,"to":1142.86,"location":2,"content":"then the question is how could we go"},{"from":1142.86,"to":1146.01,"location":2,"content":"about doing that and we'll answer number"},{"from":1146.01,"to":1147.84,"location":2,"content":"one is we kind of use the principle of"},{"from":1147.84,"to":1150.72,"location":2,"content":"compositionality since we're sure it's"},{"from":1150.72,"to":1152.82,"location":2,"content":"right and so well what the principle of"},{"from":1152.82,"to":1156.03,"location":2,"content":"compositionality essentially says if you"},{"from":1156.03,"to":1158.79,"location":2,"content":"want to work out the meaning well here"},{"from":1158.79,"to":1160.92,"location":2,"content":"it says of a sentence but the meaning of"},{"from":1160.92,"to":1163.89,"location":2,"content":"any phrase any constituent is you're"},{"from":1163.89,"to":1166.2,"location":2,"content":"going to build it by knowing the"},{"from":1166.2,"to":1169.62,"location":2,"content":"meanings of its words and then having"},{"from":1169.62,"to":1171.72,"location":2,"content":"rules that combine these meanings so"},{"from":1171.72,"to":1173.52,"location":2,"content":"starting off with the country of my"},{"from":1173.52,"to":1175.62,"location":2,"content":"birth I should be able to calculate a"},{"from":1175.62,"to":1178.02,"location":2,"content":"meaning of my birth and meaning of the"},{"from":1178.02,"to":1181.23,"location":2,"content":"country a meaning of of the my birth and"},{"from":1181.23,"to":1183.06,"location":2,"content":"then a meaning of the country of my"},{"from":1183.06,"to":1185.46,"location":2,"content":"birth so we'd have meaning composition"},{"from":1185.46,"to":1187.29,"location":2,"content":"rules which will let us calculate"},{"from":1187.29,"to":1190.89,"location":2,"content":"meanings upwards for larger constituents"},{"from":1190.89,"to":1195.87,"location":2,"content":"or sentences so that seems kind of the"},{"from":1195.87,"to":1197.85,"location":2,"content":"right thing to do and so then the"},{"from":1197.85,"to":1201.95,"location":2,"content":"question is well can we then build a"},{"from":1201.95,"to":1204.24,"location":2,"content":"model of how to do that"},{"from":1204.24,"to":1205.77,"location":2,"content":"well here's sort of a straightforward"},{"from":1205.77,"to":1210.27,"location":2,"content":"way of doing this okay so we"},{"from":1210.27,"to":1214.68,"location":2,"content":"we have word vectors for the words that"},{"from":1214.68,"to":1216.99,"location":2,"content":"we've calculated and what we'd like to"},{"from":1216.99,"to":1221.55,"location":2,"content":"do is work out then a meaning"},{"from":1221.55,"to":1223.89,"location":2,"content":"representation of this sentence and at"},{"from":1223.89,"to":1226.02,"location":2,"content":"this point we sort of have two things to"},{"from":1226.02,"to":1229.2,"location":2,"content":"do we have parsing to do where here"},{"from":1229.2,"to":1230.7,"location":2,"content":"what's the right structure of the"},{"from":1230.7,"to":1233.13,"location":2,"content":"sentence and then we have meaning"},{"from":1233.13,"to":1236.76,"location":2,"content":"computation to do of working out what is"},{"from":1236.76,"to":1238.47,"location":2,"content":"the meaning representation of this"},{"from":1238.47,"to":1242.19,"location":2,"content":"sentence so for parsing we'd sort of be"},{"from":1242.19,"to":1243.93,"location":2,"content":"building sort of noun phrase"},{"from":1243.93,"to":1245.76,"location":2,"content":"prepositional phrase verb phrase"},{"from":1245.76,"to":1248.7,"location":2,"content":"sentence kind of units to get the cat"},{"from":1248.7,"to":1251.73,"location":2,"content":"sat on the mat and then we'll what we if"},{"from":1251.73,"to":1254.79,"location":2,"content":"we had that we could then run some kind"},{"from":1254.79,"to":1257.49,"location":2,"content":"of meaning computation program and give"},{"from":1257.49,"to":1259.98,"location":2,"content":"us sort of a vector space meaning of"},{"from":1259.98,"to":1262.44,"location":2,"content":"these sentences so that's kind of what"},{"from":1262.44,"to":1264.81,"location":2,"content":"we want is to do both of those and in a"},{"from":1264.81,"to":1266.88,"location":2,"content":"little bit I'll show you an example of"},{"from":1266.88,"to":1269.13,"location":2,"content":"the kind of one way that you go about"},{"from":1269.13,"to":1271.5,"location":2,"content":"approaching that but before I do that"},{"from":1271.5,"to":1273.51,"location":2,"content":"just sort of stepping back for a moment"},{"from":1273.51,"to":1276.12,"location":2,"content":"as to what's different here write that"},{"from":1276.12,"to":1280.05,"location":2,"content":"here we had our recurrent neural network"},{"from":1280.05,"to":1281.79,"location":2,"content":"which in some sense has been our"},{"from":1281.79,"to":1284.64,"location":2,"content":"workhorse tool in this class up to now"},{"from":1284.64,"to":1287,"location":2,"content":"and it gives you it gives you a"},{"from":1287,"to":1288.9,"location":2,"content":"representation of the meaning of the"},{"from":1288.9,"to":1290.88,"location":2,"content":"country of my birth sort of you could"},{"from":1290.88,"to":1293.34,"location":2,"content":"either say that's the meaning of the"},{"from":1293.34,"to":1295.35,"location":2,"content":"country of my birth or we talked about"},{"from":1295.35,"to":1298.2,"location":2,"content":"other tricks like doing max pooling"},{"from":1298.2,"to":1300.75,"location":2,"content":"across all of these or you could have a"},{"from":1300.75,"to":1303.27,"location":2,"content":"separate node out here which sort of"},{"from":1303.27,"to":1304.95,"location":2,"content":"does attention over the ease so it does"},{"from":1304.95,"to":1309.48,"location":2,"content":"give you a sort of representation of the"},{"from":1309.48,"to":1313.38,"location":2,"content":"meaning of this of any subsequence of"},{"from":1313.38,"to":1316.98,"location":2,"content":"words as well but they sort of different"},{"from":1316.98,"to":1319.62,"location":2,"content":"right that this what the top the tree"},{"from":1319.62,"to":1323.01,"location":2,"content":"recursive neural network it requires a"},{"from":1323.01,"to":1326.67,"location":2,"content":"sentence or any kind of phrase to have a"},{"from":1326.67,"to":1328.68,"location":2,"content":"tree structure so we know what its"},{"from":1328.68,"to":1331.47,"location":2,"content":"component parts are but then we're"},{"from":1331.47,"to":1335.43,"location":2,"content":"working out meaning representations for"},{"from":1335.43,"to":1339.15,"location":2,"content":"the phrase that are sensitive to what"},{"from":1339.15,"to":1341.1,"location":2,"content":"its syntactic structure is that how the"},{"from":1341.1,"to":1344.2,"location":2,"content":"words go together to build phrases"},{"from":1344.2,"to":1346.69,"location":2,"content":"whereas for there were current neural"},{"from":1346.69,"to":1349.24,"location":2,"content":"network we're just in an oblivious way"},{"from":1349.24,"to":1351.49,"location":2,"content":"running a sequence model along and"},{"from":1351.49,"to":1354.97,"location":2,"content":"saying compute things and in the obvious"},{"from":1354.97,"to":1356.89,"location":2,"content":"it doesn't in any obvious way give a"},{"from":1356.89,"to":1359.86,"location":2,"content":"meaning representation of my birth or my"},{"from":1359.86,"to":1362.53,"location":2,"content":"birth contained inside it we sort of"},{"from":1362.53,"to":1364.87,"location":2,"content":"only have a meaning representation for"},{"from":1364.87,"to":1366.79,"location":2,"content":"the whole sequence whereas if we're"},{"from":1366.79,"to":1369.81,"location":2,"content":"doing things this way we do have meaning"},{"from":1369.81,"to":1371.83,"location":2,"content":"representations for the different"},{"from":1371.83,"to":1375.85,"location":2,"content":"meaningful parts of the sentence okay"},{"from":1375.85,"to":1377.32,"location":2,"content":"that makes sense of what we're trying to"},{"from":1377.32,"to":1381.91,"location":2,"content":"do okay so how could we do it go about"},{"from":1381.91,"to":1386.71,"location":2,"content":"doing that well the idea of how we could"},{"from":1386.71,"to":1389.02,"location":2,"content":"go about doing that is if we work"},{"from":1389.02,"to":1391.86,"location":2,"content":"bottom-up at the very bottom we have"},{"from":1391.86,"to":1395.68,"location":2,"content":"word vectors and so we want to"},{"from":1395.68,"to":1398.26,"location":2,"content":"recursively compute the meaning of"},{"from":1398.26,"to":1400.81,"location":2,"content":"bigger constituents so if we wanted to"},{"from":1400.81,"to":1403.3,"location":2,"content":"compute the meaning of on the mat what"},{"from":1403.3,"to":1406.42,"location":2,"content":"we can do is say well we have already"},{"from":1406.42,"to":1409.36,"location":2,"content":"have a meaning representation of on and"},{"from":1409.36,"to":1412.09,"location":2,"content":"mat so if we could feed those into a"},{"from":1412.09,"to":1414.49,"location":2,"content":"neural network because that's one tool"},{"from":1414.49,"to":1417.07,"location":2,"content":"we could maybe get out of it two things"},{"from":1417.07,"to":1421.51,"location":2,"content":"we could get out of it a goodness score"},{"from":1421.51,"to":1423.49,"location":2,"content":"so this is what we're going to use for"},{"from":1423.49,"to":1425.89,"location":2,"content":"parsing we're going to say do you"},{"from":1425.89,"to":1428.29,"location":2,"content":"believe do you believe you can put"},{"from":1428.29,"to":1431.86,"location":2,"content":"together on and the mat to form a good"},{"from":1431.86,"to":1434.32,"location":2,"content":"constituent that's part of a parse tree"},{"from":1434.32,"to":1437.02,"location":2,"content":"and this will be a big positive number"},{"from":1437.02,"to":1438.61,"location":2,"content":"if the answer is true and negative if"},{"from":1438.61,"to":1441.1,"location":2,"content":"that's not true and then we have a"},{"from":1441.1,"to":1443.77,"location":2,"content":"meaning composition device which says"},{"from":1443.77,"to":1446.68,"location":2,"content":"okay if you put together these two"},{"from":1446.68,"to":1448.44,"location":2,"content":"things what would be the meaning"},{"from":1448.44,"to":1451.59,"location":2,"content":"representation of what we put together"},{"from":1451.59,"to":1454.81,"location":2,"content":"and so this is the first model that we"},{"from":1454.81,"to":1457.09,"location":2,"content":"explored which was doing this in a"},{"from":1457.09,"to":1460.57,"location":2,"content":"pretty simple way right so here was our"},{"from":1460.57,"to":1464.1,"location":2,"content":"meaning composition device that we"},{"from":1464.1,"to":1466.18,"location":2,"content":"concatenated the two vectors of the"},{"from":1466.18,"to":1469.39,"location":2,"content":"constituents we multiply them by matrix"},{"from":1469.39,"to":1472.12,"location":2,"content":"atoms I'm a biased as usual put it"},{"from":1472.12,"to":1475.03,"location":2,"content":"through at NH this work is old enough"},{"from":1475.03,"to":1476.95,"location":2,"content":"it's sort of before things like values"},{"from":1476.95,"to":1477.82,"location":2,"content":"became pop"},{"from":1477.82,"to":1479.32,"location":2,"content":"pila but maybe it's better to have at an"},{"from":1479.32,"to":1482.29,"location":2,"content":"age anyway but more like recurrent"},{"from":1482.29,"to":1484.12,"location":2,"content":"neural network and so this was our"},{"from":1484.12,"to":1485.77,"location":2,"content":"meaning composition that gave the"},{"from":1485.77,"to":1488.2,"location":2,"content":"meaning of the parent and then to the"},{"from":1488.2,"to":1490.21,"location":2,"content":"side what the score if it was as to"},{"from":1490.21,"to":1492.61,"location":2,"content":"whether this was a good phrase we were"},{"from":1492.61,"to":1494.98,"location":2,"content":"taking that parent vector representation"},{"from":1494.98,"to":1498.91,"location":2,"content":"and multiplying it by another vector and"},{"from":1498.91,"to":1503.32,"location":2,"content":"that was giving us out a number if you"},{"from":1503.32,"to":1505.66,"location":2,"content":"think about it a bit while we're doing"},{"from":1505.66,"to":1507.7,"location":2,"content":"this you might think that this isn't"},{"from":1507.7,"to":1509.98,"location":2,"content":"quite a perfect model of meaning"},{"from":1509.98,"to":1511.6,"location":2,"content":"composition and later on in the class"},{"from":1511.6,"to":1514.33,"location":2,"content":"I'll talk about some more complex models"},{"from":1514.33,"to":1519.22,"location":2,"content":"that we then started to explore that"},{"from":1519.22,"to":1521.5,"location":2,"content":"this is sort of enough to get us going"},{"from":1521.5,"to":1524.44,"location":2,"content":"and this gave us a way of building a"},{"from":1524.44,"to":1527.26,"location":2,"content":"recursive neural network parser which"},{"from":1527.26,"to":1531.34,"location":2,"content":"both found parsers and worked out a"},{"from":1531.34,"to":1534.04,"location":2,"content":"meaning representation for them and so"},{"from":1534.04,"to":1536.26,"location":2,"content":"the way we did this was in the simplest"},{"from":1536.26,"to":1538.24,"location":2,"content":"possible way really which was to have a"},{"from":1538.24,"to":1539.62,"location":2,"content":"greedy parser"},{"from":1539.62,"to":1541.78,"location":2,"content":"so if we start off with the cat sat on"},{"from":1541.78,"to":1544.09,"location":2,"content":"the mat what we could do is say well"},{"from":1544.09,"to":1546.04,"location":2,"content":"maybe you should join that and cat"},{"from":1546.04,"to":1548.59,"location":2,"content":"together let's try that run it through"},{"from":1548.59,"to":1551.26,"location":2,"content":"our neural network it'll get a score and"},{"from":1551.26,"to":1553.81,"location":2,"content":"a meaning representation and while we"},{"from":1553.81,"to":1556.57,"location":2,"content":"could try doing that for cat and Sat we"},{"from":1556.57,"to":1559.03,"location":2,"content":"could try doing it for satin on we could"},{"from":1559.03,"to":1561.4,"location":2,"content":"try doing it for Anand our we could try"},{"from":1561.4,"to":1563.92,"location":2,"content":"doing it for that and mat and then at"},{"from":1563.92,"to":1566.83,"location":2,"content":"this point we'd say okay well the the"},{"from":1566.83,"to":1569.47,"location":2,"content":"best phrase that we can make combining"},{"from":1569.47,"to":1571.96,"location":2,"content":"these word vectors is the one for that"},{"from":1571.96,"to":1574.69,"location":2,"content":"cat so let's just commit to that one and"},{"from":1574.69,"to":1577.27,"location":2,"content":"it has this semantic representation and"},{"from":1577.27,"to":1580.45,"location":2,"content":"at this point we can essentially repeat"},{"from":1580.45,"to":1582.58,"location":2,"content":"now all the work we did over there we"},{"from":1582.58,"to":1584.23,"location":2,"content":"can just reuse because nothing has"},{"from":1584.23,"to":1587.2,"location":2,"content":"changed but we can also consider now"},{"from":1587.2,"to":1590.44,"location":2,"content":"joining the cat as a constituent with"},{"from":1590.44,"to":1593.44,"location":2,"content":"Sat and get a score for that and so at"},{"from":1593.44,"to":1595.66,"location":2,"content":"this point we decide okay the mat is the"},{"from":1595.66,"to":1598.02,"location":2,"content":"best constituent to build commit to that"},{"from":1598.02,"to":1600.43,"location":2,"content":"calculate a meaning representation for"},{"from":1600.43,"to":1603.25,"location":2,"content":"on the mat that looks good commit to"},{"from":1603.25,"to":1606.04,"location":2,"content":"that and kind of keep on chugging up and"},{"from":1606.04,"to":1608.53,"location":2,"content":"so we've got a mechanism for sort of"},{"from":1608.53,"to":1611.38,"location":2,"content":"choosing a pars of the sentence in a in"},{"from":1611.38,"to":1611.66,"location":2,"content":"a"},{"from":1611.66,"to":1613.4,"location":2,"content":"really manner but you know when we"},{"from":1613.4,"to":1615.38,"location":2,"content":"looked at the dependency parsing we're"},{"from":1615.38,"to":1617.66,"location":2,"content":"also doing that greedily right and"},{"from":1617.66,"to":1620.83,"location":2,"content":"coming up with a meaning representation"},{"from":1620.83,"to":1623.87,"location":2,"content":"okay so that was our first model of"},{"from":1623.87,"to":1625.76,"location":2,"content":"having a tree recursive neural network"},{"from":1625.76,"to":1629.51,"location":2,"content":"and using it for parsing there are a few"},{"from":1629.51,"to":1634.19,"location":2,"content":"more details here some of which probably"},{"from":1634.19,"to":1637.52,"location":2,"content":"aren't super important at this point"},{"from":1637.52,"to":1640.13,"location":2,"content":"right so we could score a tree by"},{"from":1640.13,"to":1643.82,"location":2,"content":"summing the scores at each node for"},{"from":1643.82,"to":1647,"location":2,"content":"working out for the optimization we were"},{"from":1647,"to":1648.98,"location":2,"content":"working out we're using this kind of Max"},{"from":1648.98,"to":1651.53,"location":2,"content":"margin loss that we've looked at in"},{"from":1651.53,"to":1655.76,"location":2,"content":"other places the simplest way to do"},{"from":1655.76,"to":1658.84,"location":2,"content":"things is completely greedily you just"},{"from":1658.84,"to":1661.7,"location":2,"content":"find the best local decision each point"},{"from":1661.7,"to":1663.26,"location":2,"content":"and make that structure and keep on"},{"from":1663.26,"to":1665.12,"location":2,"content":"going but if you want to do things a bit"},{"from":1665.12,"to":1668.53,"location":2,"content":"better and we explore this you could say"},{"from":1668.53,"to":1671.66,"location":2,"content":"we could do beam search we could explore"},{"from":1671.66,"to":1674.33,"location":2,"content":"out several good ways emerging and then"},{"from":1674.33,"to":1676.79,"location":2,"content":"the side later higher up the tree as to"},{"from":1676.79,"to":1680.39,"location":2,"content":"which was the best way to merge we"},{"from":1680.39,"to":1682.37,"location":2,"content":"haven't talked about it in this class"},{"from":1682.37,"to":1686.36,"location":2,"content":"but just to mention something in case"},{"from":1686.36,"to":1689.12,"location":2,"content":"people have seen it is traditional"},{"from":1689.12,"to":1691.43,"location":2,"content":"constituency parsing where you have"},{"from":1691.43,"to":1695.54,"location":2,"content":"symbols here like MP or VP there exist"},{"from":1695.54,"to":1698.57,"location":2,"content":"efficient dynamic programming algorithms"},{"from":1698.57,"to":1702.23,"location":2,"content":"where you can find the optimal pars of a"},{"from":1702.23,"to":1704.45,"location":2,"content":"sentence in polynomial time"},{"from":1704.45,"to":1706.7,"location":2,"content":"so in cubic time so if you have a"},{"from":1706.7,"to":1709.49,"location":2,"content":"regular context-free grammar and well so"},{"from":1709.49,"to":1711.23,"location":2,"content":"a regular probabilistic context-free"},{"from":1711.23,"to":1714.17,"location":2,"content":"grammar and you want to know what is the"},{"from":1714.17,"to":1716.06,"location":2,"content":"best part of the sentence according to"},{"from":1716.06,"to":1717.98,"location":2,"content":"the probabilistic context-free grammar"},{"from":1717.98,"to":1720.29,"location":2,"content":"you can write a cubic time dynamic"},{"from":1720.29,"to":1722.69,"location":2,"content":"programming algorithm and you can find"},{"from":1722.69,"to":1725.54,"location":2,"content":"it that's good and in the old days of"},{"from":1725.54,"to":1730.19,"location":2,"content":"cs2 24in before neural networks we used"},{"from":1730.19,"to":1732.29,"location":2,"content":"to have everyone do that the the most"},{"from":1732.29,"to":1735.77,"location":2,"content":"the most brain breaking assignment of"},{"from":1735.77,"to":1738.59,"location":2,"content":"the old cs2 24n was writing this dynamic"},{"from":1738.59,"to":1740.63,"location":2,"content":"program to do context-free grammar"},{"from":1740.63,"to":1744.32,"location":2,"content":"parsing of a sentence the slightly sad"},{"from":1744.32,"to":1745.22,"location":2,"content":"fact"},{"from":1745.22,"to":1747.05,"location":2,"content":"is once you go to these kind of neural"},{"from":1747.05,"to":1749.78,"location":2,"content":"network representations you can't write"},{"from":1749.78,"to":1752.09,"location":2,"content":"clever dynamic programming algorithms"},{"from":1752.09,"to":1754.34,"location":2,"content":"anymore because clever dynamic"},{"from":1754.34,"to":1757.01,"location":2,"content":"programming algorithms only work when"},{"from":1757.01,"to":1759.32,"location":2,"content":"you have symbols from a reasonably small"},{"from":1759.32,"to":1761.99,"location":2,"content":"set for your non terminals because if"},{"from":1761.99,"to":1764.9,"location":2,"content":"that's the case you can you kind of have"},{"from":1764.9,"to":1767.12,"location":2,"content":"collisions right you have lots of ways"},{"from":1767.12,"to":1768.98,"location":2,"content":"of parsing stuff lower down"},{"from":1768.98,"to":1771.86,"location":2,"content":"which come up turn out to be different"},{"from":1771.86,"to":1773.87,"location":2,"content":"ways to make a noun phrase or different"},{"from":1773.87,"to":1775.79,"location":2,"content":"ways to make a prepositional phrase and"},{"from":1775.79,"to":1777.68,"location":2,"content":"therefore you can save work with dynamic"},{"from":1777.68,"to":1779.87,"location":2,"content":"programming if you've got a model like"},{"from":1779.87,"to":1782.48,"location":2,"content":"this since everything that you build is"},{"from":1782.48,"to":1784.31,"location":2,"content":"going through layers of neural network"},{"from":1784.31,"to":1786.05,"location":2,"content":"and you've got a meaning representation"},{"from":1786.05,"to":1788.15,"location":2,"content":"of some high dimensional vector things"},{"from":1788.15,"to":1790.19,"location":2,"content":"are never going to collide and so you"},{"from":1790.19,"to":1791.9,"location":2,"content":"can never save work by doing dynamic"},{"from":1791.9,"to":1795.53,"location":2,"content":"programming and so you're either doing"},{"from":1795.53,"to":1797.57,"location":2,"content":"exponential work to explore out"},{"from":1797.57,"to":1799.73,"location":2,"content":"everything or else you're using some"},{"from":1799.73,"to":1801.86,"location":2,"content":"kind of beam to explore a bunch of"},{"from":1801.86,"to":1807.98,"location":2,"content":"likely stuff yeah we actually also"},{"from":1807.98,"to":1811.43,"location":2,"content":"applied this to vision at the same time"},{"from":1811.43,"to":1814.67,"location":2,"content":"so it wasn't just so completely a vague"},{"from":1814.67,"to":1818.78,"location":2,"content":"motivation of visual scenes have parts"},{"from":1818.78,"to":1821.15,"location":2,"content":"that we actually started exploring that"},{"from":1821.15,"to":1824.48,"location":2,"content":"well you could take these pieces of"},{"from":1824.48,"to":1828.5,"location":2,"content":"scenes and then work out representations"},{"from":1828.5,"to":1831.13,"location":2,"content":"for scenes using a similar form of"},{"from":1831.13,"to":1836.05,"location":2,"content":"compositionality and so in particular"},{"from":1836.05,"to":1839.24,"location":2,"content":"there was sort of this data set that was"},{"from":1839.24,"to":1842.57,"location":2,"content":"being used for multi-class segmentation"},{"from":1842.57,"to":1845.39,"location":2,"content":"envision where you start off with very"},{"from":1845.39,"to":1847.64,"location":2,"content":"small patches and then you want to"},{"from":1847.64,"to":1850.43,"location":2,"content":"combine them out into parts of a scene"},{"from":1850.43,"to":1852.53,"location":2,"content":"of sort of recognizing which part of the"},{"from":1852.53,"to":1855.32,"location":2,"content":"picture was the building the sky the"},{"from":1855.32,"to":1858.35,"location":2,"content":"road various other classes and we were"},{"from":1858.35,"to":1860.54,"location":2,"content":"actually at the time able to do this"},{"from":1860.54,"to":1863.54,"location":2,"content":"really rather well using one of these"},{"from":1863.54,"to":1865.22,"location":2,"content":"tree recursive structured neural"},{"from":1865.22,"to":1867.5,"location":2,"content":"networks better than preceding work and"},{"from":1867.5,"to":1871.66,"location":2,"content":"vision had done in the late 2000s decade"},{"from":1871.66,"to":1876.02,"location":2,"content":"okay so how can how can we build neural"},{"from":1876.02,"to":1879.26,"location":2,"content":"networks that do this kind of stuff"},{"from":1879.26,"to":1882.47,"location":2,"content":"and so when when we started off"},{"from":1882.47,"to":1884.6,"location":2,"content":"exploring this tree structured neural"},{"from":1884.6,"to":1887.42,"location":2,"content":"networks we thought that this was a cool"},{"from":1887.42,"to":1890.09,"location":2,"content":"original idea and no one had worked on"},{"from":1890.09,"to":1891.44,"location":2,"content":"tree structured neural networks"},{"from":1891.44,"to":1894.32,"location":2,"content":"successfully before but it turned out we"},{"from":1894.32,"to":1896.24,"location":2,"content":"were wrong that there are a couple of"},{"from":1896.24,"to":1900.25,"location":2,"content":"Germans in the mid-1990s who actually"},{"from":1900.25,"to":1902.27,"location":2,"content":"started looking at tree structured"},{"from":1902.27,"to":1905.15,"location":2,"content":"neural networks and had worked out the"},{"from":1905.15,"to":1907.13,"location":2,"content":"math of them so corresponding to the"},{"from":1907.13,"to":1909.23,"location":2,"content":"back propagation through time algorithm"},{"from":1909.23,"to":1911.36,"location":2,"content":"that ABI talked about when we were doing"},{"from":1911.36,"to":1913.34,"location":2,"content":"recurrent neural networks they worked"},{"from":1913.34,"to":1915.26,"location":2,"content":"out the tree structured case which they"},{"from":1915.26,"to":1917.57,"location":2,"content":"called back propagation through"},{"from":1917.57,"to":1921.83,"location":2,"content":"structure there are several slides on"},{"from":1921.83,"to":1925.01,"location":2,"content":"this in these slides but I think I'm"},{"from":1925.01,"to":1928.1,"location":2,"content":"gonna sort of skip them if anyone wants"},{"from":1928.1,"to":1929.66,"location":2,"content":"to look at them there on the web and you"},{"from":1929.66,"to":1931.88,"location":2,"content":"can look at them I mean there isn't"},{"from":1931.88,"to":1935.18,"location":2,"content":"actually anything that's new so if you"},{"from":1935.18,"to":1938.57,"location":2,"content":"remember with with bad scarring or"},{"from":1938.57,"to":1940.31,"location":2,"content":"something those early lectures of this"},{"from":1940.31,"to":1942.95,"location":2,"content":"class of working out the derivatives of"},{"from":1942.95,"to":1944.81,"location":2,"content":"neural networks and how it worked with"},{"from":1944.81,"to":1947.12,"location":2,"content":"recurrent neural networks it's sort of"},{"from":1947.12,"to":1949.25,"location":2,"content":"the same right you gave this recurrent"},{"from":1949.25,"to":1951.59,"location":2,"content":"matrix had different levels of tree"},{"from":1951.59,"to":1954.05,"location":2,"content":"structure you're summing the derivatives"},{"from":1954.05,"to":1957.62,"location":2,"content":"of everywhere it turns up the only"},{"from":1957.62,"to":1959,"location":2,"content":"difference is sort of because we now"},{"from":1959,"to":1960.86,"location":2,"content":"have tree structure you're sort of"},{"from":1960.86,"to":1965.51,"location":2,"content":"splitting things downwards so yeah so"},{"from":1965.51,"to":1968.03,"location":2,"content":"for drop we find the computer forwards"},{"from":1968.03,"to":1971.48,"location":2,"content":"and then when we're doing back crop when"},{"from":1971.48,"to":1973.58,"location":2,"content":"we've have the backward propagation we"},{"from":1973.58,"to":1975.32,"location":2,"content":"have the error signal coming from above"},{"from":1975.32,"to":1979.67,"location":2,"content":"we then combine it with the calculations"},{"from":1979.67,"to":1981.02,"location":2,"content":"that this node and then we're sort of"},{"from":1981.02,"to":1983.51,"location":2,"content":"sending it back in a tree structure down"},{"from":1983.51,"to":1987.47,"location":2,"content":"to each of the branches underneath us so"},{"from":1987.47,"to":1990.53,"location":2,"content":"that was our first version of things you"},{"from":1990.53,"to":1992.15,"location":2,"content":"know we got some decent result we got"},{"from":1992.15,"to":1994.01,"location":2,"content":"this good vision results and I showed"},{"from":1994.01,"to":1998.15,"location":2,"content":"you and it sort of seemed to do some"},{"from":1998.15,"to":2001.81,"location":2,"content":"good for language both for parsing and"},{"from":2001.81,"to":2003.97,"location":2,"content":"doing we had some results I haven't"},{"from":2003.97,"to":2006.01,"location":2,"content":"actually included here of sort of doing"},{"from":2006.01,"to":2009.37,"location":2,"content":"paraphrase judgment between sentences"},{"from":2009.37,"to":2012.7,"location":2,"content":"and and model things fail"},{"from":2012.7,"to":2015.91,"location":2,"content":"well but once we started thinking about"},{"from":2015.91,"to":2018.34,"location":2,"content":"it more it seemed like that very simple"},{"from":2018.34,"to":2021.61,"location":2,"content":"neural net function couldn't possibly"},{"from":2021.61,"to":2023.71,"location":2,"content":"compute the kind of meanings that we"},{"from":2023.71,"to":2026.47,"location":2,"content":"wanted to compute for sentence meanings"},{"from":2026.47,"to":2029.14,"location":2,"content":"and so we then sort of said about trying"},{"from":2029.14,"to":2030.88,"location":2,"content":"to come up with some more complex ways"},{"from":2030.88,"to":2033.94,"location":2,"content":"of working out kind of meaning"},{"from":2033.94,"to":2036.16,"location":2,"content":"composition functions and modes that"},{"from":2036.16,"to":2038.14,"location":2,"content":"could then be used to build a better"},{"from":2038.14,"to":2041.02,"location":2,"content":"neural network and sort of sudden some"},{"from":2041.02,"to":2044.11,"location":2,"content":"of the essence of that is on this slide"},{"from":2044.11,"to":2046.93,"location":2,"content":"that you know for the first version we"},{"from":2046.93,"to":2048.58,"location":2,"content":"just didn't have enough complexity of"},{"from":2048.58,"to":2050.8,"location":2,"content":"neural network frankly right so when we"},{"from":2050.8,"to":2053.65,"location":2,"content":"had two constituents we concatenated"},{"from":2053.65,"to":2056.55,"location":2,"content":"them and multiply that by a weight"},{"from":2056.55,"to":2060.01,"location":2,"content":"weight matrix and that was sort of"},{"from":2060.01,"to":2064.36,"location":2,"content":"essentially all we had and as I hope"},{"from":2064.36,"to":2067.15,"location":2,"content":"you've gone more of a sense of in this"},{"from":2067.15,"to":2069.13,"location":2,"content":"class if you just concatenate and"},{"from":2069.13,"to":2071.98,"location":2,"content":"multiply by weight matrix you're not"},{"from":2071.98,"to":2073.99,"location":2,"content":"actually modeling the interaction"},{"from":2073.99,"to":2076.18,"location":2,"content":"between these two vectors right because"},{"from":2076.18,"to":2078.49,"location":2,"content":"you can think of this weight matrix is"},{"from":2078.49,"to":2080.98,"location":2,"content":"just sort of being divided in two and"},{"from":2080.98,"to":2083.44,"location":2,"content":"half of it multiplies that this vector"},{"from":2083.44,"to":2085.81,"location":2,"content":"and half of it multiplies this vector so"},{"from":2085.81,"to":2088.18,"location":2,"content":"the meanings of these two things don't"},{"from":2088.18,"to":2090.61,"location":2,"content":"act on each other and so somehow you"},{"from":2090.61,"to":2092.95,"location":2,"content":"have to make your neural network more"},{"from":2092.95,"to":2095.56,"location":2,"content":"complex than that but the other way in"},{"from":2095.56,"to":2098.65,"location":2,"content":"which this seemed too simple is in the"},{"from":2098.65,"to":2101.92,"location":2,"content":"first model we had just one weight"},{"from":2101.92,"to":2105.27,"location":2,"content":"matrix which we use for everything and"},{"from":2105.27,"to":2108.07,"location":2,"content":"at least if you're a linguist and you're"},{"from":2108.07,"to":2109.39,"location":2,"content":"thinking about the structure of language"},{"from":2109.39,"to":2112.21,"location":2,"content":"you might start thinking of well wait a"},{"from":2112.21,"to":2114.28,"location":2,"content":"minute sometimes you're going to be"},{"from":2114.28,"to":2115.75,"location":2,"content":"putting together a verb and an object"},{"from":2115.75,"to":2120.01,"location":2,"content":"noun phrase hit the ball sometimes"},{"from":2120.01,"to":2121.39,"location":2,"content":"you're going to be putting together an"},{"from":2121.39,"to":2124.42,"location":2,"content":"article and a noun our ball sometimes"},{"from":2124.42,"to":2126.84,"location":2,"content":"you're going to be doing adjective or"},{"from":2126.84,"to":2130.03,"location":2,"content":"modification blue ball these things are"},{"from":2130.03,"to":2133.06,"location":2,"content":"very different in their semantics can it"},{"from":2133.06,"to":2135.16,"location":2,"content":"really be the case that you can just"},{"from":2135.16,"to":2137.05,"location":2,"content":"have one weight matrix that is this"},{"from":2137.05,"to":2139.51,"location":2,"content":"universal composition function for"},{"from":2139.51,"to":2141.31,"location":2,"content":"putting together meaning of phrases"},{"from":2141.31,"to":2143.83,"location":2,"content":"could that possibly work and you sort of"},{"from":2143.83,"to":2146.64,"location":2,"content":"might suspect it doesn't work"},{"from":2146.64,"to":2151.33,"location":2,"content":"and so I'm going to go on and show some"},{"from":2151.33,"to":2153.64,"location":2,"content":"of those different things but really"},{"from":2153.64,"to":2158.44,"location":2,"content":"before I show the different things I'm"},{"from":2158.44,"to":2160.84,"location":2,"content":"going to show one more version that sort"},{"from":2160.84,"to":2163.03,"location":2,"content":"of related to the first thing which"},{"from":2163.03,"to":2165.88,"location":2,"content":"actually gave a pretty successful and"},{"from":2165.88,"to":2170.83,"location":2,"content":"good parser or edit for doing context"},{"from":2170.83,"to":2175.06,"location":2,"content":"free style constituency parsing and so"},{"from":2175.06,"to":2177.85,"location":2,"content":"this was another way of getting away"},{"from":2177.85,"to":2181.68,"location":2,"content":"from the parsing being completely greedy"},{"from":2181.68,"to":2185.38,"location":2,"content":"which was to actually split apart the"},{"from":2185.38,"to":2188.2,"location":2,"content":"two parts of gee we have to come up with"},{"from":2188.2,"to":2191.65,"location":2,"content":"a tree structure for our sentence from"},{"from":2191.65,"to":2193.78,"location":2,"content":"let's compute the meaning of the"},{"from":2193.78,"to":2197.38,"location":2,"content":"sentence and so the thinking was well in"},{"from":2197.38,"to":2201.01,"location":2,"content":"terms of deciding what's a good tree"},{"from":2201.01,"to":2203.62,"location":2,"content":"structure for a sentence that's actually"},{"from":2203.62,"to":2205.69,"location":2,"content":"something you can do pretty well with a"},{"from":2205.69,"to":2208.12,"location":2,"content":"symbolic grammar that the problems with"},{"from":2208.12,"to":2210.85,"location":2,"content":"symbolic grammars aren't that they can't"},{"from":2210.85,"to":2213.43,"location":2,"content":"put tree structures over sentences the"},{"from":2213.43,"to":2215.29,"location":2,"content":"problems you have with those grammars is"},{"from":2215.29,"to":2217.12,"location":2,"content":"that they can't compute meaning"},{"from":2217.12,"to":2219.97,"location":2,"content":"representation and they're not very good"},{"from":2219.97,"to":2222.61,"location":2,"content":"at choosing between alternative tree"},{"from":2222.61,"to":2225.58,"location":2,"content":"structures but we could divide up the"},{"from":2225.58,"to":2228.28,"location":2,"content":"two parts so what we could do is say"},{"from":2228.28,"to":2230.34,"location":2,"content":"well let's just use a regular"},{"from":2230.34,"to":2233.35,"location":2,"content":"probabilistic context-free grammar to"},{"from":2233.35,"to":2235.39,"location":2,"content":"generate possible tree structures for"},{"from":2235.39,"to":2237.94,"location":2,"content":"sentences we can generate a key best"},{"from":2237.94,"to":2240.75,"location":2,"content":"list and say what are the 50 best"},{"from":2240.75,"to":2243.55,"location":2,"content":"context-free grammar structures for this"},{"from":2243.55,"to":2245.23,"location":2,"content":"sentence and that's something we can do"},{"from":2245.23,"to":2247.09,"location":2,"content":"very efficiently with dynamic"},{"from":2247.09,"to":2250.44,"location":2,"content":"programming algorithms and then we can"},{"from":2250.44,"to":2255.28,"location":2,"content":"work out a neural net that will work out"},{"from":2255.28,"to":2256.9,"location":2,"content":"the meaning representation of the"},{"from":2256.9,"to":2262.09,"location":2,"content":"sentence and so that led to this what's"},{"from":2262.09,"to":2265.27,"location":2,"content":"called syntactically untied recursive"},{"from":2265.27,"to":2268.93,"location":2,"content":"neural network so essentially what this"},{"from":2268.93,"to":2272.56,"location":2,"content":"is saying is that we had for each node"},{"from":2272.56,"to":2276.43,"location":2,"content":"in the sentence it's got a category of a"},{"from":2276.43,"to":2279.31,"location":2,"content":"symbolic context-free grammar so there"},{"from":2279.31,"to":2279.89,"location":2,"content":"category"},{"from":2279.89,"to":2283.31,"location":2,"content":"and B and C so when we put things"},{"from":2283.31,"to":2286.7,"location":2,"content":"together we'll be able to say okay we've"},{"from":2286.7,"to":2292.79,"location":2,"content":"got a rule that says X goes to BC so"},{"from":2292.79,"to":2295.94,"location":2,"content":"that licenses this node here so that"},{"from":2295.94,"to":2299.3,"location":2,"content":"part of the parsing is symbolic then"},{"from":2299.3,"to":2302.63,"location":2,"content":"then we want to work out the meaning of"},{"from":2302.63,"to":2306.62,"location":2,"content":"this phrase and well the second problem"},{"from":2306.62,"to":2309.95,"location":2,"content":"I talked about was surely just having"},{"from":2309.95,"to":2313.03,"location":2,"content":"one way of doing composition is"},{"from":2313.03,"to":2315.62,"location":2,"content":"expecting a lot too much to be able to"},{"from":2315.62,"to":2317.57,"location":2,"content":"have sort of urban object versus"},{"from":2317.57,"to":2320.03,"location":2,"content":"adjective and noun composed the same way"},{"from":2320.03,"to":2323.54,"location":2,"content":"so we had this idea of well since we've"},{"from":2323.54,"to":2325.97,"location":2,"content":"now know about the syntactic categories"},{"from":2325.97,"to":2328.28,"location":2,"content":"of the children that we maybe know that"},{"from":2328.28,"to":2330.53,"location":2,"content":"this is an adjective and this is a noun"},{"from":2330.53,"to":2333.47,"location":2,"content":"what we could do is have different"},{"from":2333.47,"to":2335.71,"location":2,"content":"weight matrices for composition"},{"from":2335.71,"to":2339.11,"location":2,"content":"depending on what the categories are so"},{"from":2339.11,"to":2342.02,"location":2,"content":"rather than where before there was just"},{"from":2342.02,"to":2344.75,"location":2,"content":"this one universal weight matrix which"},{"from":2344.75,"to":2346.94,"location":2,"content":"was meant to do all meaning composition"},{"from":2346.94,"to":2349.61,"location":2,"content":"here we can have this is the weight"},{"from":2349.61,"to":2351.95,"location":2,"content":"matrix for combining together the"},{"from":2351.95,"to":2354.17,"location":2,"content":"meanings of an adjective and a noun and"},{"from":2354.17,"to":2356.63,"location":2,"content":"it will compute the meaning of this"},{"from":2356.63,"to":2358.82,"location":2,"content":"constituent but then we'll have a"},{"from":2358.82,"to":2361.37,"location":2,"content":"different weight matrix for combining"},{"from":2361.37,"to":2364.1,"location":2,"content":"together the meanings of a determiner"},{"from":2364.1,"to":2370.45,"location":2,"content":"and a noun phrase or something like that"},{"from":2370.45,"to":2376.19,"location":2,"content":"okay yeah so I sort of already said this"},{"from":2376.19,"to":2378.74,"location":2,"content":"one I guess we wanted to be able to do"},{"from":2378.74,"to":2382.97,"location":2,"content":"things quickly and so our solution to be"},{"from":2382.97,"to":2385.4,"location":2,"content":"able to do that is we sort of used a"},{"from":2385.4,"to":2387.53,"location":2,"content":"probabilistic context-free grammar to"},{"from":2387.53,"to":2391.22,"location":2,"content":"find likely parsers and then only worked"},{"from":2391.22,"to":2394.4,"location":2,"content":"out our meaning for ones that were quite"},{"from":2394.4,"to":2396.53,"location":2,"content":"probable and so we call this result a"},{"from":2396.53,"to":2399.02,"location":2,"content":"compositional vector grammar which was a"},{"from":2399.02,"to":2401.93,"location":2,"content":"combination of a pcfg and a tree"},{"from":2401.93,"to":2408.64,"location":2,"content":"recursive neural network and yes so"},{"from":2408.64,"to":2411.59,"location":2,"content":"essentially at the time this actually"},{"from":2411.59,"to":2413.36,"location":2,"content":"gave a pretty good constituent"},{"from":2413.36,"to":2415.85,"location":2,"content":"see Pazza so there are sort of lots of"},{"from":2415.85,"to":2417.98,"location":2,"content":"results here the top ones are kind of"},{"from":2417.98,"to":2421.28,"location":2,"content":"our classic older Stanford parser which"},{"from":2421.28,"to":2424.07,"location":2,"content":"is a pcfg are the kind of parsers that"},{"from":2424.07,"to":2425.78,"location":2,"content":"people have built this is our"},{"from":2425.78,"to":2427.55,"location":2,"content":"compositional vector grammar"},{"from":2427.55,"to":2431.96,"location":2,"content":"at the time of this being done in 2013"},{"from":2431.96,"to":2434.06,"location":2,"content":"it wasn't the very best paths are"},{"from":2434.06,"to":2436.04,"location":2,"content":"available there'd been some better work"},{"from":2436.04,"to":2438.41,"location":2,"content":"by Eugene charniak at Brown but we"},{"from":2438.41,"to":2440.21,"location":2,"content":"actually had a pretty good parser"},{"from":2440.21,"to":2442.67,"location":2,"content":"coming out of that system but what was"},{"from":2442.67,"to":2446.66,"location":2,"content":"perhaps a bit more interesting was we we"},{"from":2446.66,"to":2448.79,"location":2,"content":"didn't only have a part of those meant"},{"from":2448.79,"to":2451.19,"location":2,"content":"to give the right parse trees we're also"},{"from":2451.19,"to":2453.86,"location":2,"content":"computing meaning representations of"},{"from":2453.86,"to":2457.82,"location":2,"content":"nodes and as a kind of a consequence of"},{"from":2457.82,"to":2460.16,"location":2,"content":"that you could look at not only meaning"},{"from":2460.16,"to":2462.71,"location":2,"content":"representations of nodes you could learn"},{"from":2462.71,"to":2464.72,"location":2,"content":"about the weight matrices that these"},{"from":2464.72,"to":2467.42,"location":2,"content":"models were learning when they combine"},{"from":2467.42,"to":2469.79,"location":2,"content":"together meanings so remember we have"},{"from":2469.79,"to":2472.4,"location":2,"content":"these sort of category specific W"},{"from":2472.4,"to":2474.74,"location":2,"content":"matrices that were going together with"},{"from":2474.74,"to":2478.28,"location":2,"content":"the children to work out the meaning so"},{"from":2478.28,"to":2480.2,"location":2,"content":"these are a little bit hard to interpret"},{"from":2480.2,"to":2483.23,"location":2,"content":"but the deal is when weilert these"},{"from":2483.23,"to":2485.84,"location":2,"content":"matrices we initialize them as a pair of"},{"from":2485.84,"to":2488.33,"location":2,"content":"diagonal matrices so these are so two by"},{"from":2488.33,"to":2490.7,"location":2,"content":"one rectangular matrices because there"},{"from":2490.7,"to":2495.04,"location":2,"content":"are two children so half of it is"},{"from":2495.04,"to":2497.3,"location":2,"content":"multiplying the left child the other"},{"from":2497.3,"to":2499.49,"location":2,"content":"half is multiplying the right child and"},{"from":2499.49,"to":2501.98,"location":2,"content":"we initialize them as sort of like a"},{"from":2501.98,"to":2505.25,"location":2,"content":"complete two identity matrices next to"},{"from":2505.25,"to":2507.35,"location":2,"content":"each other which would give us the sort"},{"from":2507.35,"to":2510.23,"location":2,"content":"of default semantics of just averaging"},{"from":2510.23,"to":2512.45,"location":2,"content":"until something different was learnt in"},{"from":2512.45,"to":2516.32,"location":2,"content":"the in the in the weight vectors and to"},{"from":2516.32,"to":2518.81,"location":2,"content":"the extent that sort of nothing"},{"from":2518.81,"to":2521.66,"location":2,"content":"interesting has been learnt by the model"},{"from":2521.66,"to":2524.42,"location":2,"content":"you'll get yellow along the diagonal and"},{"from":2524.42,"to":2527.45,"location":2,"content":"this sort of sky-blue in the rest of the"},{"from":2527.45,"to":2530.48,"location":2,"content":"field and the extent that it's learnt"},{"from":2530.48,"to":2532.61,"location":2,"content":"something interesting to take out of the"},{"from":2532.61,"to":2535.13,"location":2,"content":"semantics of a child you'll then start"},{"from":2535.13,"to":2537.8,"location":2,"content":"to see reds and oranges on the diagonal"},{"from":2537.8,"to":2541.13,"location":2,"content":"on dark blues and greens and stuff in"},{"from":2541.13,"to":2543.83,"location":2,"content":"the rest of the field so what you find"},{"from":2543.83,"to":2546.38,"location":2,"content":"is that if you train this model it's"},{"from":2546.38,"to":2547.07,"location":2,"content":"learn"},{"from":2547.07,"to":2552.17,"location":2,"content":"about which children of a phrase are"},{"from":2552.17,"to":2554.84,"location":2,"content":"actually the important ones so these"},{"from":2554.84,"to":2556.91,"location":2,"content":"ones are saying that if you're combining"},{"from":2556.91,"to":2558.32,"location":2,"content":"together a noun phrase and the"},{"from":2558.32,"to":2560.78,"location":2,"content":"coordination so something like the cat"},{"from":2560.78,"to":2563.63,"location":2,"content":"and that most of the semantics is to be"},{"from":2563.63,"to":2566.12,"location":2,"content":"found in the cat and not much of the"},{"from":2566.12,"to":2568.13,"location":2,"content":"semantics is going to be found in and"},{"from":2568.13,"to":2570.68,"location":2,"content":"whereas if you're combining together a"},{"from":2570.68,"to":2573.98,"location":2,"content":"possessive pronoun something like her or"},{"from":2573.98,"to":2578.77,"location":2,"content":"hears with a noun phrase insider like"},{"from":2578.77,"to":2582.53,"location":2,"content":"her tabby cat or something like that and"},{"from":2582.53,"to":2584.12,"location":2,"content":"most of the meaning is to be found"},{"from":2584.12,"to":2586.91,"location":2,"content":"inside the tabby cat constituent so it's"},{"from":2586.91,"to":2588.56,"location":2,"content":"actually learning where the important"},{"from":2588.56,"to":2593.3,"location":2,"content":"semantics of sentences is and there are"},{"from":2593.3,"to":2598.94,"location":2,"content":"lots of examples of that yeah this one"},{"from":2598.94,"to":2601.73,"location":2,"content":"sort of so this one shows a variety of"},{"from":2601.73,"to":2604.34,"location":2,"content":"modification structures where adjectives"},{"from":2604.34,"to":2609.44,"location":2,"content":"or adverbs modify either a noun phrase"},{"from":2609.44,"to":2612.4,"location":2,"content":"or an adjective phrase or just a single"},{"from":2612.4,"to":2615.26,"location":2,"content":"adjective ISM multiplying a noun phrase"},{"from":2615.26,"to":2617.93,"location":2,"content":"and the thing that you seem to notice is"},{"from":2617.93,"to":2619.67,"location":2,"content":"that there are particular dimensions"},{"from":2619.67,"to":2622.54,"location":2,"content":"which are kind of capturing sort of"},{"from":2622.54,"to":2625.22,"location":2,"content":"modification meaning so dimension six"},{"from":2625.22,"to":2628.73,"location":2,"content":"and dimension eleven is sort of showing"},{"from":2628.73,"to":2631.97,"location":2,"content":"up in these different combinations here"},{"from":2631.97,"to":2633.86,"location":2,"content":"sort of capturing meaning component so"},{"from":2633.86,"to":2636.74,"location":2,"content":"that was kind of neat and so this"},{"from":2636.74,"to":2638.8,"location":2,"content":"slightly more complex model actually"},{"from":2638.8,"to":2642.11,"location":2,"content":"worked pretty well at capturing a"},{"from":2642.11,"to":2645.41,"location":2,"content":"meaning of phrases and sentences so in"},{"from":2645.41,"to":2648.47,"location":2,"content":"this test here we were giving it at the"},{"from":2648.47,"to":2651.65,"location":2,"content":"system a test sentence and saying well"},{"from":2651.65,"to":2655.31,"location":2,"content":"what are the other what are sentences"},{"from":2655.31,"to":2658.43,"location":2,"content":"that are most similar in meaning nearest"},{"from":2658.43,"to":2661.49,"location":2,"content":"to paraphrases in our corpus for this"},{"from":2661.49,"to":2663.65,"location":2,"content":"sentence so far all the figures are"},{"from":2663.65,"to":2666.38,"location":2,"content":"adjusted for seasonal variations the two"},{"from":2666.38,"to":2668.66,"location":2,"content":"most similar other sentences in the"},{"from":2668.66,"to":2671.03,"location":2,"content":"corpus were all the numbers are adjusted"},{"from":2671.03,"to":2673.28,"location":2,"content":"for seasonal vet fluctuation that's a"},{"from":2673.28,"to":2675.2,"location":2,"content":"pretty easy one or all the figures are"},{"from":2675.2,"to":2677.18,"location":2,"content":"adjusted to remove usual seasonal"},{"from":2677.18,"to":2679.07,"location":2,"content":"patterns so that seems to be working"},{"from":2679.07,"to":2680.63,"location":2,"content":"pretty well night"},{"from":2680.63,"to":2683.63,"location":2,"content":"you know one comment on the offer haskó"},{"from":2683.63,"to":2685.67,"location":2,"content":"declined to say what country placed the"},{"from":2685.67,"to":2687.86,"location":2,"content":"order the semantics they're a bit more"},{"from":2687.86,"to":2689.3,"location":2,"content":"different but it seems like it is"},{"from":2689.3,"to":2691.64,"location":2,"content":"capturing something similar"},{"from":2691.64,"to":2693.59,"location":2,"content":"coastal wouldn't disclose the terms"},{"from":2693.59,"to":2695.45,"location":2,"content":"that's kind of a really interesting one"},{"from":2695.45,"to":2697.4,"location":2,"content":"because that one is actually very"},{"from":2697.4,"to":2699.92,"location":2,"content":"similar in meaning but it's expressed in"},{"from":2699.92,"to":2702.14,"location":2,"content":"a very different way in terms of the"},{"from":2702.14,"to":2704.27,"location":2,"content":"words and the syntactic structure that"},{"from":2704.27,"to":2705.88,"location":2,"content":"are used"},{"from":2705.88,"to":2709.7,"location":2,"content":"okay so that was progress because now we"},{"from":2709.7,"to":2712.34,"location":2,"content":"could have different matrices for"},{"from":2712.34,"to":2715.82,"location":2,"content":"different constituent types but there's"},{"from":2715.82,"to":2719.03,"location":2,"content":"still some reason to think that we"},{"from":2719.03,"to":2722.87,"location":2,"content":"didn't have enough power and that was we"},{"from":2722.87,"to":2725.18,"location":2,"content":"are still at heart using this very"},{"from":2725.18,"to":2728.66,"location":2,"content":"simple compositional structure we were"},{"from":2728.66,"to":2733.01,"location":2,"content":"just concatenated to children's vectors"},{"from":2733.01,"to":2735.32,"location":2,"content":"and multiplying it by a matrix so that"},{"from":2735.32,"to":2739.07,"location":2,"content":"means the two words didn't interact with"},{"from":2739.07,"to":2743.71,"location":2,"content":"each other in terms of their meaning but"},{"from":2743.71,"to":2747.14,"location":2,"content":"it seems like we want to have them"},{"from":2747.14,"to":2749.63,"location":2,"content":"interact in their meaning right so in"},{"from":2749.63,"to":2753.8,"location":2,"content":"particular if you if you think about"},{"from":2753.8,"to":2756.71,"location":2,"content":"human languages and the kind of things"},{"from":2756.71,"to":2758.27,"location":2,"content":"that people look at at linguistic"},{"from":2758.27,"to":2761.63,"location":2,"content":"semantics you get words that appear to"},{"from":2761.63,"to":2765.23,"location":2,"content":"be kind of modifiers or operators so the"},{"from":2765.23,"to":2768.86,"location":2,"content":"word very sort of doesn't mean much by"},{"from":2768.86,"to":2770.77,"location":2,"content":"itself I mean it means something like"},{"from":2770.77,"to":2774.14,"location":2,"content":"strengthening or more so or something"},{"from":2774.14,"to":2776.45,"location":2,"content":"like that but you know it doesn't really"},{"from":2776.45,"to":2778.76,"location":2,"content":"have a meaning right it doesn't have any"},{"from":2778.76,"to":2781.82,"location":2,"content":"D notation you can't show me very things"},{"from":2781.82,"to":2783.77,"location":2,"content":"right you can show me chairs and pins"},{"from":2783.77,"to":2787.22,"location":2,"content":"and children but you can't show me very"},{"from":2787.22,"to":2789.77,"location":2,"content":"things that the meaning of very seems to"},{"from":2789.77,"to":2792.26,"location":2,"content":"be that something comes after it good"},{"from":2792.26,"to":2795.44,"location":2,"content":"and this has a sort of an operator"},{"from":2795.44,"to":2798.68,"location":2,"content":"meaning of increase on the scale this"},{"from":2798.68,"to":2801.02,"location":2,"content":"thing and it can increase on the scale"},{"from":2801.02,"to":2803.15,"location":2,"content":"in either direction you can have very"},{"from":2803.15,"to":2806.48,"location":2,"content":"good or very bad so if we want to"},{"from":2806.48,"to":2810.47,"location":2,"content":"capture that kind of semantics it seems"},{"from":2810.47,"to":2812.15,"location":2,"content":"like we can't capture that kind of"},{"from":2812.15,"to":2813.93,"location":2,"content":"semantics by just can-can"},{"from":2813.93,"to":2816.66,"location":2,"content":"needing two vectors and multiplying them"},{"from":2816.66,"to":2819.9,"location":2,"content":"by a matrix it seems like what we really"},{"from":2819.9,"to":2824.13,"location":2,"content":"want to say is very is going to grab"},{"from":2824.13,"to":2827.4,"location":2,"content":"hold of the meaning of good and modify"},{"from":2827.4,"to":2829.68,"location":2,"content":"it in some ways to produce a new meaning"},{"from":2829.68,"to":2832.98,"location":2,"content":"for very good and indeed that's the kind"},{"from":2832.98,"to":2835.98,"location":2,"content":"of approach that's typically been done"},{"from":2835.98,"to":2839.43,"location":2,"content":"in linguistic semantics so in linguistic"},{"from":2839.43,"to":2841.23,"location":2,"content":"theories of semantics you'd normally say"},{"from":2841.23,"to":2844.26,"location":2,"content":"okay good has a meaning very as a"},{"from":2844.26,"to":2846.15,"location":2,"content":"function that takes in the meaning of"},{"from":2846.15,"to":2849.39,"location":2,"content":"good and returns a meaning very good and"},{"from":2849.39,"to":2852.87,"location":2,"content":"so we wanted to have a way of putting"},{"from":2852.87,"to":2855.81,"location":2,"content":"that into a neural network and so to try"},{"from":2855.81,"to":2857.91,"location":2,"content":"and come up with a new composition"},{"from":2857.91,"to":2861.45,"location":2,"content":"function as to how to do that and there"},{"from":2861.45,"to":2863.97,"location":2,"content":"are various ways that you could think"},{"from":2863.97,"to":2866.28,"location":2,"content":"about doing that and other people have"},{"from":2866.28,"to":2868.94,"location":2,"content":"had a couple of different attempts but"},{"from":2868.94,"to":2872.37,"location":2,"content":"essentially what was in our head is well"},{"from":2872.37,"to":2875.82,"location":2,"content":"we have word vectors and if we want to"},{"from":2875.82,"to":2879.15,"location":2,"content":"say that 'very takes the meaning of good"},{"from":2879.15,"to":2882.75,"location":2,"content":"and returns a new meaning the kind of"},{"from":2882.75,"to":2885.6,"location":2,"content":"obvious thing to do is to say very has a"},{"from":2885.6,"to":2887.88,"location":2,"content":"matrix attached to it because then we"},{"from":2887.88,"to":2890.88,"location":2,"content":"can use the the very matrix and multiply"},{"from":2890.88,"to":2893.81,"location":2,"content":"it by the good vector and we get a new"},{"from":2893.81,"to":2899.73,"location":2,"content":"vector coming out and so then well the"},{"from":2899.73,"to":2903.21,"location":2,"content":"problem is which in which words have"},{"from":2903.21,"to":2905.4,"location":2,"content":"vectors and which words have matrices"},{"from":2905.4,"to":2908.85,"location":2,"content":"and that's kind of hard to know the"},{"from":2908.85,"to":2913.47,"location":2,"content":"answer to I mean in particular words"},{"from":2913.47,"to":2916.5,"location":2,"content":"that act as operators can often"},{"from":2916.5,"to":2924.03,"location":2,"content":"themselves be modified and so that you"},{"from":2924.03,"to":2928.07,"location":2,"content":"know good can also good also is a"},{"from":2928.07,"to":2931.64,"location":2,"content":"operator right so that from a sort of a"},{"from":2931.64,"to":2934.38,"location":2,"content":"person you can have a good person that's"},{"from":2934.38,"to":2936.84,"location":2,"content":"sort of also an operator and very is"},{"from":2936.84,"to":2939.81,"location":2,"content":"modifying that good so the idea we came"},{"from":2939.81,"to":2941.61,"location":2,"content":"up with is let's not try and"},{"from":2941.61,"to":2944.01,"location":2,"content":"predetermine all of this why don't we"},{"from":2944.01,"to":2946.69,"location":2,"content":"say that every word and"},{"from":2946.69,"to":2950.5,"location":2,"content":"phrase has connected to it both a matrix"},{"from":2950.5,"to":2954.28,"location":2,"content":"and the vector so here's our very good"},{"from":2954.28,"to":2955.06,"location":2,"content":"movie"},{"from":2955.06,"to":2957.43,"location":2,"content":"so for each word we have a vector"},{"from":2957.43,"to":2959.86,"location":2,"content":"meaning and it has a matrix meaning and"},{"from":2959.86,"to":2962.2,"location":2,"content":"then as we start to build up phrases"},{"from":2962.2,"to":2965.11,"location":2,"content":"like very good they're also going to"},{"from":2965.11,"to":2967.36,"location":2,"content":"have a vector meaning and a matrix"},{"from":2967.36,"to":2973.36,"location":2,"content":"meaning and so what we proposed was so"},{"from":2973.36,"to":2975.94,"location":2,"content":"first of all we we would like to be able"},{"from":2975.94,"to":2982.57,"location":2,"content":"to calculate the vector meanings so to"},{"from":2982.57,"to":2985.39,"location":2,"content":"work out the vector meaning of a phrase"},{"from":2985.39,"to":2988.6,"location":2,"content":"like very good each word has a matrix"},{"from":2988.6,"to":2990.88,"location":2,"content":"meaning and so we're going to combine"},{"from":2990.88,"to":2992.83,"location":2,"content":"their opposing matrix and vector"},{"from":2992.83,"to":2995.62,"location":2,"content":"meanings so we're going to take the"},{"from":2995.62,"to":2998.23,"location":2,"content":"matrix meaning of good and multiply it"},{"from":2998.23,"to":3000.93,"location":2,"content":"by the vector meaning of 'very and we're"},{"from":3000.93,"to":3002.91,"location":2,"content":"going to take the matrix meaning of"},{"from":3002.91,"to":3005.31,"location":2,"content":"'very and multiply it by the vector"},{"from":3005.31,"to":3008.37,"location":2,"content":"meaning of good and so we're going to"},{"from":3008.37,"to":3011.67,"location":2,"content":"have both of those two things and then"},{"from":3011.67,"to":3013.35,"location":2,"content":"we're going to have a neural network"},{"from":3013.35,"to":3016.68,"location":2,"content":"layer like before that combine those"},{"from":3016.68,"to":3018.63,"location":2,"content":"together and so that sort of in the red"},{"from":3018.63,"to":3020.79,"location":2,"content":"box then those two things were"},{"from":3020.79,"to":3023.04,"location":2,"content":"concatenated and put through the kind of"},{"from":3023.04,"to":3025.38,"location":2,"content":"neural network layer we had before to"},{"from":3025.38,"to":3027.99,"location":2,"content":"give us a final vector meaning for this"},{"from":3027.99,"to":3032.16,"location":2,"content":"for the phrase and then we also needed a"},{"from":3032.16,"to":3035.31,"location":2,"content":"matrix meaning for the phrase and so for"},{"from":3035.31,"to":3039.3,"location":2,"content":"the matrix meaning for the fray is we"},{"from":3039.3,"to":3041.13,"location":2,"content":"did this kind of simple model which"},{"from":3041.13,"to":3043.23,"location":2,"content":"maybe actually wasn't very good which"},{"from":3043.23,"to":3047.55,"location":2,"content":"was to say let's just concatenate the"},{"from":3047.55,"to":3051.92,"location":2,"content":"two matrices of the the constituents"},{"from":3051.92,"to":3054.96,"location":2,"content":"multiply them by another matrix and"},{"from":3054.96,"to":3057.05,"location":2,"content":"that's then going to give us a matrix"},{"from":3057.05,"to":3060.66,"location":2,"content":"version of the parent node and so this"},{"from":3060.66,"to":3063.21,"location":2,"content":"was gave us our new more company more"},{"from":3063.21,"to":3066.87,"location":2,"content":"powerful composition procedure this did"},{"from":3066.87,"to":3069.99,"location":2,"content":"seem like it could do some kind of good"},{"from":3069.99,"to":3073.29,"location":2,"content":"things that captured a sort of operator"},{"from":3073.29,"to":3076.2,"location":2,"content":"semantics where one word modified the"},{"from":3076.2,"to":3079.83,"location":2,"content":"meaning of another word so here's a kind"},{"from":3079.83,"to":3080.65,"location":2,"content":"of a neat"},{"from":3080.65,"to":3086.89,"location":2,"content":"that we were able to do with this that"},{"from":3086.89,"to":3089.8,"location":2,"content":"we are wanting to be able to work out"},{"from":3089.8,"to":3093.31,"location":2,"content":"the semantics of an operator modifying"},{"from":3093.31,"to":3096.36,"location":2,"content":"another word so unbelievably annoying"},{"from":3096.36,"to":3100.44,"location":2,"content":"unbelievably awesome unbelievably sad"},{"from":3100.44,"to":3104.29,"location":2,"content":"not annoying not awesome not sad"},{"from":3104.29,"to":3110.65,"location":2,"content":"and so this was contrasting our old"},{"from":3110.65,"to":3114.67,"location":2,"content":"model versus the new model and this"},{"from":3114.67,"to":3117.54,"location":2,"content":"scale is a scale of positive to negative"},{"from":3117.54,"to":3120.76,"location":2,"content":"so this is completely negative to"},{"from":3120.76,"to":3123.88,"location":2,"content":"completely positive right and so the"},{"from":3123.88,"to":3130.15,"location":2,"content":"kind of contrasts you get that for not"},{"from":3130.15,"to":3133.27,"location":2,"content":"annoying that the simple model thought"},{"from":3133.27,"to":3136,"location":2,"content":"that this is pretty negative whereas the"},{"from":3136,"to":3138.13,"location":2,"content":"new model thinks this is pretty neutral"},{"from":3138.13,"to":3140.28,"location":2,"content":"and meaning and that seems to be"},{"from":3140.28,"to":3145.12,"location":2,"content":"reasonably correct for not sad that"},{"from":3145.12,"to":3147.52,"location":2,"content":"means it's a little bit positive and"},{"from":3147.52,"to":3150.52,"location":2,"content":"both models we're trying to kept capture"},{"from":3150.52,"to":3153.46,"location":2,"content":"that then you know the results here a"},{"from":3153.46,"to":3155.59,"location":2,"content":"little bit ambivalent but it sort of"},{"from":3155.59,"to":3157.45,"location":2,"content":"seems that they sort of go a little bit"},{"from":3157.45,"to":3163.29,"location":2,"content":"in the direction of what we want yes"},{"from":3163.29,"to":3167.05,"location":2,"content":"this ground truth was we actually asked"},{"from":3167.05,"to":3169.38,"location":2,"content":"a whole bunch of human beings to say"},{"from":3169.38,"to":3173.68,"location":2,"content":"race the meaning of not sad on this"},{"from":3173.68,"to":3176.02,"location":2,"content":"scale of 1 to 10 maybe this wasn't a"},{"from":3176.02,"to":3177.91,"location":2,"content":"very good clear task because as you can"},{"from":3177.91,"to":3179.95,"location":2,"content":"see it bounced around a lot"},{"from":3179.95,"to":3184.18,"location":2,"content":"uh-huh what kind of ratings we were"},{"from":3184.18,"to":3186.28,"location":2,"content":"getting for things but yeah that was"},{"from":3186.28,"to":3190.02,"location":2,"content":"actually going to gain human judgments"},{"from":3190.02,"to":3195.16,"location":2,"content":"we also then used this model to say well"},{"from":3195.16,"to":3198.19,"location":2,"content":"could we do semantic classification"},{"from":3198.19,"to":3201.03,"location":2,"content":"tasks so if we wanted to understand"},{"from":3201.03,"to":3204.18,"location":2,"content":"relations between different noun phrases"},{"from":3204.18,"to":3208.96,"location":2,"content":"so this was a data set where there were"},{"from":3208.96,"to":3211.63,"location":2,"content":"relations marked between two noun"},{"from":3211.63,"to":3214,"location":2,"content":"phrases my apartment has a"},{"from":3214,"to":3216.67,"location":2,"content":"pretty large kitchen that was seen as an"},{"from":3216.67,"to":3220.45,"location":2,"content":"example of a component whole a a part of"},{"from":3220.45,"to":3222.67,"location":2,"content":"relationship between the two noun"},{"from":3222.67,"to":3224.56,"location":2,"content":"phrases and there were other"},{"from":3224.56,"to":3227.65,"location":2,"content":"relationships between different kinds of"},{"from":3227.65,"to":3230.56,"location":2,"content":"noun phrases so for us the movie showed"},{"from":3230.56,"to":3234.19,"location":2,"content":"wars that that was then a message topic"},{"from":3234.19,"to":3236.5,"location":2,"content":"so there's some communication medium"},{"from":3236.5,"to":3239.11,"location":2,"content":"that contains some topic relationship"},{"from":3239.11,"to":3242.2,"location":2,"content":"and so we were using this kind of neural"},{"from":3242.2,"to":3244.14,"location":2,"content":"network to sort of build our meaning"},{"from":3244.14,"to":3246.58,"location":2,"content":"representations and then putting them"},{"from":3246.58,"to":3249.01,"location":2,"content":"through another neural network layer as"},{"from":3249.01,"to":3252.31,"location":2,"content":"a classifier to see how well we did and"},{"from":3252.31,"to":3254.53,"location":2,"content":"so we got some sort of fairly good"},{"from":3254.53,"to":3256.75,"location":2,"content":"results on that so this was a data set"},{"from":3256.75,"to":3258.88,"location":2,"content":"that people had worked on with"},{"from":3258.88,"to":3261.97,"location":2,"content":"traditional NLP systems of different"},{"from":3261.97,"to":3264.46,"location":2,"content":"kinds of machine learning methods but in"},{"from":3264.46,"to":3266.74,"location":2,"content":"some sense you know what we're"},{"from":3266.74,"to":3268.6,"location":2,"content":"interested in was we seem to be making"},{"from":3268.6,"to":3270.61,"location":2,"content":"progress in having a better semantic"},{"from":3270.61,"to":3272.88,"location":2,"content":"composition system that our old"},{"from":3272.88,"to":3275.38,"location":2,"content":"recursive neural network was getting"},{"from":3275.38,"to":3278.38,"location":2,"content":"about 75% and then our new one was"},{"from":3278.38,"to":3281.08,"location":2,"content":"getting about 79% which we could sort of"},{"from":3281.08,"to":3283.48,"location":2,"content":"push up further by putting more features"},{"from":3283.48,"to":3288.73,"location":2,"content":"into our system so that was progress but"},{"from":3288.73,"to":3290.59,"location":2,"content":"we didn't stop there"},{"from":3290.59,"to":3292.87,"location":2,"content":"and we kept on trying to come up with"},{"from":3292.87,"to":3296.56,"location":2,"content":"better ways of doing things and so even"},{"from":3296.56,"to":3299.74,"location":2,"content":"though things worked fairly well here it"},{"from":3299.74,"to":3302.94,"location":2,"content":"sort of seemed like this way of doing"},{"from":3302.94,"to":3307.63,"location":2,"content":"matrices wasn't necessarily very good it"},{"from":3307.63,"to":3311.17,"location":2,"content":"sort of had two problems one problem was"},{"from":3311.17,"to":3314.68,"location":2,"content":"it introduced a humongous number of"},{"from":3314.68,"to":3316.72,"location":2,"content":"parameters because you know for just"},{"from":3316.72,"to":3318.06,"location":2,"content":"about everything that we've done"},{"from":3318.06,"to":3321.7,"location":2,"content":"otherwise words have had a vector and"},{"from":3321.7,"to":3324.73,"location":2,"content":"well maybe sometimes we use quite high"},{"from":3324.73,"to":3326.77,"location":2,"content":"dimensional vectors like a thousand and"},{"from":3326.77,"to":3329.5,"location":2,"content":"twenty four but you know that's a"},{"from":3329.5,"to":3331.54,"location":2,"content":"relatively modest number of parameters"},{"from":3331.54,"to":3334.45,"location":2,"content":"whereas once we introduced this matrix"},{"from":3334.45,"to":3337.26,"location":2,"content":"here we've got that number of squared"},{"from":3337.26,"to":3340.95,"location":2,"content":"additional parameters for every word and"},{"from":3340.95,"to":3343.24,"location":2,"content":"essentially because of that number of"},{"from":3343.24,"to":3345.52,"location":2,"content":"parameters to be able to compute this"},{"from":3345.52,"to":3347.82,"location":2,"content":"model at all we were making the vet"},{"from":3347.82,"to":3349.89,"location":2,"content":"the size small so what we're actually"},{"from":3349.89,"to":3351.9,"location":2,"content":"using was of these were just 25"},{"from":3351.9,"to":3354.9,"location":2,"content":"dimensional vectors so that the 25"},{"from":3354.9,"to":3358.92,"location":2,"content":"squared 625 still saved sort of decently"},{"from":3358.92,"to":3360.63,"location":2,"content":"within the range in which we could"},{"from":3360.63,"to":3363.21,"location":2,"content":"compute so that was the first problem"},{"from":3363.21,"to":3366.12,"location":2,"content":"the second problem is we didn't really"},{"from":3366.12,"to":3369.87,"location":2,"content":"have very good ways I'm sort of building"},{"from":3369.87,"to":3372.72,"location":2,"content":"up the matrix meaning of bigger phrases"},{"from":3372.72,"to":3375.03,"location":2,"content":"I mean you know this sort of seems"},{"from":3375.03,"to":3377.34,"location":2,"content":"something simple we could do but it"},{"from":3377.34,"to":3379.59,"location":2,"content":"didn't you know feel a very good way of"},{"from":3379.59,"to":3382.2,"location":2,"content":"getting a matrix meaning of a phrase so"},{"from":3382.2,"to":3383.67,"location":2,"content":"he sort of wanted to come up with some"},{"from":3383.67,"to":3385.35,"location":2,"content":"other way of doing things"},{"from":3385.35,"to":3387.65,"location":2,"content":"that could fix both of those problems"},{"from":3387.65,"to":3390.92,"location":2,"content":"and then that led in to work on"},{"from":3390.92,"to":3394.29,"location":2,"content":"recursive neural tenser networks and"},{"from":3394.29,"to":3396.84,"location":2,"content":"there's a kind of a nice idea here of"},{"from":3396.84,"to":3400.29,"location":2,"content":"these neural tensors which is an idea"},{"from":3400.29,"to":3401.85,"location":2,"content":"that's actually being used in other"},{"from":3401.85,"to":3406.41,"location":2,"content":"places including work on sort of putting"},{"from":3406.41,"to":3408.54,"location":2,"content":"vector embeddings of knowledge graphs"},{"from":3408.54,"to":3410.7,"location":2,"content":"and so on which is a kind of a bit of a"},{"from":3410.7,"to":3413.58,"location":2,"content":"nice idea so I wanted to sort of show a"},{"from":3413.58,"to":3417.15,"location":2,"content":"bit of how this model works and but just"},{"from":3417.15,"to":3420.03,"location":2,"content":"to say first a place where we applied"},{"from":3420.03,"to":3422.16,"location":2,"content":"this model was on the problem of"},{"from":3422.16,"to":3424.77,"location":2,"content":"sentiment analysis now I think the term"},{"from":3424.77,"to":3426.78,"location":2,"content":"sentiment analysis has come up a few"},{"from":3426.78,"to":3429.36,"location":2,"content":"times as something you can do and"},{"from":3429.36,"to":3432.44,"location":2,"content":"actually Richard and mention the last"},{"from":3432.44,"to":3435.21,"location":2,"content":"lecture but I think we've never really"},{"from":3435.21,"to":3438.21,"location":2,"content":"talked for five minutes in this class on"},{"from":3438.21,"to":3440.97,"location":2,"content":"sentiment analysis so I'll give you this"},{"from":3440.97,"to":3444.09,"location":2,"content":"as an example of that sentiment analysis"},{"from":3444.09,"to":3446.22,"location":2,"content":"has actually been a really common and"},{"from":3446.22,"to":3449.34,"location":2,"content":"important application in natural"},{"from":3449.34,"to":3452.13,"location":2,"content":"language processing you're looking at a"},{"from":3452.13,"to":3454.47,"location":2,"content":"piece of text and you're sort of saying"},{"from":3454.47,"to":3459.15,"location":2,"content":"is it positive or negative and that's"},{"from":3459.15,"to":3460.86,"location":2,"content":"just something that's very useful for"},{"from":3460.86,"to":3463.68,"location":2,"content":"lots of commercial applications of"},{"from":3463.68,"to":3466.05,"location":2,"content":"looking at product reviews or doing"},{"from":3466.05,"to":3468.96,"location":2,"content":"brand awareness and things like that of"},{"from":3468.96,"to":3470.58,"location":2,"content":"sort of looking at sentiment connected"},{"from":3470.58,"to":3473.1,"location":2,"content":"to things and to some extent doing"},{"from":3473.1,"to":3475.68,"location":2,"content":"sentiment analysis is easy right that"},{"from":3475.68,"to":3478.11,"location":2,"content":"you can kind of say well look at a piece"},{"from":3478.11,"to":3480.54,"location":2,"content":"of text if you see words like loved"},{"from":3480.54,"to":3481.66,"location":2,"content":"great impressed"},{"from":3481.66,"to":3484.42,"location":2,"content":"and it's positive it's a positive review"},{"from":3484.42,"to":3486.97,"location":2,"content":"and if it's saying bad and awful then"},{"from":3486.97,"to":3489.07,"location":2,"content":"it's a negative review and to some"},{"from":3489.07,"to":3491.2,"location":2,"content":"extent that's the baseline of sentiment"},{"from":3491.2,"to":3494.5,"location":2,"content":"analysis that you can use just either"},{"from":3494.5,"to":3497.59,"location":2,"content":"selected word features or all words in a"},{"from":3497.59,"to":3500.17,"location":2,"content":"bag of words and if you do that you"},{"from":3500.17,"to":3503.46,"location":2,"content":"don't actually do that badly in"},{"from":3503.46,"to":3505.78,"location":2,"content":"sentiment analysis if you have longer"},{"from":3505.78,"to":3508,"location":2,"content":"documents just looking at bags of words"},{"from":3508,"to":3510.4,"location":2,"content":"can give you 90% in sentiment analysis"},{"from":3510.4,"to":3513.4,"location":2,"content":"but on the other hand things often do"},{"from":3513.4,"to":3516.7,"location":2,"content":"get trickier right so this is from"},{"from":3516.7,"to":3519.25,"location":2,"content":"Rotten Tomatoes with this cast and this"},{"from":3519.25,"to":3521.26,"location":2,"content":"subject matter the movie should have"},{"from":3521.26,"to":3523.39,"location":2,"content":"been funnier and more entertaining and"},{"from":3523.39,"to":3525.79,"location":2,"content":"if you sort of pretend you're a bag of"},{"from":3525.79,"to":3529.03,"location":2,"content":"words model the only words in this that"},{"from":3529.03,"to":3531.55,"location":2,"content":"are sort of clearly sentiment Laden"},{"from":3531.55,"to":3535.99,"location":2,"content":"words entertaining and funnier and both"},{"from":3535.99,"to":3538.99,"location":2,"content":"of those are pretty positive words but"},{"from":3538.99,"to":3541.27,"location":2,"content":"it's fairly obvious that this actually"},{"from":3541.27,"to":3543.73,"location":2,"content":"isn't meant to be a bad review of the"},{"from":3543.73,"to":3546.64,"location":2,"content":"movie and so well how are we meant to"},{"from":3546.64,"to":3548.68,"location":2,"content":"know that well it sort of seems again"},{"from":3548.68,"to":3550.3,"location":2,"content":"like what we have to do as meaning"},{"from":3550.3,"to":3553.26,"location":2,"content":"composition we have to get sort of"},{"from":3553.26,"to":3555.73,"location":2,"content":"phrases like should have been funnier"},{"from":3555.73,"to":3558.46,"location":2,"content":"and then realized that that's actually a"},{"from":3558.46,"to":3561.61,"location":2,"content":"negative meaning for a phrase and so we"},{"from":3561.61,"to":3564.07,"location":2,"content":"wanted to explore how we could look at"},{"from":3564.07,"to":3567.58,"location":2,"content":"those sort of meanings for phrases and"},{"from":3567.58,"to":3569.92,"location":2,"content":"explore building up those meanings as"},{"from":3569.92,"to":3573.97,"location":2,"content":"doing meaning composition over trees um"},{"from":3573.97,"to":3577.51,"location":2,"content":"so the first thing we did was we built a"},{"from":3577.51,"to":3580.69,"location":2,"content":"tree Bank of sentiment trees where we"},{"from":3580.69,"to":3583.03,"location":2,"content":"got people to rate sentiment and so this"},{"from":3583.03,"to":3585.61,"location":2,"content":"led to the Stanford sentiment tree Bank"},{"from":3585.61,"to":3588.19,"location":2,"content":"which is still a data set you often see"},{"from":3588.19,"to":3592.48,"location":2,"content":"used in various of evaluations with a"},{"from":3592.48,"to":3594.28,"location":2,"content":"whole bunch of data sets indeed that"},{"from":3594.28,"to":3597.73,"location":2,"content":"showed up in Decorah NLP last week so"},{"from":3597.73,"to":3601.59,"location":2,"content":"what we were doing this was taking"},{"from":3601.59,"to":3603.96,"location":2,"content":"sentences which were Rotten Tomatoes"},{"from":3603.96,"to":3607.36,"location":2,"content":"sentences from movies we were parsing"},{"from":3607.36,"to":3609.85,"location":2,"content":"them to give tree structure and then we"},{"from":3609.85,"to":3613.69,"location":2,"content":"were asking Mechanical Turk URLs to rate"},{"from":3613.69,"to":3614.57,"location":2,"content":"the"},{"from":3614.57,"to":3616.67,"location":2,"content":"different free the different words and"},{"from":3616.67,"to":3619.67,"location":2,"content":"phrases on a sentiment scale of very"},{"from":3619.67,"to":3622.13,"location":2,"content":"positive to very negative so lots of"},{"from":3622.13,"to":3624.35,"location":2,"content":"stuff is white because it's just not"},{"from":3624.35,"to":3626.75,"location":2,"content":"sentiment Laden right there's words that"},{"from":3626.75,"to":3629.09,"location":2,"content":"are there and there's phrases like the"},{"from":3629.09,"to":3631.7,"location":2,"content":"movie and the movie was which don't"},{"from":3631.7,"to":3633.8,"location":2,"content":"really have any sentiment but then you"},{"from":3633.8,"to":3635.9,"location":2,"content":"have pieces of sort of very positive"},{"from":3635.9,"to":3639.11,"location":2,"content":"pieces of tree and negative pieces of"},{"from":3639.11,"to":3641.3,"location":2,"content":"tree that are then shown in the blue and"},{"from":3641.3,"to":3644.21,"location":2,"content":"the red and so typically in sentiment"},{"from":3644.21,"to":3647.24,"location":2,"content":"datasets people have only labeled the"},{"from":3647.24,"to":3650.24,"location":2,"content":"entire sentence to say this is a"},{"from":3650.24,"to":3652.37,"location":2,"content":"positive sentence or a very positive"},{"from":3652.37,"to":3654.35,"location":2,"content":"sentence this is a negative sentence or"},{"from":3654.35,"to":3656.9,"location":2,"content":"a very negative sentence crucially what"},{"from":3656.9,"to":3659.72,"location":2,"content":"we were doing differently here is every"},{"from":3659.72,"to":3662.42,"location":2,"content":"phrase in the sentence according to our"},{"from":3662.42,"to":3665.03,"location":2,"content":"tree structure was being given a label"},{"from":3665.03,"to":3668.93,"location":2,"content":"for its positivity or negativity and"},{"from":3668.93,"to":3671.6,"location":2,"content":"perhaps not surprisingly just the fact"},{"from":3671.6,"to":3673.19,"location":2,"content":"that you have a lot more annotations"},{"from":3673.19,"to":3677.54,"location":2,"content":"like that just improves the behavior of"},{"from":3677.54,"to":3680.54,"location":2,"content":"classifiers because you kind of can do"},{"from":3680.54,"to":3681.04,"location":2,"content":"better"},{"from":3681.04,"to":3683.9,"location":2,"content":"attribution of which words in a sentence"},{"from":3683.9,"to":3688.7,"location":2,"content":"a positive or negative so these were"},{"from":3688.7,"to":3692.39,"location":2,"content":"were results of sort of preceding models"},{"from":3692.39,"to":3695.98,"location":2,"content":"so the green is a naive Bayes model"},{"from":3695.98,"to":3699.35,"location":2,"content":"except it not only uses individual words"},{"from":3699.35,"to":3702.32,"location":2,"content":"but it uses pairs of words it turns out"},{"from":3702.32,"to":3703.75,"location":2,"content":"if you're building at a traditional"},{"from":3703.75,"to":3706.61,"location":2,"content":"classifier and you want to do sentiment"},{"from":3706.61,"to":3708.32,"location":2,"content":"analysis as opposed to something like"},{"from":3708.32,"to":3711.05,"location":2,"content":"topic classification you get a lot"},{"from":3711.05,"to":3713.51,"location":2,"content":"better results if you also use word pair"},{"from":3713.51,"to":3715.49,"location":2,"content":"features and that's because it does a"},{"from":3715.49,"to":3719,"location":2,"content":"baby bit of composition for you you"},{"from":3719,"to":3720.92,"location":2,"content":"don't only have features for not an"},{"from":3720.92,"to":3722.63,"location":2,"content":"interesting but you can have a feature"},{"from":3722.63,"to":3725.39,"location":2,"content":"for not interesting and that lets you"},{"from":3725.39,"to":3727.91,"location":2,"content":"model a certain amount of stuff and then"},{"from":3727.91,"to":3729.74,"location":2,"content":"these are our older generations of"},{"from":3729.74,"to":3732.05,"location":2,"content":"neural networks our original tree"},{"from":3732.05,"to":3733.91,"location":2,"content":"structured neural network and our matrix"},{"from":3733.91,"to":3737.54,"location":2,"content":"vector one and so simply having for"},{"from":3737.54,"to":3739.79,"location":2,"content":"these so fixed models simply having the"},{"from":3739.79,"to":3742.64,"location":2,"content":"richer supervision that comes from our"},{"from":3742.64,"to":3744.92,"location":2,"content":"new tree Bank it sort of moved up the"},{"from":3744.92,"to":3747.31,"location":2,"content":"performance of every model so even"},{"from":3747.31,"to":3750.61,"location":2,"content":"for just the naivebayes models"},{"from":3750.61,"to":3754.03,"location":2,"content":"performances going up about 4% because"},{"from":3754.03,"to":3757.6,"location":2,"content":"of the fact that it now knows more about"},{"from":3757.6,"to":3760.33,"location":2,"content":"which particular words are positive or"},{"from":3760.33,"to":3764.08,"location":2,"content":"negative in the sentences but still none"},{"from":3764.08,"to":3766.65,"location":2,"content":"of these performances are really great"},{"from":3766.65,"to":3769.84,"location":2,"content":"so we still thought that well can we"},{"from":3769.84,"to":3773.68,"location":2,"content":"build better models of how to do this in"},{"from":3773.68,"to":3776.32,"location":2,"content":"particular if you look at sentences with"},{"from":3776.32,"to":3779.23,"location":2,"content":"sort of various kinds of negation you"},{"from":3779.23,"to":3780.67,"location":2,"content":"know things like should have been"},{"from":3780.67,"to":3783.22,"location":2,"content":"funnier these models in general still"},{"from":3783.22,"to":3785.47,"location":2,"content":"couldn't capture the right meanings for"},{"from":3785.47,"to":3788.38,"location":2,"content":"them and so that led into our fourth"},{"from":3788.38,"to":3792.04,"location":2,"content":"model of how to do this which is this"},{"from":3792.04,"to":3795.48,"location":2,"content":"idea of recursive neural tenser networks"},{"from":3795.48,"to":3799.3,"location":2,"content":"and so what we wanted to be able to do"},{"from":3799.3,"to":3803.7,"location":2,"content":"is go back to just having meanings of"},{"from":3803.7,"to":3807.4,"location":2,"content":"words be vectors but nevertheless"},{"from":3807.4,"to":3810.58,"location":2,"content":"despite that to be able to have a"},{"from":3810.58,"to":3812.65,"location":2,"content":"meaning for a phrase where the two"},{"from":3812.65,"to":3817.63,"location":2,"content":"vectors acted on each other and well you"},{"from":3817.63,"to":3820.36,"location":2,"content":"know this kind of this is the picture of"},{"from":3820.36,"to":3822.85,"location":2,"content":"what we did when we were doing attention"},{"from":3822.85,"to":3825.34,"location":2,"content":"in a bilinear way right we had vectors"},{"from":3825.34,"to":3827.89,"location":2,"content":"for two words we stuck a matrix in"},{"from":3827.89,"to":3831.79,"location":2,"content":"between and we use that and gave an"},{"from":3831.79,"to":3834.37,"location":2,"content":"attention and got an attention score out"},{"from":3834.37,"to":3838.12,"location":2,"content":"so that let these two vectors interact"},{"from":3838.12,"to":3840.82,"location":2,"content":"with each other but it only produced one"},{"from":3840.82,"to":3843.31,"location":2,"content":"number as the output but there's a way"},{"from":3843.31,"to":3846.73,"location":2,"content":"to fix that which is to say well rather"},{"from":3846.73,"to":3851.02,"location":2,"content":"than having matrix here what we could"},{"from":3851.02,"to":3853.86,"location":2,"content":"stick here is a three dimensional cube"},{"from":3853.86,"to":3856.15,"location":2,"content":"which physicists and deep learning"},{"from":3856.15,"to":3859.27,"location":2,"content":"people caught now call a tensor right so"},{"from":3859.27,"to":3860.47,"location":2,"content":"a tensor is just higher"},{"from":3860.47,"to":3863.08,"location":2,"content":"multi-dimensional array in computer"},{"from":3863.08,"to":3867.19,"location":2,"content":"science terms so if we sort of made that"},{"from":3867.19,"to":3870.43,"location":2,"content":"a tensor you know it's like we have sort"},{"from":3870.43,"to":3873.01,"location":2,"content":"of multiple layers and matrix here and"},{"from":3873.01,"to":3876.22,"location":2,"content":"so the end result of that is we get one"},{"from":3876.22,"to":3878.7,"location":2,"content":"number here and one number here"},{"from":3878.7,"to":3881.88,"location":2,"content":"so in total we get out of size two"},{"from":3881.88,"to":3885.24,"location":2,"content":"vector which is all we need and my baby"},{"from":3885.24,"to":3887.82,"location":2,"content":"example where we baby examples where we"},{"from":3887.82,"to":3889.41,"location":2,"content":"only have these two component vectors"},{"from":3889.41,"to":3891.21,"location":2,"content":"for words but in general we have a"},{"from":3891.21,"to":3894.06,"location":2,"content":"tensor with the extra mention dimension"},{"from":3894.06,"to":3896.4,"location":2,"content":"of the size of our word vector and so"},{"from":3896.4,"to":3898.32,"location":2,"content":"therefore we'll get a word vector at"},{"from":3898.32,"to":3900.69,"location":2,"content":"work we'll get a phrase vector out from"},{"from":3900.69,"to":3903.42,"location":2,"content":"the composition that's the same size of"},{"from":3903.42,"to":3906.5,"location":2,"content":"the input vectors and will allow them to"},{"from":3906.5,"to":3909.69,"location":2,"content":"interact with each other in working out"},{"from":3909.69,"to":3916.28,"location":2,"content":"the meaning of the entire thing okay"},{"from":3916.28,"to":3920.55,"location":2,"content":"right at that point we use the resulting"},{"from":3920.55,"to":3927.78,"location":2,"content":"vectors so we had our new tensor network"},{"from":3927.78,"to":3930.09,"location":2,"content":"we actually combined it together with"},{"from":3930.09,"to":3932.46,"location":2,"content":"this sort of previous kind of layer we"},{"from":3932.46,"to":3936.15,"location":2,"content":"used to have our first R and then maybe"},{"from":3936.15,"to":3937.5,"location":2,"content":"you didn't need to do this but we just"},{"from":3937.5,"to":3939.99,"location":2,"content":"decided we're betting as well put things"},{"from":3939.99,"to":3942.18,"location":2,"content":"through a non-linearity and that was"},{"from":3942.18,"to":3944.76,"location":2,"content":"then giving us our new representation of"},{"from":3944.76,"to":3947.61,"location":2,"content":"Fraser's we built that up the tree and"},{"from":3947.61,"to":3950.34,"location":2,"content":"then at the end we could classify the"},{"from":3950.34,"to":3954.27,"location":2,"content":"meaning of any phrase in the same kind"},{"from":3954.27,"to":3956.7,"location":2,"content":"of way with softmax regression and we"},{"from":3956.7,"to":3958.41,"location":2,"content":"could train these weights with gradient"},{"from":3958.41,"to":3961.17,"location":2,"content":"descent to predict sentiment and so this"},{"from":3961.17,"to":3964.29,"location":2,"content":"actually worked pretty nicely I'm in"},{"from":3964.29,"to":3966.57,"location":2,"content":"particular it didn't show really work"},{"from":3966.57,"to":3969.63,"location":2,"content":"any better with just the sentence labels"},{"from":3969.63,"to":3972.72,"location":2,"content":"but if we train the model with our tree"},{"from":3972.72,"to":3974.75,"location":2,"content":"bank we could then get a kind of"},{"from":3974.75,"to":3977.22,"location":2,"content":"whatever that is about another couple of"},{"from":3977.22,"to":3979.14,"location":2,"content":"percent and performance and so that"},{"from":3979.14,"to":3981.9,"location":2,"content":"seemed good and so in particular it"},{"from":3981.9,"to":3984,"location":2,"content":"seemed to do a much better job of"},{"from":3984,"to":3985.62,"location":2,"content":"actually understanding meaning"},{"from":3985.62,"to":3988.17,"location":2,"content":"composition so here's the kind of"},{"from":3988.17,"to":3990.93,"location":2,"content":"sentence where you have there a slow and"},{"from":3990.93,"to":3993,"location":2,"content":"repetitive parts but it has just enough"},{"from":3993,"to":3995.52,"location":2,"content":"spice to keep it interesting and the"},{"from":3995.52,"to":3997.47,"location":2,"content":"models seem you know pretty good at"},{"from":3997.47,"to":3999.72,"location":2,"content":"understanding ok this part of the"},{"from":3999.72,"to":4002.36,"location":2,"content":"sentence is negative this part of the"},{"from":4002.36,"to":4004.46,"location":2,"content":"sentence is positive and actually when"},{"from":4004.46,"to":4006.5,"location":2,"content":"you stick the two halves together the"},{"from":4006.5,"to":4008.54,"location":2,"content":"end result is the sentence there's"},{"from":4008.54,"to":4010.53,"location":2,"content":"positive and meaning"},{"from":4010.53,"to":4013.65,"location":2,"content":"but focusing in a little bit more what"},{"from":4013.65,"to":4016.92,"location":2,"content":"seems like it's especially good was for"},{"from":4016.92,"to":4019.14,"location":2,"content":"the first time this actually did seem"},{"from":4019.14,"to":4022.86,"location":2,"content":"like it could do a better job of working"},{"from":4022.86,"to":4025.53,"location":2,"content":"out sort of what happens when you do"},{"from":4025.53,"to":4028.17,"location":2,"content":"things like negation so here we have"},{"from":4028.17,"to":4030.51,"location":2,"content":"it's just incredibly dull and it's"},{"from":4030.51,"to":4032.55,"location":2,"content":"definitely not dull so if it's"},{"from":4032.55,"to":4034.86,"location":2,"content":"definitely not dull that actually means"},{"from":4034.86,"to":4037.86,"location":2,"content":"it's good right can we work out that the"},{"from":4037.86,"to":4041.24,"location":2,"content":"meaning of it's definitely not dull and"},{"from":4041.24,"to":4046.8,"location":2,"content":"so these this is sort of showing a sort"},{"from":4046.8,"to":4050.94,"location":2,"content":"of what happens when you have a negative"},{"from":4050.94,"to":4053.61,"location":2,"content":"and negative sentence that's further"},{"from":4053.61,"to":4061.41,"location":2,"content":"negated so if you go from so if you sort"},{"from":4061.41,"to":4064.77,"location":2,"content":"of do and the in X and the Gatien of a"},{"from":4064.77,"to":4067.53,"location":2,"content":"negative things should become moderately"},{"from":4067.53,"to":4071.25,"location":2,"content":"positive right so that if you have dull"},{"from":4071.25,"to":4074.19,"location":2,"content":"is negative and if you say not dull it"},{"from":4074.19,"to":4076.53,"location":2,"content":"doesn't mean it's fantastic but it means"},{"from":4076.53,"to":4079.2,"location":2,"content":"it's moderately positive and so for"},{"from":4079.2,"to":4082.38,"location":2,"content":"either a kind of naive Bayes model or"},{"from":4082.38,"to":4084.54,"location":2,"content":"our preceding models they weren't"},{"from":4084.54,"to":4086.61,"location":2,"content":"capable of capturing that of sort of"},{"from":4086.61,"to":4090.08,"location":2,"content":"going from dull to not dull your your"},{"from":4090.08,"to":4092.91,"location":2,"content":"meaning computation did not come out any"},{"from":4092.91,"to":4095.49,"location":2,"content":"more positive whereas this sort of neuro"},{"from":4095.49,"to":4098.31,"location":2,"content":"tenser network was capturing the fact"},{"from":4098.31,"to":4100.92,"location":2,"content":"that not dull meant it was reasonably"},{"from":4100.92,"to":4107.31,"location":2,"content":"good so that was progress yeah so I"},{"from":4107.31,"to":4109.83,"location":2,"content":"think that's as much as I'll show you"},{"from":4109.83,"to":4111.84,"location":2,"content":"really now about applying these tree"},{"from":4111.84,"to":4115.68,"location":2,"content":"structured neural networks to natural"},{"from":4115.68,"to":4120.3,"location":2,"content":"language you know I think the summary I"},{"from":4120.3,"to":4123.06,"location":2,"content":"sort of said at the beginning is that I"},{"from":4123.06,"to":4124.35,"location":2,"content":"think you know they're kind of"},{"from":4124.35,"to":4126.51,"location":2,"content":"interesting ideas and linguistic"},{"from":4126.51,"to":4131.16,"location":2,"content":"connections here I mean for various"},{"from":4131.16,"to":4132.18,"location":2,"content":"reasons"},{"from":4132.18,"to":4136.44,"location":2,"content":"these ideas haven't been pursued a ton"},{"from":4136.44,"to":4138.45,"location":2,"content":"in recent years of natural language"},{"from":4138.45,"to":4141.69,"location":2,"content":"processing you know one is in all"},{"from":4141.69,"to":4144.19,"location":2,"content":"honesty people have found that"},{"from":4144.19,"to":4147.61,"location":2,"content":"once you have high dimensional vectors"},{"from":4147.61,"to":4149.95,"location":2,"content":"in things like the kind of sequence"},{"from":4149.95,"to":4151.81,"location":2,"content":"models that we've looked at whether it's"},{"from":4151.81,"to":4153.67,"location":2,"content":"meaning things like the sort of lsdm"},{"from":4153.67,"to":4156.4,"location":2,"content":"models or any of the more recent"},{"from":4156.4,"to":4158.92,"location":2,"content":"contextual language models those work"},{"from":4158.92,"to":4162.07,"location":2,"content":"incredibly well and it's not it's not"},{"from":4162.07,"to":4163.72,"location":2,"content":"clear that over all these models work"},{"from":4163.72,"to":4166.6,"location":2,"content":"better the second reason is sort of a"},{"from":4166.6,"to":4171.91,"location":2,"content":"computational reason which is GPUs work"},{"from":4171.91,"to":4174.13,"location":2,"content":"great when you're doing uniform"},{"from":4174.13,"to":4176.74,"location":2,"content":"computation and the beauty of having"},{"from":4176.74,"to":4178.66,"location":2,"content":"something like a sequence model is that"},{"from":4178.66,"to":4182.02,"location":2,"content":"there's there's just one determinate"},{"from":4182.02,"to":4184,"location":2,"content":"computation you're doing along the"},{"from":4184,"to":4186.55,"location":2,"content":"sequence or in the convolutional neural"},{"from":4186.55,"to":4188.91,"location":2,"content":"network there's one determinate"},{"from":4188.91,"to":4191.83,"location":2,"content":"computation you're doing up through your"},{"from":4191.83,"to":4194.08,"location":2,"content":"convolutional layers and therefore"},{"from":4194.08,"to":4196.96,"location":2,"content":"things can be represented and computed"},{"from":4196.96,"to":4199.72,"location":2,"content":"efficiently on a GPU the huge problem"},{"from":4199.72,"to":4201.88,"location":2,"content":"with these kind of models was what"},{"from":4201.88,"to":4203.59,"location":2,"content":"computations you were going to do"},{"from":4203.59,"to":4205.75,"location":2,"content":"depended on which structure you're"},{"from":4205.75,"to":4208.03,"location":2,"content":"assigning to the sentence and every"},{"from":4208.03,"to":4210.04,"location":2,"content":"sentence was going to have a different"},{"from":4210.04,"to":4212.35,"location":2,"content":"structure and so therefore there was no"},{"from":4212.35,"to":4214.75,"location":2,"content":"way to batch the computations over a"},{"from":4214.75,"to":4217.27,"location":2,"content":"group of sentences and have the same"},{"from":4217.27,"to":4219.19,"location":2,"content":"computations being done for different"},{"from":4219.19,"to":4221.74,"location":2,"content":"sentences which sort of undermined the"},{"from":4221.74,"to":4223.57,"location":2,"content":"ability to sort of efficiently build"},{"from":4223.57,"to":4227.32,"location":2,"content":"these models in the large the thing I"},{"from":4227.32,"to":4229.51,"location":2,"content":"thought I'd just sort of say a moment"},{"from":4229.51,"to":4232.99,"location":2,"content":"about at the end the funny thing is that"},{"from":4232.99,"to":4235.36,"location":2,"content":"although these haven't been used much"},{"from":4235.36,"to":4238.98,"location":2,"content":"for language in the last few years that"},{"from":4238.98,"to":4242.2,"location":2,"content":"they've actually had some use and found"},{"from":4242.2,"to":4244.45,"location":2,"content":"different applications in different"},{"from":4244.45,"to":4247.39,"location":2,"content":"places which is just sort of seen kind"},{"from":4247.39,"to":4249.91,"location":2,"content":"of cute so this is actually an"},{"from":4249.91,"to":4253.96,"location":2,"content":"application from physics and I think I'm"},{"from":4253.96,"to":4256.21,"location":2,"content":"going to just have to read this since I"},{"from":4256.21,"to":4258.45,"location":2,"content":"have no idea what half the words mean"},{"from":4258.45,"to":4262.06,"location":2,"content":"but what it says is by far the most"},{"from":4262.06,"to":4264.25,"location":2,"content":"common structures seen in collisions at"},{"from":4264.25,"to":4266.77,"location":2,"content":"the Large Hadron Collider a collimated"},{"from":4266.77,"to":4269.32,"location":2,"content":"sprays of energetic hydrants referred to"},{"from":4269.32,"to":4271.78,"location":2,"content":"as Jets these Jets are produced from the"},{"from":4271.78,"to":4273.85,"location":2,"content":"fragmentation and hadron ization of"},{"from":4273.85,"to":4276.52,"location":2,"content":"quarks and gluons as described by"},{"from":4276.52,"to":4277.6,"location":2,"content":"quantum chromo"},{"from":4277.6,"to":4281.59,"location":2,"content":"dynamics anyone knows what that means I"},{"from":4281.59,"to":4284.35,"location":2,"content":"hope you're following along here one"},{"from":4284.35,"to":4286.42,"location":2,"content":"compelling physics challenge is to"},{"from":4286.42,"to":4288.79,"location":2,"content":"search for highly boosted standard model"},{"from":4288.79,"to":4291.96,"location":2,"content":"particles decaying hydraulically"},{"from":4291.96,"to":4294.01,"location":2,"content":"unfortunately there's a large background"},{"from":4294.01,"to":4297.82,"location":2,"content":"from jets produced by more mundane QCD"},{"from":4297.82,"to":4299.71,"location":2,"content":"that's quantum chromodynamics that"},{"from":4299.71,"to":4302.62,"location":2,"content":"processes in this work we propose"},{"from":4302.62,"to":4304.89,"location":2,"content":"instead a solution for jet"},{"from":4304.89,"to":4307.36,"location":2,"content":"classification based on an analogy"},{"from":4307.36,"to":4310.18,"location":2,"content":"between quantum chromodynamics and"},{"from":4310.18,"to":4312.88,"location":2,"content":"natural languages as inspired by several"},{"from":4312.88,"to":4316.11,"location":2,"content":"works from natural language processing"},{"from":4316.11,"to":4319.03,"location":2,"content":"much like a sentence is composed of"},{"from":4319.03,"to":4321.43,"location":2,"content":"words for syntactic structure organized"},{"from":4321.43,"to":4323.77,"location":2,"content":"as the past tree and Jettas are also"},{"from":4323.77,"to":4326.62,"location":2,"content":"composed of four momenta following a"},{"from":4326.62,"to":4328.9,"location":2,"content":"structure dictated by key CD and"},{"from":4328.9,"to":4331.03,"location":2,"content":"organized by the clustering history of a"},{"from":4331.03,"to":4334.66,"location":2,"content":"sequential combination jet algorithm so"},{"from":4334.66,"to":4337.45,"location":2,"content":"anyway um yeah with these jets you see"},{"from":4337.45,"to":4339.88,"location":2,"content":"they've got a tree structure over them"},{"from":4339.88,"to":4342.73,"location":2,"content":"and they're using a tree recursive"},{"from":4342.73,"to":4346.78,"location":2,"content":"neural network to model it well that's a"},{"from":4346.78,"to":4348.76,"location":2,"content":"little bit far afield but to show you"},{"from":4348.76,"to":4353.05,"location":2,"content":"just one more example that another place"},{"from":4353.05,"to":4354.76,"location":2,"content":"where these models have actually been"},{"from":4354.76,"to":4358.03,"location":2,"content":"quite useful is for doing things in"},{"from":4358.03,"to":4360.91,"location":2,"content":"programming languages and I think in"},{"from":4360.91,"to":4363.79,"location":2,"content":"part this is because the application is"},{"from":4363.79,"to":4366.79,"location":2,"content":"easier in programming languages so"},{"from":4366.79,"to":4369.22,"location":2,"content":"unlike a natural language where we have"},{"from":4369.22,"to":4371.23,"location":2,"content":"this uncertainty as to what is the"},{"from":4371.23,"to":4373.6,"location":2,"content":"correct parse tree because there's a lot"},{"from":4373.6,"to":4376.02,"location":2,"content":"of ambiguity in natural language in"},{"from":4376.02,"to":4379.33,"location":2,"content":"programming languages the parse trees"},{"from":4379.33,"to":4381.67,"location":2,"content":"are actually pretty determinate so a"},{"from":4381.67,"to":4384.43,"location":2,"content":"group of people at Berkeley dorm song"},{"from":4384.43,"to":4387.49,"location":2,"content":"and her students have worked on doing"},{"from":4387.49,"to":4390.16,"location":2,"content":"programming language translation by"},{"from":4390.16,"to":4392.38,"location":2,"content":"building tree recursive neural network"},{"from":4392.38,"to":4396.31,"location":2,"content":"encoder decoders so that you're building"},{"from":4396.31,"to":4398.04,"location":2,"content":"up a tree structured neural network"},{"from":4398.04,"to":4401.44,"location":2,"content":"representation of a program in one"},{"from":4401.44,"to":4403.66,"location":2,"content":"language this is a CoffeeScript program"},{"from":4403.66,"to":4406.12,"location":2,"content":"and then you're wanting to build a tree"},{"from":4406.12,"to":4408.7,"location":2,"content":"a tree model which is then translating"},{"from":4408.7,"to":4410.74,"location":2,"content":"that to a program in a different"},{"from":4410.74,"to":4411.52,"location":2,"content":"language"},{"from":4411.52,"to":4413.95,"location":2,"content":"and they've been able to do that and get"},{"from":4413.95,"to":4417.88,"location":2,"content":"good results I was too lazy to retype"},{"from":4417.88,"to":4420.82,"location":2,"content":"this table so this is probably a bit bit"},{"from":4420.82,"to":4423.46,"location":2,"content":"hard to read but what's contrasting is"},{"from":4423.46,"to":4426.25,"location":2,"content":"for a number of programs this is a sort"},{"from":4426.25,"to":4430.26,"location":2,"content":"of CoffeeScript to JavaScript"},{"from":4430.26,"to":4433.42,"location":2,"content":"translation they're comparing using tree"},{"from":4433.42,"to":4436.39,"location":2,"content":"to tree models and then using sequence"},{"from":4436.39,"to":4438.25,"location":2,"content":"to sequence models and then they try"},{"from":4438.25,"to":4441.1,"location":2,"content":"both other combinations sequence to tree"},{"from":4441.1,"to":4444.91,"location":2,"content":"and tree to sequence and what they find"},{"from":4444.91,"to":4447.34,"location":2,"content":"is you can get the best results with the"},{"from":4447.34,"to":4450.13,"location":2,"content":"tree to tree neural network models and"},{"from":4450.13,"to":4452.92,"location":2,"content":"in particular these tree to tree models"},{"from":4452.92,"to":4455.59,"location":2,"content":"are augmented with attention so they"},{"from":4455.59,"to":4457.24,"location":2,"content":"have attention like we talked about the"},{"from":4457.24,"to":4459.64,"location":2,"content":"sequence to sequence models where you're"},{"from":4459.64,"to":4461.26,"location":2,"content":"then being able to do attention back to"},{"from":4461.26,"to":4463.75,"location":2,"content":"nodes in the tree structure which is a"},{"from":4463.75,"to":4466.66,"location":2,"content":"pretty natural way of doing translation"},{"from":4466.66,"to":4469.48,"location":2,"content":"and indeed what these results show is if"},{"from":4469.48,"to":4471.31,"location":2,"content":"you don't have I'll sorry these results"},{"from":4471.31,"to":4473.17,"location":2,"content":"show is if you don't have the attention"},{"from":4473.17,"to":4476.2,"location":2,"content":"operation it doesn't work at all it's"},{"from":4476.2,"to":4479.77,"location":2,"content":"too difficult to get things so done if"},{"from":4479.77,"to":4481,"location":2,"content":"you've just sort of trying to create a"},{"from":4481,"to":4482.98,"location":2,"content":"single tree representation then say"},{"from":4482.98,"to":4485.65,"location":2,"content":"generate that the translation from that"},{"from":4485.65,"to":4487.09,"location":2,"content":"but if you can do it with this sort of"},{"from":4487.09,"to":4489.1,"location":2,"content":"putting attention into the different"},{"from":4489.1,"to":4493.57,"location":2,"content":"nodes that's great you might if you know"},{"from":4493.57,"to":4495.7,"location":2,"content":"about what CoffeeScript is you might"},{"from":4495.7,"to":4498.31,"location":2,"content":"feel like wait that's cheating slightly"},{"from":4498.31,"to":4500.32,"location":2,"content":"because copy script is a bit too similar"},{"from":4500.32,"to":4504.07,"location":2,"content":"to JavaScript but they've also done it"},{"from":4504.07,"to":4506.05,"location":2,"content":"in other languages so this is going"},{"from":4506.05,"to":4510.16,"location":2,"content":"between Java and c-sharp and this is a"},{"from":4510.16,"to":4513.13,"location":2,"content":"sort of a handwritten Java to c-sharp"},{"from":4513.13,"to":4515.05,"location":2,"content":"converter that you can download from"},{"from":4515.05,"to":4517.09,"location":2,"content":"github if you want but it doesn't"},{"from":4517.09,"to":4520,"location":2,"content":"actually work that well and they're able"},{"from":4520,"to":4521.83,"location":2,"content":"to show that they're able to build a far"},{"from":4521.83,"to":4527.14,"location":2,"content":"better Java to c-sharp translator doing"},{"from":4527.14,"to":4530.47,"location":2,"content":"that so that's actually kind of cool and"},{"from":4530.47,"to":4531.73,"location":2,"content":"it's good to know the tree structure"},{"from":4531.73,"to":4533.83,"location":2,"content":"recursive neural networks are good for"},{"from":4533.83,"to":4536.74,"location":2,"content":"some things so I'm pleased oh I see work"},{"from":4536.74,"to":4540.28,"location":2,"content":"like this okay I'm just about done but I"},{"from":4540.28,"to":4543.91,"location":2,"content":"thought before finishing I just"},{"from":4543.91,"to":4544.87,"location":2,"content":"mentioned one of"},{"from":4544.87,"to":4547.06,"location":2,"content":"thing which is sort of nothing to do"},{"from":4547.06,"to":4548.25,"location":2,"content":"with natural language processing"},{"from":4548.25,"to":4551.95,"location":2,"content":"precisely but it's about AI but I wanted"},{"from":4551.95,"to":4553.57,"location":2,"content":"to sort of put in a little bit of"},{"from":4553.57,"to":4556.24,"location":2,"content":"advertisement that something that a"},{"from":4556.24,"to":4558.22,"location":2,"content":"number of us have been working on very"},{"from":4558.22,"to":4560.95,"location":2,"content":"hard for the last year or so is"},{"from":4560.95,"to":4564.25,"location":2,"content":"developing a new Stanford Institute for"},{"from":4564.25,"to":4566.2,"location":2,"content":"human centered artificial intelligence"},{"from":4566.2,"to":4568.3,"location":2,"content":"and actually the launch of this"},{"from":4568.3,"to":4570.58,"location":2,"content":"Institute is going to be on Monday of"},{"from":4570.58,"to":4573.07,"location":2,"content":"exam week just when you're maximally"},{"from":4573.07,"to":4575.5,"location":2,"content":"concentrating oh and the things such as"},{"from":4575.5,"to":4579.19,"location":2,"content":"this but our hope is that we can have a"},{"from":4579.19,"to":4582.43,"location":2,"content":"lot of new activity around artificial"},{"from":4582.43,"to":4584.62,"location":2,"content":"intelligence taking a much broader"},{"from":4584.62,"to":4586.86,"location":2,"content":"perspective to artificial intelligence"},{"from":4586.86,"to":4590.05,"location":2,"content":"which is centrally viewing it from the"},{"from":4590.05,"to":4594.84,"location":2,"content":"viewpoint of humans and working out"},{"from":4594.84,"to":4597.22,"location":2,"content":"exploring a large broader range of"},{"from":4597.22,"to":4599.02,"location":2,"content":"issues that embrace a lot of the"},{"from":4599.02,"to":4601.06,"location":2,"content":"interests of the rest of the university"},{"from":4601.06,"to":4602.98,"location":2,"content":"where there's the social sciences and"},{"from":4602.98,"to":4605.2,"location":2,"content":"humanities are also various in"},{"from":4605.2,"to":4607.12,"location":2,"content":"professional schools like the law school"},{"from":4607.12,"to":4609.76,"location":2,"content":"and the business school so to just"},{"from":4609.76,"to":4614.92,"location":2,"content":"quickly say a minute about that that the"},{"from":4614.92,"to":4617.5,"location":2,"content":"the so motivating idea is that sort of"},{"from":4617.5,"to":4620.65,"location":2,"content":"for most of my life sort of AI seem like"},{"from":4620.65,"to":4623.74,"location":2,"content":"a kind of a fun intellectual quest as to"},{"from":4623.74,"to":4625,"location":2,"content":"whether you could write bits of software"},{"from":4625,"to":4628.03,"location":2,"content":"that did anything halfway intelligent"},{"from":4628.03,"to":4629.74,"location":2,"content":"but that's clearly not what's going to"},{"from":4629.74,"to":4632.11,"location":2,"content":"be what's happening for the next 25"},{"from":4632.11,"to":4634.24,"location":2,"content":"years that we're now at this point in"},{"from":4634.24,"to":4636.43,"location":2,"content":"which artificial intelligence systems"},{"from":4636.43,"to":4640.06,"location":2,"content":"are being unleashed on society"},{"from":4640.06,"to":4642.58,"location":2,"content":"and well hopefully they do some good"},{"from":4642.58,"to":4644.53,"location":2,"content":"things but as we've increasingly been"},{"from":4644.53,"to":4646.84,"location":2,"content":"seeing there are lots of also lots of"},{"from":4646.84,"to":4648.82,"location":2,"content":"opportunities for them to do bad things"},{"from":4648.82,"to":4650.74,"location":2,"content":"and even if we're not imagining"},{"from":4650.74,"to":4653.23,"location":2,"content":"Terminator scenarios there are just lots"},{"from":4653.23,"to":4655.84,"location":2,"content":"of places where people are using machine"},{"from":4655.84,"to":4658.66,"location":2,"content":"learning and AI algorithms for making"},{"from":4658.66,"to":4660.7,"location":2,"content":"decisions some of the worst ones or"},{"from":4660.7,"to":4662.53,"location":2,"content":"things like sentencing guidelines and"},{"from":4662.53,"to":4664.45,"location":2,"content":"courts where you have very biased"},{"from":4664.45,"to":4666.64,"location":2,"content":"algorithms making bad decisions and"},{"from":4666.64,"to":4668.8,"location":2,"content":"people are only starting to become a lot"},{"from":4668.8,"to":4671.35,"location":2,"content":"more aware of the issues and so"},{"from":4671.35,"to":4672.88,"location":2,"content":"effectively we're wanting to have this"},{"from":4672.88,"to":4675.28,"location":2,"content":"Institutes of embracing a lot of the"},{"from":4675.28,"to":4677.68,"location":2,"content":"work of social scientists ethicists and"},{"from":4677.68,"to":4678.64,"location":2,"content":"other people"},{"from":4678.64,"to":4680.98,"location":2,"content":"to actually explore how to have an AI"},{"from":4680.98,"to":4683.92,"location":2,"content":"that's really improving human lives"},{"from":4683.92,"to":4685.6,"location":2,"content":"rather than having the opposite effect"},{"from":4685.6,"to":4689.16,"location":2,"content":"and so the three themes that we're"},{"from":4689.16,"to":4692.35,"location":2,"content":"mainly emphasizing for this Institute is"},{"from":4692.35,"to":4695.49,"location":2,"content":"the first one in the top left is"},{"from":4695.49,"to":4698.38,"location":2,"content":"developing AI technologies but we're"},{"from":4698.38,"to":4700.36,"location":2,"content":"particularly interested in making"},{"from":4700.36,"to":4702.88,"location":2,"content":"linkages back to human intelligence so"},{"from":4702.88,"to":4705.79,"location":2,"content":"cognitive science and neuroscience that"},{"from":4705.79,"to":4707.59,"location":2,"content":"when a lot of the early formative work"},{"from":4707.59,"to":4711.49,"location":2,"content":"in AI was done including all of the"},{"from":4711.49,"to":4713.44,"location":2,"content":"early work in neural networks like the"},{"from":4713.44,"to":4715.6,"location":2,"content":"development of backpropagation it was"},{"from":4715.6,"to":4717.37,"location":2,"content":"actually largely done in the context of"},{"from":4717.37,"to":4718.96,"location":2,"content":"cognitive science right and that was a"},{"from":4718.96,"to":4721.3,"location":2,"content":"sort of a linkage that tended to get"},{"from":4721.3,"to":4725.29,"location":2,"content":"lost in the 90s and 2000's statistical"},{"from":4725.29,"to":4727.45,"location":2,"content":"machine learning emphasis and I think"},{"from":4727.45,"to":4730.41,"location":2,"content":"it'd be good to renew that the top right"},{"from":4730.41,"to":4733.78,"location":2,"content":"there's paying much more attention to"},{"from":4733.78,"to":4736.36,"location":2,"content":"the human and societal impact of AI and"},{"from":4736.36,"to":4738.61,"location":2,"content":"so this is looking at legal issues"},{"from":4738.61,"to":4742.56,"location":2,"content":"economic issues labor forces ethics"},{"from":4742.56,"to":4745.42,"location":2,"content":"great power politics whatever you are"},{"from":4745.42,"to":4748.42,"location":2,"content":"but then down the bottom is something"},{"from":4748.42,"to":4750.31,"location":2,"content":"where it seems like this just kind of"},{"from":4750.31,"to":4752.95,"location":2,"content":"enormous opportunities to do more which"},{"from":4752.95,"to":4756.22,"location":2,"content":"is how can we build technology that"},{"from":4756.22,"to":4759.22,"location":2,"content":"actually augments human lives like to"},{"from":4759.22,"to":4762.21,"location":2,"content":"some extent hero tech we've got"},{"from":4762.21,"to":4765.28,"location":2,"content":"technology with AI augmenting human"},{"from":4765.28,"to":4767.53,"location":2,"content":"lives so all of your cell phones have"},{"from":4767.53,"to":4769.78,"location":2,"content":"speech recognition in them now so you"},{"from":4769.78,"to":4773.02,"location":2,"content":"know that's AI that can augment your"},{"from":4773.02,"to":4775.06,"location":2,"content":"human lives but there's a sense of which"},{"from":4775.06,"to":4778.54,"location":2,"content":"not very much of artificial intelligence"},{"from":4778.54,"to":4780.76,"location":2,"content":"has actually been put into the service"},{"from":4780.76,"to":4783.67,"location":2,"content":"of augmenting human lives like most of"},{"from":4783.67,"to":4785.68,"location":2,"content":"what a cell phone has on it is still"},{"from":4785.68,"to":4787.99,"location":2,"content":"sort of clever and cute stuff done by"},{"from":4787.99,"to":4790.54,"location":2,"content":"HCI people and designers which is very"},{"from":4790.54,"to":4792.52,"location":2,"content":"nice a lot of the time when you're using"},{"from":4792.52,"to":4794.71,"location":2,"content":"your map program or something but we"},{"from":4794.71,"to":4797.71,"location":2,"content":"don't really have much AI inside these"},{"from":4797.71,"to":4799.57,"location":2,"content":"devices helping to make people's lives"},{"from":4799.57,"to":4802.72,"location":2,"content":"better and so we're hoping not only for"},{"from":4802.72,"to":4804.94,"location":2,"content":"individuals but applications like health"},{"from":4804.94,"to":4807.85,"location":2,"content":"care to be doing much more sort of"},{"from":4807.85,"to":4810.1,"location":2,"content":"putting artificial intelligence into"},{"from":4810.1,"to":4812.35,"location":2,"content":"human centered applications"},{"from":4812.35,"to":4814.88,"location":2,"content":"anyway that's my brief advertisement I'm"},{"from":4814.88,"to":4816.29,"location":2,"content":"look out for this while you're not"},{"from":4816.29,"to":4818.27,"location":2,"content":"studying for your exams and I think"},{"from":4818.27,"to":4819.14,"location":2,"content":"there'll be sort of lots of"},{"from":4819.14,"to":4821.75,"location":2,"content":"opportunities for students and others to"},{"from":4821.75,"to":4823.34,"location":2,"content":"be getting more involved in this in the"},{"from":4823.34,"to":4826.85,"location":2,"content":"coming months okay thank you very much"},{"from":4826.85,"to":4829.08,"location":2,"content":"and I will see you later"},{"from":4829.08,"to":4833.76,"location":2,"content":"[Applause]"}]}