diff --git a/demo/semantic_role_labeling/data/extract_dict_feature.py b/demo/semantic_role_labeling/data/extract_dict_feature.py index 2982e54c665b41400aab0a893ff3c76335404988..daca5f01cf2b3bd231bf530f17ec760272ce93e0 100644 --- a/demo/semantic_role_labeling/data/extract_dict_feature.py +++ b/demo/semantic_role_labeling/data/extract_dict_feature.py @@ -17,24 +17,15 @@ import os from optparse import OptionParser -def extract_dict_features(pair_file, feature_file, src_dict_file, - tgt_dict_file): - src_dict = set() - tgt_dict = set() - - with open(pair_file) as fin, open(feature_file, 'w') as feature_out, open( - src_dict_file, 'w') as src_dict_out, open(tgt_dict_file, - 'w') as tgt_dict_out: +def extract_dict_features(pair_file, feature_file): + + with open(pair_file) as fin, open(feature_file, 'w') as feature_out: for line in fin: - sentence, labels = line.strip().split('\t') + sentence, predicate, labels = line.strip().split('\t') sentence_list = sentence.split() labels_list = labels.split() - src_dict.update(sentence_list) - tgt_dict.update(labels_list) - verb_index = labels_list.index('B-V') - verb_feature = sentence_list[verb_index] mark = [0] * len(labels_list) if verb_index > 0: @@ -42,47 +33,50 @@ def extract_dict_features(pair_file, feature_file, src_dict_file, ctx_n1 = sentence_list[verb_index - 1] else: ctx_n1 = 'bos' - ctx_n1_feature = ctx_n1 + + if verb_index > 1: + mark[verb_index - 2] = 1 + ctx_n2 = sentence_list[verb_index - 2] + else: + ctx_n2 = 'bos' mark[verb_index] = 1 - ctx_0_feature = sentence_list[verb_index] + ctx_0 = sentence_list[verb_index] if verb_index < len(labels_list) - 2: mark[verb_index + 1] = 1 ctx_p1 = sentence_list[verb_index + 1] else: ctx_p1 = 'eos' - ctx_p1_feature = ctx_p1 + + if verb_index < len(labels_list) - 3: + mark[verb_index + 2] = 1 + ctx_p2 = sentence_list[verb_index + 2] + else: + ctx_p2 = 'eos' + feature_str = sentence + '\t' \ - + verb_feature + '\t' \ - + ctx_n1_feature + '\t' \ - + ctx_0_feature + '\t' \ - + ctx_p1_feature + '\t' \ + + predicate + '\t' \ + + ctx_n2 + '\t' \ + + ctx_n1 + '\t' \ + + ctx_0 + '\t' \ + + ctx_p1 + '\t' \ + + ctx_p2 + '\t' \ + ' '.join([str(i) for i in mark]) + '\t' \ + labels feature_out.write(feature_str + '\n') - src_dict_out.write('\n') - src_dict_out.write('\n'.join(list(src_dict))) - - tgt_dict_out.write('\n'.join(list(tgt_dict))) if __name__ == '__main__': - usage = '-p pair_file -f feature_file -s source dictionary -t target dictionary ' + usage = '-p pair_file -f feature_file' parser = OptionParser(usage) parser.add_option('-p', dest='pair_file', help='the pair file') - parser.add_option( - '-f', dest='feature_file', help='the file to store feature') - parser.add_option( - '-s', dest='src_dict', help='the file to store source dictionary') - parser.add_option( - '-t', dest='tgt_dict', help='the file to store target dictionary') + parser.add_option('-f', dest='feature_file', help='the feature file') (options, args) = parser.parse_args() - extract_dict_features(options.pair_file, options.feature_file, - options.src_dict, options.tgt_dict) + extract_dict_features(options.pair_file, options.feature_file) diff --git a/demo/semantic_role_labeling/data/extract_pairs.py b/demo/semantic_role_labeling/data/extract_pairs.py index 4d1bef8f958a62be9941d474a0b67542dcc5cfab..86ab00ce41723169de035a841d9e129a1b9e82a3 100644 --- a/demo/semantic_role_labeling/data/extract_pairs.py +++ b/demo/semantic_role_labeling/data/extract_pairs.py @@ -51,7 +51,7 @@ def read_sentences(words_file): for line in fin: line = line.strip() if line == '': - sentences.append(s.lower()) + sentences.append(s) s = '' else: s += line + ' ' @@ -64,6 +64,11 @@ def transform_labels(sentences, labels): if len(labels[i]) == 1: continue else: + verb_list = [] + for x in labels[i][0]: + if x !='-': + verb_list.append(x) + for j in xrange(1, len(labels[i])): label_list = labels[i][j] current_tag = 'O' @@ -88,8 +93,7 @@ def transform_labels(sentences, labels): is_in_bracket = True else: print 'error:', ll - - sen_lab_pair.append((sentences[i], label_seq)) + sen_lab_pair.append((sentences[i], verb_list[j-1], label_seq)) return sen_lab_pair @@ -97,9 +101,9 @@ def write_file(sen_lab_pair, output_file): with open(output_file, 'w') as fout: for x in sen_lab_pair: sentence = x[0] - label_seq = ' '.join(x[1]) - assert len(sentence.split()) == len(x[1]) - fout.write(sentence + '\t' + label_seq + '\n') + label_seq = ' '.join(x[2]) + assert len(sentence.split()) == len(x[2]) + fout.write(sentence + '\t' + x[1]+'\t' +label_seq + '\n') if __name__ == '__main__': diff --git a/demo/semantic_role_labeling/data/get_data.sh b/demo/semantic_role_labeling/data/get_data.sh index 268c0995e27006ec62f38bdda9b0a0994dab096c..55e33f4685627ed483aa6642c518a33558091531 100644 --- a/demo/semantic_role_labeling/data/get_data.sh +++ b/demo/semantic_role_labeling/data/get_data.sh @@ -14,6 +14,10 @@ # limitations under the License. set -e wget http://www.cs.upc.edu/~srlconll/conll05st-tests.tar.gz +wget https://www.googledrive.com/host/0B7Q8d52jqeI9ejh6Q1RpMTFQT1k/semantic_role_labeling/verbDict.txt --no-check-certificate +wget https://www.googledrive.com/host/0B7Q8d52jqeI9ejh6Q1RpMTFQT1k/semantic_role_labeling/targetDict.txt --no-check-certificate +wget https://www.googledrive.com/host/0B7Q8d52jqeI9ejh6Q1RpMTFQT1k/semantic_role_labeling/wordDict.txt --no-check-certificate +wget https://www.googledrive.com/host/0B7Q8d52jqeI9ejh6Q1RpMTFQT1k/semantic_role_labeling/emb --no-check-certificate tar -xzvf conll05st-tests.tar.gz rm conll05st-tests.tar.gz cp ./conll05st-release/test.wsj/words/test.wsj.words.gz . @@ -22,4 +26,4 @@ gunzip test.wsj.words.gz gunzip test.wsj.props.gz python extract_pairs.py -w test.wsj.words -p test.wsj.props -o test.wsj.seq_pair -python extract_dict_feature.py -p test.wsj.seq_pair -f feature -s src.dict -t tgt.dict +python extract_dict_feature.py -p test.wsj.seq_pair -f feature diff --git a/demo/semantic_role_labeling/dataprovider.py b/demo/semantic_role_labeling/dataprovider.py index 5c003584a52d459f13b7942ebe3a7147ac58a42f..d4c137ef42c4e2ec609f3e6f809363e602dfd8dd 100644 --- a/demo/semantic_role_labeling/dataprovider.py +++ b/demo/semantic_role_labeling/dataprovider.py @@ -17,11 +17,15 @@ from paddle.trainer.PyDataProvider2 import * UNK_IDX = 0 -def hook(settings, word_dict, label_dict, **kwargs): +def hook(settings, word_dict, label_dict, predicate_dict, **kwargs): settings.word_dict = word_dict settings.label_dict = label_dict + settings.predicate_dict = predicate_dict + #all inputs are integral and sequential type settings.slots = [ + integer_value_sequence(len(word_dict)), + integer_value_sequence(len(predicate_dict)), integer_value_sequence(len(word_dict)), integer_value_sequence(len(word_dict)), integer_value_sequence(len(word_dict)), @@ -31,27 +35,33 @@ def hook(settings, word_dict, label_dict, **kwargs): ] -@provider(init_hook=hook) -def process(obj, file_name): +def get_batch_size(yeild_data): + return len(yeild_data[0]) + + +@provider(init_hook=hook, should_shuffle=True, calc_batch_size=get_batch_size, + can_over_batch_size=False, cache=CacheType.CACHE_PASS_IN_MEM) +def process(settings, file_name): with open(file_name, 'r') as fdata: for line in fdata: - sentence, predicate, ctx_n1, ctx_0, ctx_p1, mark, label = \ + sentence, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, label = \ line.strip().split('\t') - + words = sentence.split() sen_len = len(words) - word_slot = [obj.word_dict.get(w, UNK_IDX) for w in words] + word_slot = [settings.word_dict.get(w, UNK_IDX) for w in words] - predicate_slot = [obj.word_dict.get(predicate, UNK_IDX)] * sen_len - ctx_n1_slot = [obj.word_dict.get(ctx_n1, UNK_IDX)] * sen_len - ctx_0_slot = [obj.word_dict.get(ctx_0, UNK_IDX)] * sen_len - ctx_p1_slot = [obj.word_dict.get(ctx_p1, UNK_IDX)] * sen_len + predicate_slot = [settings.predicate_dict.get(predicate)] * sen_len + ctx_n2_slot = [settings.word_dict.get(ctx_n2, UNK_IDX)] * sen_len + ctx_n1_slot = [settings.word_dict.get(ctx_n1, UNK_IDX)] * sen_len + ctx_0_slot = [settings.word_dict.get(ctx_0, UNK_IDX)] * sen_len + ctx_p1_slot = [settings.word_dict.get(ctx_p1, UNK_IDX)] * sen_len + ctx_p2_slot = [settings.word_dict.get(ctx_p2, UNK_IDX)] * sen_len marks = mark.split() mark_slot = [int(w) for w in marks] label_list = label.split() - label_slot = [obj.label_dict.get(w) for w in label_list] - - yield word_slot, predicate_slot, ctx_n1_slot, \ - ctx_0_slot, ctx_p1_slot, mark_slot, label_slot + label_slot = [settings.label_dict.get(w) for w in label_list] + yield word_slot, predicate_slot, ctx_n2_slot, ctx_n1_slot, \ + ctx_0_slot, ctx_p1_slot, ctx_p2_slot, mark_slot, label_slot diff --git a/demo/semantic_role_labeling/db_lstm.py b/demo/semantic_role_labeling/db_lstm.py index e3f6edad6972112ed04e173a9b714e3fec13d402..54ceff0e724220cc9ea96b9e0ec6844947a8343e 100644 --- a/demo/semantic_role_labeling/db_lstm.py +++ b/demo/semantic_role_labeling/db_lstm.py @@ -18,8 +18,9 @@ import sys from paddle.trainer_config_helpers import * #file paths -word_dict_file = './data/src.dict' -label_dict_file = './data/tgt.dict' +word_dict_file = './data/wordDict.txt' +label_dict_file = './data/targetDict.txt' +predicate_file= './data/verbDict.txt' train_list_file = './data/train.list' test_list_file = './data/test.list' @@ -30,8 +31,10 @@ if not is_predict: #load dictionaries word_dict = dict() label_dict = dict() + predicate_dict = dict() with open(word_dict_file, 'r') as f_word, \ - open(label_dict_file, 'r') as f_label: + open(label_dict_file, 'r') as f_label, \ + open(predicate_file, 'r') as f_pre: for i, line in enumerate(f_word): w = line.strip() word_dict[w] = i @@ -40,6 +43,11 @@ if not is_predict: w = line.strip() label_dict[w] = i + for i, line in enumerate(f_pre): + w = line.strip() + predicate_dict[w] = i + + if is_test: train_list_file = None @@ -50,91 +58,157 @@ if not is_predict: module='dataprovider', obj='process', args={'word_dict': word_dict, - 'label_dict': label_dict}) + 'label_dict': label_dict, + 'predicate_dict': predicate_dict }) word_dict_len = len(word_dict) label_dict_len = len(label_dict) + pred_len = len(predicate_dict) else: word_dict_len = get_config_arg('dict_len', int) label_dict_len = get_config_arg('label_len', int) + pred_len = get_config_arg('pred_len', int) +############################## Hyper-parameters ################################## mark_dict_len = 2 word_dim = 32 mark_dim = 5 -hidden_dim = 128 +hidden_dim = 512 depth = 8 -emb_lr = 1e-2 -fc_lr = 1e-2 -lstm_lr = 2e-2 + + + +########################### Optimizer ####################################### + settings( batch_size=150, - learning_method=AdamOptimizer(), - learning_rate=1e-3, + learning_method=MomentumOptimizer(momentum=0), + learning_rate=2e-2, regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) + is_async=False, + model_average=ModelAverage(average_window=0.5, + max_average_window=10000), + +) -#6 features + + + +####################################### network ############################## +#8 features and 1 target word = data_layer(name='word_data', size=word_dict_len) -predicate = data_layer(name='verb_data', size=word_dict_len) +predicate = data_layer(name='verb_data', size=pred_len) + +ctx_n2 = data_layer(name='ctx_n2_data', size=word_dict_len) ctx_n1 = data_layer(name='ctx_n1_data', size=word_dict_len) ctx_0 = data_layer(name='ctx_0_data', size=word_dict_len) ctx_p1 = data_layer(name='ctx_p1_data', size=word_dict_len) +ctx_p2 = data_layer(name='ctx_p2_data', size=word_dict_len) mark = data_layer(name='mark_data', size=mark_dict_len) + if not is_predict: target = data_layer(name='target', size=label_dict_len) -ptt = ParameterAttribute(name='src_emb', learning_rate=emb_lr) -layer_attr = ExtraLayerAttribute(drop_rate=0.5) -fc_para_attr = ParameterAttribute(learning_rate=fc_lr) -lstm_para_attr = ParameterAttribute(initial_std=0., learning_rate=lstm_lr) -para_attr = [fc_para_attr, lstm_para_attr] -word_embedding = embedding_layer(size=word_dim, input=word, param_attr=ptt) -predicate_embedding = embedding_layer( - size=word_dim, input=predicate, param_attr=ptt) -ctx_n1_embedding = embedding_layer(size=word_dim, input=ctx_n1, param_attr=ptt) -ctx_0_embedding = embedding_layer(size=word_dim, input=ctx_0, param_attr=ptt) -ctx_p1_embedding = embedding_layer(size=word_dim, input=ctx_p1, param_attr=ptt) -mark_embedding = embedding_layer(size=mark_dim, input=mark) +default_std=1/math.sqrt(hidden_dim)/3.0 + +emb_para = ParameterAttribute(name='emb', initial_std=0., learning_rate=0.) +std_0 = ParameterAttribute(initial_std=0.) +std_default = ParameterAttribute(initial_std=default_std) + +predicate_embedding = embedding_layer(size=word_dim, input=predicate, param_attr=ParameterAttribute(name='vemb',initial_std=default_std)) +mark_embedding = embedding_layer(name='word_ctx-in_embedding', size=mark_dim, input=mark, param_attr=std_0) + +word_input=[word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] +emb_layers = [embedding_layer(size=word_dim, input=x, param_attr=emb_para) for x in word_input] +emb_layers.append(predicate_embedding) +emb_layers.append(mark_embedding) hidden_0 = mixed_layer( + name='hidden0', size=hidden_dim, - input=[ - full_matrix_projection(input=word_embedding), - full_matrix_projection(input=predicate_embedding), - full_matrix_projection(input=ctx_n1_embedding), - full_matrix_projection(input=ctx_0_embedding), - full_matrix_projection(input=ctx_p1_embedding), - full_matrix_projection(input=mark_embedding), - ]) + bias_attr=std_default, + input=[ full_matrix_projection(input=emb, param_attr=std_default ) for emb in emb_layers ]) + -lstm_0 = lstmemory(input=hidden_0, layer_attr=layer_attr) +mix_hidden_lr = 1e-3 +lstm_para_attr = ParameterAttribute(initial_std=0.0, learning_rate=1.0) +hidden_para_attr = ParameterAttribute(initial_std=default_std, learning_rate=mix_hidden_lr) + +lstm_0 = lstmemory(name='lstm0', + input=hidden_0, + act=ReluActivation(), + gate_act=SigmoidActivation(), + state_act=SigmoidActivation(), + bias_attr=std_0, + param_attr=lstm_para_attr) #stack L-LSTM and R-LSTM with direct edges input_tmp = [hidden_0, lstm_0] + for i in range(1, depth): - fc = fc_layer(input=input_tmp, size=hidden_dim, param_attr=para_attr) + mix_hidden = mixed_layer(name='hidden'+str(i), + size=hidden_dim, + bias_attr=std_default, + input=[full_matrix_projection(input=input_tmp[0], param_attr=hidden_para_attr), + full_matrix_projection(input=input_tmp[1], param_attr=lstm_para_attr) + ] + ) + + lstm = lstmemory(name='lstm'+str(i), + input=mix_hidden, + act=ReluActivation(), + gate_act=SigmoidActivation(), + state_act=SigmoidActivation(), + reverse=((i % 2)==1), + bias_attr=std_0, + param_attr=lstm_para_attr) + + input_tmp = [mix_hidden, lstm] + +feature_out = mixed_layer(name='output', + size=label_dict_len, + bias_attr=std_default, + input=[full_matrix_projection(input=input_tmp[0], param_attr=hidden_para_attr), + full_matrix_projection(input=input_tmp[1], param_attr=lstm_para_attr) + ], + ) - lstm = lstmemory( - input=fc, - act=ReluActivation(), - reverse=(i % 2) == 1, - layer_attr=layer_attr) - input_tmp = [fc, lstm] -prob = fc_layer( - input=input_tmp, - size=label_dict_len, - act=SoftmaxActivation(), - param_attr=para_attr) if not is_predict: - cls = classification_cost(input=prob, label=target) - outputs(cls) + crf_l = crf_layer( name = 'crf', + size = label_dict_len, + input = feature_out, + label = target, + param_attr=ParameterAttribute(name='crfw',initial_std=default_std, learning_rate=mix_hidden_lr) + + ) + + + crf_dec_l = crf_decoding_layer(name = 'crf_dec_l', + size = label_dict_len, + input = feature_out, + label = target, + param_attr=ParameterAttribute(name='crfw') + ) + + + eval = sum_evaluator(input=crf_dec_l) + + outputs(crf_l) + else: - outputs(prob) + crf_dec_l = crf_decoding_layer(name = 'crf_dec_l', + size = label_dict_len, + input = feature_out, + param_attr=ParameterAttribute(name='crfw') + ) + + outputs(crf_dec_l) + diff --git a/demo/semantic_role_labeling/predict.py b/demo/semantic_role_labeling/predict.py index f051d4175cf6fff43bd7f84b457ab9dd12405a15..2761814e1811e701122e0be4850526c5b290c457 100644 --- a/demo/semantic_role_labeling/predict.py +++ b/demo/semantic_role_labeling/predict.py @@ -26,7 +26,7 @@ UNK_IDX = 0 class Prediction(): - def __init__(self, train_conf, dict_file, model_dir, label_file): + def __init__(self, train_conf, dict_file, model_dir, label_file, predicate_dict_file): """ train_conf: trainer configure. dict_file: word dictionary file name. @@ -35,26 +35,41 @@ class Prediction(): self.dict = {} self.labels = {} + self.predicate_dict={} self.labels_reverse = {} - self.load_dict_label(dict_file, label_file) + self.load_dict_label(dict_file, label_file, predicate_dict_file) len_dict = len(self.dict) len_label = len(self.labels) - - conf = parse_config(train_conf, 'dict_len=' + str(len_dict) + - ',label_len=' + str(len_label) + ',is_predict=True') + len_pred = len(self.predicate_dict) + + conf = parse_config( + train_conf, + 'dict_len=' + str(len_dict) + + ',label_len=' + str(len_label) + + ',pred_len=' + str(len_pred) + + ',is_predict=True') self.network = swig_paddle.GradientMachine.createFromConfigProto( conf.model_config) self.network.loadParameters(model_dir) slots = [ + integer_value_sequence(len_dict), + integer_value_sequence(len_pred), + integer_value_sequence(len_dict), + integer_value_sequence(len_dict), + integer_value_sequence(len_dict), + integer_value_sequence(len_dict), + integer_value_sequence(len_dict), + integer_value_sequence(2) + ] integer_value_sequence(len_dict), integer_value_sequence(len_dict), integer_value_sequence(len_dict), integer_value_sequence(len_dict), integer_value_sequence(len_dict), integer_value_sequence(2) ] self.converter = DataProviderConverter(slots) - def load_dict_label(self, dict_file, label_file): + def load_dict_label(self, dict_file, label_file, predicate_dict_file): """ Load dictionary from self.dict_file. """ @@ -65,39 +80,42 @@ class Prediction(): self.labels[line.strip()] = line_count self.labels_reverse[line_count] = line.strip() + for line_count, line in enumerate(open(predicate_dict_file, 'r')): + self.predicate_dict[line.strip()] = line_count def get_data(self, data_file): """ Get input data of paddle format. """ with open(data_file, 'r') as fdata: for line in fdata: - sentence, predicate, ctx_n1, ctx_0, ctx_p1, mark, label = line.strip( + sentence, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, label = line.strip( ).split('\t') words = sentence.split() sen_len = len(words) - + word_slot = [self.dict.get(w, UNK_IDX) for w in words] - predicate_slot = [self.dict.get(predicate, UNK_IDX)] * sen_len + predicate_slot = [self.predicate_dict.get(predicate, UNK_IDX)] * sen_len + ctx_n2_slot = [self.dict.get(ctx_n2, UNK_IDX)] * sen_len ctx_n1_slot = [self.dict.get(ctx_n1, UNK_IDX)] * sen_len ctx_0_slot = [self.dict.get(ctx_0, UNK_IDX)] * sen_len ctx_p1_slot = [self.dict.get(ctx_p1, UNK_IDX)] * sen_len + ctx_p2_slot = [self.dict.get(ctx_p2, UNK_IDX)] * sen_len marks = mark.split() mark_slot = [int(w) for w in marks] + + yield word_slot, predicate_slot, ctx_n2_slot, ctx_n1_slot, \ + ctx_0_slot, ctx_p1_slot, ctx_p2_slot, mark_slot - yield word_slot, predicate_slot, ctx_n1_slot, \ - ctx_0_slot, ctx_p1_slot, mark_slot - - def predict(self, data_file): + def predict(self, data_file, output_file): """ data_file: file name of input data. """ input = self.converter(self.get_data(data_file)) output = self.network.forwardTest(input) - prob = output[0]["value"] - lab = list(np.argsort(-prob)[:, 0]) + lab = output[0]["id"].tolist() - with open(data_file, 'r') as fin, open('predict.res', 'w') as fout: + with open(data_file, 'r') as fin, open(output_file, 'w') as fout: index = 0 for line in fin: sen = line.split('\t')[0] @@ -109,8 +127,8 @@ class Prediction(): def option_parser(): - usage = ("python predict.py -c config -w model_dir " - "-d word dictionary -l label_file -i input_file") + usage = ("python predict.py -c config -w model_dir " + "-d word dictionary -l label_file -i input_file -p pred_dict_file") parser = OptionParser(usage="usage: %s [options]" % usage) parser.add_option( "-c", @@ -131,6 +149,13 @@ def option_parser(): dest="label_file", default=None, help="label file") + parser.add_option( + "-p", + "--predict_dict_file", + action="store", + dest="predict_dict_file", + default=None, + help="predict_dict_file") parser.add_option( "-i", "--data", @@ -144,6 +169,14 @@ def option_parser(): dest="model_path", default=None, help="model path") + + parser.add_option( + "-o", + "--output_file", + action="store", + dest="output_file", + default=None, + help="output file") return parser.parse_args() @@ -154,10 +187,12 @@ def main(): dict_file = options.dict_file model_path = options.model_path label_file = options.label_file + predict_dict_file = options.predict_dict_file + output_file = options.output_file swig_paddle.initPaddle("--use_gpu=0") - predict = Prediction(train_conf, dict_file, model_path, label_file) - predict.predict(data_file) + predict = Prediction(train_conf, dict_file, model_path, label_file, predict_dict_file) + predict.predict(data_file,output_file) if __name__ == '__main__': diff --git a/demo/semantic_role_labeling/predict.sh b/demo/semantic_role_labeling/predict.sh index a545b9a5d591b41bdbd54905cbbffc410abc8fb0..d0acdb0bd093974485475cf796c6d41ac7899135 100644 --- a/demo/semantic_role_labeling/predict.sh +++ b/demo/semantic_role_labeling/predict.sh @@ -26,15 +26,18 @@ LOG=`get_best_pass $log` LOG=(${LOG}) best_model_path="output/pass-${LOG[1]}" - config_file=db_lstm.py -dict_file=./data/src.dict -label_file=./data/tgt.dict +dict_file=./data/wordDict.txt +label_file=./data/targetDict.txt +predicate_dict_file=./data/verbDict.txt input_file=./data/feature +output_file=predict.res python predict.py \ -c $config_file \ -w $best_model_path \ -l $label_file \ + -p $predicate_dict_file \ -d $dict_file \ - -i $input_file + -i $input_file \ + -o $output_file diff --git a/demo/semantic_role_labeling/test.sh b/demo/semantic_role_labeling/test.sh index 844649e8c0f6867dc0766e4ec6f250c5a4a004d9..c4ab44f5ca08aefd18f2851a1410aa08563925a9 100644 --- a/demo/semantic_role_labeling/test.sh +++ b/demo/semantic_role_labeling/test.sh @@ -36,4 +36,5 @@ paddle train \ --job=test \ --use_gpu=false \ --config_args=is_test=1 \ + --test_all_data_in_one_period=1 \ 2>&1 | tee 'test.log' diff --git a/demo/semantic_role_labeling/train.sh b/demo/semantic_role_labeling/train.sh index c3a22b644be0ca08a2af73a57c09657014e49bfc..420768bb2b4ebed7b135a49c5eee5e5538426ae1 100644 --- a/demo/semantic_role_labeling/train.sh +++ b/demo/semantic_role_labeling/train.sh @@ -16,11 +16,14 @@ set -e paddle train \ --config=./db_lstm.py \ + --use_gpu=0 \ + --log_period=5000 \ + --trainer_count=1 \ + --show_parameter_stats_period=5000 \ --save_dir=./output \ - --trainer_count=4 \ - --log_period=10 \ - --num_passes=500 \ - --use_gpu=false \ - --show_parameter_stats_period=10 \ + --num_passes=10000 \ + --average_test_period=10000000 \ + --init_model_path=./data \ + --load_missing_parameter_strategy=rand \ --test_all_data_in_one_period=1 \ -2>&1 | tee 'train.log' + 2>&1 | tee 'train.log' diff --git a/doc/demo/semantic_role_labeling/curve.jpg b/doc/demo/semantic_role_labeling/curve.jpg new file mode 100644 index 0000000000000000000000000000000000000000..baa35ae7f0a0b6c246f3a0d331735477ab8bcd70 Binary files /dev/null and b/doc/demo/semantic_role_labeling/curve.jpg differ diff --git a/doc/demo/semantic_role_labeling/semantic_role_labeling.md b/doc/demo/semantic_role_labeling/semantic_role_labeling.md index 890f7314582c65e9add50664006b57aa4e0709eb..e2793b2b3494160a7a80f07ec2127bd1f1a4f2e4 100644 --- a/doc/demo/semantic_role_labeling/semantic_role_labeling.md +++ b/doc/demo/semantic_role_labeling/semantic_role_labeling.md @@ -1,183 +1,200 @@ -# Semantic Role labeling Tutorial # - -Semantic role labeling (SRL) is a form of shallow semantic parsing whose goal is to discover the predicate-argument structure of each predicate in a given input sentence. SRL is useful as an intermediate step in a wide range of natural language processing tasks, such as information extraction. automatic document categorization and question answering. An instance is as following [1]: - - [ A0 He ] [ AM-MOD would ][ AM-NEG n’t ] [ V accept] [ A1 anything of value ] from [A2 those he was writing about ]. - -- V: verb -- A0: acceptor -- A1: thing accepted -- A2: accepted-from -- A3: Attribute -- AM-MOD: modal -- AM-NEG: negation - -Given the verb "accept", the chunks in sentence would play certain semantic roles. Here, the label scheme is from Penn Proposition Bank. - -To this date, most of the successful SRL systems are built on top of some form of parsing results where pre-defined feature templates over the syntactic structure are used. This tutorial will present an end-to-end system using deep bidirectional long short-term memory (DB-LSTM)[2] for solving the SRL task, which largely outperforms the previous state-of-the-art systems. The system regards SRL task as the sequence labelling problem. - -## Data Description -The relevant paper[2] takes the data set in CoNLL-2005&2012 Shared Task for training and testing. Accordingto data license, the demo adopts the test data set of CoNLL-2005, which can be reached on website. - -To download and process the original data, user just need to execute the following command: - -```bash -cd data -./get_data.sh -``` -Several new files appear in the `data `directory as follows. -```bash -conll05st-release：the test data set of CoNll-2005 shared task -test.wsj.words：the Wall Street Journal data sentences -test.wsj.props: the propositional arguments -src.dict：the dictionary of words in sentences -tgt.dict：the labels dictionary -feature: the extracted features from data set -``` - -## Training -### DB-LSTM -Please refer to the Sentiment Analysis demo to learn more about the long short-term memory unit. - -Unlike Bidirectional-LSTM that used in Sentiment Analysis demo, the DB-LSTM adopts another way to stack LSTM layer. First a standard LSTM processes the sequence in forward direction. The input and output of this LSTM layer are taken by the next LSTM layer as input, processed in reversed direction. These two standard LSTM layers compose a pair of LSTM. Then we stack LSTM layers pair after pair to obtain the deep LSTM model. - -The following figure shows a temporal expanded 2-layer DB-LSTM network. -
-![pic](./network_arch.png) -
- -### Features -Two input features play an essential role in this pipeline: predicate (pred) and argument (argu). Two other features: predicate context (ctx-p) and region mark (mr) are also adopted. Because a single predicate word can not exactly describe the predicate information, especially when the same words appear more than one times in a sentence. With the predicate context, the ambiguity can be largely eliminated. Similarly, we use region mark mr = 1 to denote the argument position if it locates in the predicate context region, or mr = 0 if does not. These four simple features are all we need for our SRL system. Features of one sample with context size set to 1 is showed as following[2]: -
-![pic](./feature.jpg) -
- -In this sample, the coresponding labelled sentence is: - -[ A1 A record date ] has [ AM-NEG n't ] been [ V set ] . - -In the demo, we adopt the feature template as above, consists of : `argument`, `predicate`, `ctx-p (p=-1,0,1)`, `mark` and use `B/I/O` scheme to label each argument. These features and labels are stored in `feature` file, and separated by `\t`. - -### Data Provider - -`dataprovider.py` is the python file to wrap data. `hook()` function is to define the data slots for network. The Six features and label are all IndexSlots. -``` -def hook(settings, word_dict, label_dict, **kwargs): - settings.word_dict = word_dict - settings.label_dict = label_dict - #all inputs are integral and sequential type - settings.slots = [ - integer_value_sequence(len(word_dict)), - integer_value_sequence(len(word_dict)), - integer_value_sequence(len(word_dict)), - integer_value_sequence(len(word_dict)), - integer_value_sequence(len(word_dict)), - integer_value_sequence(2), - integer_value_sequence(len(label_dict))] -``` -The corresponding data iterator is as following: -``` -@provider(use_seq=True, init_hook=hook) -def process(obj, file_name): - with open(file_name, 'r') as fdata: - for line in fdata: - sentence, predicate, ctx_n1, ctx_0, ctx_p1, mark, label = line.strip().split('\t') - words = sentence.split() - sen_len = len(words) - word_slot = [obj.word_dict.get(w, UNK_IDX) for w in words] - - predicate_slot = [obj.word_dict.get(predicate, UNK_IDX)] * sen_len - ctx_n1_slot = [obj.word_dict.get(ctx_n1, UNK_IDX) ] * sen_len - ctx_0_slot = [obj.word_dict.get(ctx_0, UNK_IDX) ] * sen_len - ctx_p1_slot = [obj.word_dict.get(ctx_p1, UNK_IDX) ] * sen_len - - marks = mark.split() - mark_slot = [int(w) for w in marks] - - label_list = label.split() - label_slot = [obj.label_dict.get(w) for w in label_list] - - yield word_slot, predicate_slot, ctx_n1_slot, ctx_0_slot, ctx_p1_slot, mark_slot, label_slot -``` -The `process`function yield 7 lists which are six features and labels. - -### Neural Network Config -`db_lstm.py` is the neural network config file to load the dictionaries and define the data provider module and network architecture during the training procedure. - -Seven `data_layer` load instances from data provider. Six features are transformed into embedddings respectively, and mixed by `mixed_layer` . Deep bidirectional LSTM layers extract features for the softmax layer. The objective function is cross entropy of labels. - -### Run Training -The script for training is `train.sh`, user just need to execute: -```bash - ./train.sh -``` -The content in `train.sh`: -``` -paddle train \ - --config=./db_lstm.py \ - --save_dir=./output \ - --trainer_count=4 \ - --log_period=10 \ - --num_passes=500 \ - --use_gpu=false \ - --show_parameter_stats_period=10 \ - --test_all_data_in_one_period=1 \ -2>&1 | tee 'train.log' -``` - -- \--config=./db_lstm.py : network config file. -- \--save_di=./output: output path to save models. -- \--trainer_count=4 : set thread number (or GPU count). -- \--log_period=10 : print log every 20 batches. -- \--num_passes=500: set pass number, one pass in PaddlePaddle means training all samples in dataset one time. -- \--use_gpu=false: use CPU to train, set true, if you install GPU version of PaddlePaddle and want to use GPU to train. -- \--show_parameter_stats_period=10: show parameter statistic every 100 batches. -- \--test_all_data_in_one_period=1: test all data in every testing. - - -After training, the models will be saved in directory `output`. - -### Run testing -The script for testing is `test.sh`, user just need to execute: -```bash - ./test.sh -``` -The main part in `tesh.sh` -``` -paddle train \ - --config=./db_lstm.py \ - --model_list=$model_list \ - --job=test \ - --config_args=is_test=1 \ -``` - - - \--config=./db_lstm.py: network config file - - \--model_list=$model_list.list: model list file - - \--job=test: indicate the test job - - \--config_args=is_test=1: flag to indicate test - - -### Run prediction -The script for prediction is `predict.sh`, user just need to execute: -```bash - ./predict.sh - -``` -In `predict.sh`, user should offer the network config file, model path, label file, word dictionary file, feature file -``` -python predict.py - -c $config_file - -w $model_path - -l $label_file - -d $dict_file - -i $input_file -``` - -`predict.py` is the main executable python script, which includes functions: load model, load data, data prediction. The network model will output the probability distribution of labels. In the demo, we take the label with maximum probability as result. User can also implement the beam search or viterbi decoding upon the probability distribution matrix. - -After prediction, the result is saved in `predict.res`. - -## Reference -[1] Martha Palmer, Dan Gildea, and Paul Kingsbury. The Proposition Bank: An Annotated Corpus of Semantic Roles , Computational Linguistics, 31(1), 2005. - -[2] Zhou, Jie, and Wei Xu. "End-to-end learning of semantic role labeling using recurrent neural networks." Proceedings of the Annual Meeting of the Association for Computational Linguistics. 2015. +# Semantic Role labeling Tutorial # + +Semantic role labeling (SRL) is a form of shallow semantic parsing whose goal is to discover the predicate-argument structure of each predicate in a given input sentence. SRL is useful as an intermediate step in a wide range of natural language processing tasks, such as information extraction. automatic document categorization and question answering. An instance is as following [1]: + + [ A0 He ] [ AM-MOD would ][ AM-NEG n’t ] [ V accept] [ A1 anything of value ] from [A2 those he was writing about ]. + +- V: verb +- A0: acceptor +- A1: thing accepted +- A2: accepted-from +- A3: Attribute +- AM-MOD: modal +- AM-NEG: negation + +Given the verb "accept", the chunks in sentence would play certain semantic roles. Here, the label scheme is from Penn Proposition Bank. + +To this date, most of the successful SRL systems are built on top of some form of parsing results where pre-defined feature templates over the syntactic structure are used. This tutorial will present an end-to-end system using deep bidirectional long short-term memory (DB-LSTM)[2] for solving the SRL task, which largely outperforms the previous state-of-the-art systems. The system regards SRL task as the sequence labelling problem. + +## Data Description +The relevant paper[2] takes the data set in CoNLL-2005&2012 Shared Task for training and testing. Accordingto data license, the demo adopts the test data set of CoNLL-2005, which can be reached on website. + +To download and process the original data, user just need to execute the following command: + +```bash +cd data +./get_data.sh +``` +Several new files appear in the `data `directory as follows. +```bash +conll05st-release：the test data set of CoNll-2005 shared task +test.wsj.words：the Wall Street Journal data sentences +test.wsj.props: the propositional arguments +feature: the extracted features from data set +``` + +## Training +### DB-LSTM +Please refer to the Sentiment Analysis demo to learn more about the long short-term memory unit. + +Unlike Bidirectional-LSTM that used in Sentiment Analysis demo, the DB-LSTM adopts another way to stack LSTM layer. First a standard LSTM processes the sequence in forward direction. The input and output of this LSTM layer are taken by the next LSTM layer as input, processed in reversed direction. These two standard LSTM layers compose a pair of LSTM. Then we stack LSTM layers pair after pair to obtain the deep LSTM model. + +The following figure shows a temporal expanded 2-layer DB-LSTM network. +
+![pic](./network_arch.png) +
+ +### Features +Two input features play an essential role in this pipeline: predicate (pred) and argument (argu). Two other features: predicate context (ctx-p) and region mark (mr) are also adopted. Because a single predicate word can not exactly describe the predicate information, especially when the same words appear more than one times in a sentence. With the predicate context, the ambiguity can be largely eliminated. Similarly, we use region mark mr = 1 to denote the argument position if it locates in the predicate context region, or mr = 0 if does not. These four simple features are all we need for our SRL system. Features of one sample with context size set to 1 is showed as following[2]: +
+![pic](./feature.jpg) +
+ +In this sample, the coresponding labelled sentence is: + +[ A1 A record date ] has [ AM-NEG n't ] been [ V set ] . + +In the demo, we adopt the feature template as above, consists of : `argument`, `predicate`, `ctx-p (p=-1,0,1)`, `mark` and use `B/I/O` scheme to label each argument. These features and labels are stored in `feature` file, and separated by `\t`. + +### Data Provider + +`dataprovider.py` is the python file to wrap data. `hook()` function is to define the data slots for network. The Six features and label are all IndexSlots. +``` +def hook(settings, word_dict, label_dict, **kwargs): + settings.word_dict = word_dict + settings.label_dict = label_dict + #all inputs are integral and sequential type + settings.slots = [ + integer_value_sequence(len(word_dict)), + integer_value_sequence(len(predicate_dict)), + integer_value_sequence(len(word_dict)), + integer_value_sequence(len(word_dict)), + integer_value_sequence(len(word_dict)), + integer_value_sequence(len(word_dict)), + integer_value_sequence(len(word_dict)), + integer_value_sequence(2), + integer_value_sequence(len(label_dict))] +``` +The corresponding data iterator is as following: +``` +@provider(init_hook=hook, should_shuffle=True, calc_batch_size=get_batch_size, + can_over_batch_size=False, cache=CacheType.CACHE_PASS_IN_MEM) +def process(settings, file_name): + with open(file_name, 'r') as fdata: + for line in fdata: + sentence, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, label = \ + line.strip().split('\t') + + words = sentence.split() + sen_len = len(words) + word_slot = [settings.word_dict.get(w, UNK_IDX) for w in words] + + predicate_slot = [settings.predicate_dict.get(predicate)] * sen_len + ctx_n2_slot = [settings.word_dict.get(ctx_n2, UNK_IDX)] * sen_len + ctx_n1_slot = [settings.word_dict.get(ctx_n1, UNK_IDX)] * sen_len + ctx_0_slot = [settings.word_dict.get(ctx_0, UNK_IDX)] * sen_len + ctx_p1_slot = [settings.word_dict.get(ctx_p1, UNK_IDX)] * sen_len + ctx_p2_slot = [settings.word_dict.get(ctx_p2, UNK_IDX)] * sen_len + + marks = mark.split() + mark_slot = [int(w) for w in marks] + + label_list = label.split() + label_slot = [settings.label_dict.get(w) for w in label_list] + yield word_slot, predicate_slot, ctx_n2_slot, ctx_n1_slot, \ + ctx_0_slot, ctx_p1_slot, ctx_p2_slot, mark_slot, label_slot +``` +The `process`function yield 9 lists which are 8 features and label. + +### Neural Network Config +`db_lstm.py` is the neural network config file to load the dictionaries and define the data provider module and network architecture during the training procedure. + +Nine `data_layer` load instances from data provider. Eight features are transformed into embedddings respectively, and mixed by `mixed_layer` . Deep bidirectional LSTM layers extract features for the softmax layer. The objective function is cross entropy of labels. + +### Run Training +The script for training is `train.sh`, user just need to execute: +```bash + ./train.sh +``` +The content in `train.sh`: +``` +paddle train \ + --config=./db_lstm.py \ + --use_gpu=0 \ + --log_period=5000 \ + --trainer_count=1 \ + --show_parameter_stats_period=5000 \ + --save_dir=./output \ + --num_passes=10000 \ + --average_test_period=10000000 \ + --init_model_path=./data \ + --load_missing_parameter_strategy=rand \ + --test_all_data_in_one_period=1 \ +2>&1 | tee 'train.log' +``` + +- \--config=./db_lstm.py : network config file. +- \--use_gpu=false: use CPU to train, set true, if you install GPU version of PaddlePaddle and want to use GPU to train, until now crf_layer do not support GPU +- \--log_period=500: print log every 20 batches. +- \--trainer_count=1: set thread number (or GPU count). +- \--show_parameter_stats_period=5000: show parameter statistic every 100 batches. +- \--save_dir=./output: output path to save models. +- \--num_passes=10000: set pass number, one pass in PaddlePaddle means training all samples in dataset one time. +- \--average_test_period=10000000: do test on average parameter every average_test_period batches +- \--init_model_path=./data: parameter initialization path +- \--load_missing_parameter_strategy=rand: random initialization unexisted parameters +- \--test_all_data_in_one_period=1: test all data in one period + + +After training, the models will be saved in directory `output`. Our training curve is as following: +
+![pic](./curve.jpg) +
+ +### Run testing +The script for testing is `test.sh`, user just need to execute: +```bash + ./test.sh +``` +The main part in `tesh.sh` +``` +paddle train \ + --config=./db_lstm.py \ + --model_list=$model_list \ + --job=test \ + --config_args=is_test=1 \ +``` + + - \--config=./db_lstm.py: network config file + - \--model_list=$model_list.list: model list file + - \--job=test: indicate the test job + - \--config_args=is_test=1: flag to indicate test + - \--test_all_data_in_one_period=1: test all data in 1 period + + +### Run prediction +The script for prediction is `predict.sh`, user just need to execute: +```bash + ./predict.sh + +``` +In `predict.sh`, user should offer the network config file, model path, label file, word dictionary file, feature file +``` +python predict.py + -c $config_file \ + -w $best_model_path \ + -l $label_file \ + -p $predicate_dict_file \ + -d $dict_file \ + -i $input_file \ + -o $output_file +``` + +`predict.py` is the main executable python script, which includes functions: load model, load data, data prediction. The network model will output the probability distribution of labels. In the demo, we take the label with maximum probability as result. User can also implement the beam search or viterbi decoding upon the probability distribution matrix. + +After prediction, the result is saved in `predict.res`. + +## Reference +[1] Martha Palmer, Dan Gildea, and Paul Kingsbury. The Proposition Bank: An Annotated Corpus of Semantic Roles , Computational Linguistics, 31(1), 2005. + +[2] Zhou, Jie, and Wei Xu. "End-to-end learning of semantic role labeling using recurrent neural networks." Proceedings of the Annual Meeting of the Association for Computational Linguistics. 2015.