@@ -37,6 +37,7 @@ Text classification is one of the most basic tasks in natural language processin
For text classification, we provide a non-sequential text classification model based on DNN and CNN. (For LSTM-based model, please refer to PaddleBook [Sentiment Analysis](http://www.paddlepaddle.org/docs/develop/book/06.understand_sentiment/index.html)).
- 4.1 [Sentiment analysis based on DNN / CNN](https://github.com/PaddlePaddle/models/tree/develop/text_classification)
- 4.2 [Text classification model based on Nested sequence](https://github.com/PaddlePaddle/models/tree/develop/nested_sequence/text_classification)
## 5. Learning to rank
...
...
@@ -69,18 +70,65 @@ Sequence-to-sequence model has a wide range of applications. This includes machi
As an example for sequence-to-sequence learning, we take the machine translation task. We demonstrate the sequence-to-sequence mapping model without attention mechanism, which is the basis for all sequence-to-sequence learning models. We will use scheduled sampling to improve the problem of error accumulation in the RNN model, and machine translation with external memory mechanism.
- 8.2 [Improve translation quality using Scheduled Sampling](https://github.com/PaddlePaddle/models/tree/develop/scheduled_sampling)
- 8.3 [Neural machine translation with external memory mechanism](https://github.com/PaddlePaddle/models/tree/develop/mt_with_external_memory)
- 8.4 [Generate chinese poetry](https://github.com/PaddlePaddle/models/tree/develop/generate_chinese_poetry)
## 9. Image classification
## 9. Reading comprehension
When deep learning and various new technologies continue to push forward the field of natural language processing, we cannot help but ask: How should we confirm that the model truly understands human-specific natural language and has a certain ability to understand and reason? Looking at various classic issues in the field of NLP: lexical analysis, syntactic analysis, emotional classification, writing poetry, etc. From the technical principle, the classic solutions of these problems still have a certain distance from the “language understanding”. In order to measure the gap between the existing NLP technology and the ultimate goal of “language comprehension,” we need a task that is difficult enough, quantifiable, and reproducible. This is also the original intention of reading comprehension. Although the current research status indicates that the models that perform well on the current reading comprehension dataset still do not achieve true language comprehension, machine-reading comprehension is still regarded as an important task for the test model to understand the language.
Reading comprehension is essentially a kind of question answering. The model answers the given question after reading a paragraph of text, in this task, we introduce the use of the Learning to Search method, which translates reading comprehension into a multi-step decision process, which looking for the sentence where the answer lies from the paragraph, the starting and ending position of the answer in the sentence.
The Question Answering system uses computer to automatically answer the questions raised by users. It is one of the important tasks to verify whether the machine has natural language understanding ability. Its research history can be traced back to the origin of artificial intelligence. Compared with the retrieval system, the question answering system is an advanced form of the information service. The system returns to the user no longer the sorted keyword-based retrieval results, but an accurate natural language answer.
In an automated question answering task, we demonstrate an end-to-end question-answering system based on deep learning, which translates automated question answering into a sequence annotation problem. The end-to-end question answering system attempts to build a joint learning model by learning from high-quality "question-evidence-answer" data, and at the same time learns the semantic mapping relationship between corpora, knowledge bases, and semantic representations of question sentences. The system transform traditional question semantic analysis, text retrieval, answer extraction and generation into a learnable process
- 10.1 [A Factual Auto Answers Model Based on Sequence Labeling](https://github.com/PaddlePaddle/models/tree/develop/neural_qa)
## 11. Image classification
Compared with text, images can provide more vivid, easy to understand and more artistic information, which is an important source of people's transfer and exchange of information. Image classification is to distinguish different types of images based on the semantic information of the image. It is an important basic problem in computer vision and is also the basis of other high-level visual tasks such as image detection, image segmentation, object tracking, and behavior analysis. It is widely used in many fields. Applications. For example, face recognition and intelligent video analysis in the field of security, traffic scene recognition in the traffic field, content-based image retrieval and album auto-categorization in the Internet field, and image recognition in the medical field.
For the example of image classification, we show you how to train AlexNet, VGG, GoogLeNet, ResNet, Inception-v4, Inception-Resnet-V2 and Xception models in PaddlePaddle. It also provides model conversion tools that convert Caffe or TensorFlow trained model files into PaddlePaddle model files.
- 9.1 [convert Caffe model file to PaddlePaddle model file](https://github.com/PaddlePaddle/models/tree/develop/image_classification/caffe2paddle)
- 9.2 [convert TensorFlow model file to PaddlePaddle model file](https://github.com/PaddlePaddle/models/tree/develop/image_classification/tf2paddle)
The goal of the target detection task is to give an image or video frame, let the computer find the location of all the targets, and give each target a specific category. Target detection is a very simple task for humans. However, computers can only “see” a matrix with values between 0 and 255. It is difficult to solve high-level semantic concepts such as humans or objects in images or video frames, and it is even more difficult to locate which area in the image the target appears in. At the same time, because the target will appear in any position in the image or video frame, the shape of the target is ever-changing, and the background of the image or video frame is very different. Many factors make the target detection to be a challenging problem for the computer.
In the target detection task, we perform to use SSD method to complete the target detection. SSD(Single Shot MultiBox Detector) is one of the newer and better-performing detection algorithms in the field of target detection. It has the features of high detection speed and detection accuracy.
Many scene images contain rich text information, which plays an important role in understanding image information and can greatly help people to cognize and understand the content of scene images. Scene text recognition is a process of converting image information into a text sequence with complex image background, low resolution, various fonts, and random distribution. It can be considered as a special translation process: translating image input into natural language output. The development of scene image text recognition technology has also promoted the emergence of new applications such as helping Street View applications to obtain more accurate address information by automatically recognizing words in street signs.
In the scene text recognition task, we describe how to combine CNN-based image feature extraction and RNN-based sequence translation techniques to eliminate artificially defined features, avoid character segmentation, and use automatically learned image features to achieve end-to-end unconstrainedness Character positioning and recognition.
- 13.1 [Scene Text Recognition](https://github.com/PaddlePaddle/models/tree/develop/scene_text_recognition)
## 14. Speech Recognize
Auto Speech Recognize(ASR) translates vocabulary content in human speech into computer-readable input, allowing the machine to “understand” human speech and play an important role in applications such as voice assistant, voice input, and voice interaction. Deep learning has achieved remarkable achievements in the field of speech recognition. The end-to-end deep learning method integrates traditional acoustic models, dictionaries, language models and other modules into a whole. It no longer depends on various conditional independence in hidden Markov models, and the model becomes more concise. a neural network model takes speech features as input and directly outputs the recognized text, which has become the most important means of speech recognition.
In the speech recognition task, we provide a complete pipeline based on the DeepSpeech2 model, including: feature extraction, data enhancement, model training, language model, decoding module, etc. At the same time, we provide a trained model and experience example. Everyone can use their own Voice to experience the fun of speech recognition.
