import nltk nltk.download('punkt') nltk.download('averaged_perceptron_tagger') nltk.download('wordnet') import owlready2 as OWL import nltk import sklearn import tensorflow as tf import numpy as np import os import datetime import hyperopt import pickle import json import torch from pytorch_pretrained_bert import BertTokenizer, BertModel from DataReader import Reader class ALDONAr(): def __init__(self, FLAGS): self.ontology = OWL.get_ontology(FLAGS.ontology_path) self.ontology.base_iri = FLAGS.ontology_path self.ontology = self.ontology.load() self.polarity_categories = {} self.polarity_categories['positive'] = self.ontology.search(iri='*Positive')[0] self.polarity_categories['negative'] = self.ontology.search(iri='*Negative')[0] self.type1, self.type2, self.type3 = {}, {}, {} classes = set(self.ontology.classes()) self.classes_dict = {onto_class: onto_class.lex for onto_class in classes} self.classesIntoTypes(classes) self.FLAGS = FLAGS self.tokenizer = BertTokenizer.from_pretrained(self.FLAGS.model_type) self.bert = BertModel.from_pretrained(self.FLAGS.model_type) self.bert.eval() self.d = self.getVectorLength() self.UNUSED = 0 def getVectorLength(self): d = '[CLS] [SEP]' d = self.tokenizer.tokenize(d) d = self.tokenizer.convert_tokens_to_ids(d) d = torch.tensor([d]) with torch.no_grad(): encoded_layers, _ = self.bert(d) return encoded_layers[0].shape[2] def getIndices(self, sentence, aspect): index = sentence.find(aspect) left = list(range(len(sentence[:index].split()))) aspect = list(range(len(left), len(left) + len(aspect.split()))) right = list(range(len(left) + len(aspect), len(sentence.split()))) return left, aspect, right def aggLayers(self, sentence_all): if self.FLAGS.agg_type == 'first': return sentence_all[0,:,:,:] elif self.FLAGS.agg_type == 'second_to_last': return sentence_all[-2,:,:,:] elif self.FLAGS.agg_type == 'last': return sentence_all[-1,:,:,:] elif self.FLAGS.agg_type == 'sum': return sentence_all.sum(axis=0) elif self.FLAGS.agg_type == 'sum_last_four': return sentence_all[-4:,:,:,:].sum(axis=0) elif self.FLAGS.agg_type == 'avg': return sentence_all.mean(axis=0) elif self.FLAGS.agg_type == 'avg_last_four': return sentence_all[-4:,:,:,:].mean(axis=0) else: raise NotImplemented('{} is not implemented'.format(self.FLAGS.agg_type)) def embedSentence(self, sentence): marked_sentence = '[CLS] ' + sentence + ' [SEP]' tokenized_sentence = self.tokenizer.tokenize(marked_sentence) indexed_tokens = self.tokenizer.convert_tokens_to_ids(tokenized_sentence) tokens_tensor = torch.tensor([indexed_tokens]) with torch.no_grad(): encoded_layers, _ = self.bert(tokens_tensor) shape = [len(encoded_layers), 1, encoded_layers[0].shape[1]-2, self.d] sentence_all = np.zeros(shape) for i, layer in enumerate(encoded_layers): sentence_all[i,:,:,:] = layer.cpu()[:,1:-1,:] sentence = self.aggLayers(sentence_all) return sentence def getEmbeddings(self, data): final = np.zeros([len(data), self.FLAGS.N, self.d]) E_LC = np.zeros_like(final) E_A = np.zeros_like(final) E_RC = np.zeros_like(final) polarities = [] for i, (sentence, aspect, polarity) in enumerate(data): for word in aspect.split(): if word not in self.tokenizer.vocab.keys(): self.tokenizer.vocab[word] = self.tokenizer.vocab.pop(f'[unused{self.UNUSED}]') self.UNUSED += 1 for word in sentence.split(): if word not in self.tokenizer.vocab.keys(): self.tokenizer.vocab[word] = self.tokenizer.vocab.pop(f'[unused{self.UNUSED}]') self.UNUSED += 1 indices = self.getIndices(sentence, aspect) sentence_embedded = self.embedSentence(sentence) final[i, :sentence_embedded.shape[1], :] = sentence_embedded.reshape([-1, self.d]) E_LC[i, indices[0], :] = final[i, indices[0], :] E_A[i, indices[1], :] = final[i, indices[1], :] E_RC[i, indices[2], :] = final[i, indices[2], :] polarities.append(self.transformPolarity(polarity)) return E_LC, E_A, E_RC, np.array(polarities) def transformPolarity(self, polarity): if polarity == 'negative': return np.array([1,0,0]) elif polarity == 'neutral': return np.array([0,1,0]) elif polarity == 'positive': return np.array([0,0,1]) else: Exception(polarity) def classesIntoTypes(self, classes): remove_words = ['property', 'mention', 'positive', 'neutral', 'negative'] for ont_class in classes: class_name = ont_class.__name__.lower() if any(word in class_name for word in remove_words): continue names = [x.__name__ for x in ont_class.ancestors()] names.sort() for name in names: if 'Generic' in name: self.type1[class_name] = ont_class break elif any(x in name for x in ['Positive', 'Negative']): self.type2[class_name] = ont_class break elif 'PropertyMention' in name: self.type3[class_name] = ont_class break def getClassPolarity(self, word_lemma_class, negated, type3): positive, negative = False, False if type3: OWL.sync_reasoner(debug=False) # To set relations of all newly created classes if self.polarity_categories['positive'].__subclasscheck__(word_lemma_class): if negated: positive = False negative = True else: positive = True if self.polarity_categories['negative'].__subclasscheck__(word_lemma_class): if negated: positive = True negative = False else: negative = True return positive, negative def categoryMatch(self, aspect_class, word_class): if aspect_class is None: return False aspect_mentions, word_mentions = [], [] for ancestor in aspect_class.ancestors(): if 'Mention' in ancestor.__name__: aspect_mentions.append(ancestor.__name__.rsplit('Mention',1)[0]) for ancestor in word_class.ancestors(): if 'Mention' in ancestor.__name__: word_mentions.append(ancestor.__name__.rsplit('Mention',1)[0]) common = set(aspect_mentions).intersection(set(word_mentions)) # If they have more than 2 ancestors in common (ontology.Mention, ontology.EntityMention and something else) if len(common) > 2: return True else: return False def addSubclass(self, word_class, aspect_class): class_name = word_class.__name__ + aspect_class.__name__ new_class = OWL.types.new_class(class_name, (word_class, aspect_class)) self.type3[new_class.__name__.lower()] = new_class return new_class def isNegated(self, word, words_in_sentence): negation = ({"not","no","never","isnt","arent","wont","wasnt","werent", "havent","hasnt", "nt", "cant", "couldnt", "dont", "doesnt"}) negated = False index = words_in_sentence.index(word) check = set(words_in_sentence[max(index-3,0):index]) if check.intersection(negation): negated = True return negated def predictSentiment(self, sentence, aspect): lemmatizer = nltk.WordNetLemmatizer() positive_list, negative_list = [], [] sentence_classes = {} words_in_sentence = sentence.split() aspect_class = None for word, tag in np.array(nltk.pos_tag(nltk.word_tokenize(aspect))): if tag.startswith("V"): aspect_lemma = lemmatizer.lemmatize(word, "v") # Verb elif tag.startswith("J"): aspect_lemma = lemmatizer.lemmatize(word, "a") # Adjective elif tag.startswith("R"): aspect_lemma = lemmatizer.lemmatize(word, "r") # Adverb else: aspect_lemma = lemmatizer.lemmatize(word) # Other words do not change for ont_class in list(self.classes_dict.values()): if aspect_lemma in ont_class: aspect_class = list(self.classes_dict.keys())[list(self.classes_dict.values()).index(ont_class)] for word, tag in np.array(nltk.pos_tag(nltk.word_tokenize(sentence))): if tag.startswith("V"): word_lemma = lemmatizer.lemmatize(word, "v") # Verb elif tag.startswith("J"): word_lemma = lemmatizer.lemmatize(word, "a") # Adjective elif tag.startswith("R"): word_lemma = lemmatizer.lemmatize(word, "r") # Adverb else: word_lemma = lemmatizer.lemmatize(word) # Other words do not change for ont_class in list(self.classes_dict.values()): if word_lemma in ont_class: word_class = list(self.classes_dict.keys())[list(self.classes_dict.values()).index(ont_class)] sentence_classes[word] = word_class if word == aspect: aspect_class = word_class is_negated = self.isNegated(word, words_in_sentence) if word_lemma in self.type1: positive, negative = self.getClassPolarity(word_class, is_negated, False) positive_list.append(positive) negative_list.append(negative) elif word_lemma in self.type2: if self.categoryMatch(aspect_class, word_class): positive, negative = self.getClassPolarity(word_class, is_negated, False) positive_list.append(positive) negative_list.append(negative) elif word_lemma in self.type3: if (aspect_class != word_class) and (aspect_class is not None): new_class = self.addSubclass(word_class, aspect_class) positive, negative = self.getClassPolarity(new_class, is_negated, True) positive_list.append(positive) negative_list.append(negative) if (True in positive_list) and (True not in negative_list): prediction = np.array([[0,0,1]]) elif (True not in positive_list) and (True in negative_list): prediction = np.array([[1,0,0]]) else: prediction = None return prediction def getMinibatches(self, batch_size, E_LC, E_A, E_RC, Y): with tf.name_scope('Mini_batch'): batch_size = tf.cond(batch_size > tf.cast(tf.shape(Y)[0], tf.int64), lambda: tf.cast(tf.shape(Y)[0], tf.int64), lambda: batch_size) df = tf.data.Dataset.from_tensor_slices((E_LC, E_A, E_RC, Y)).repeat().batch(batch_size) iterator = df.make_initializable_iterator() next_batch = iterator.get_next() nr_batches = tf.identity(tf.cast(tf.math.ceil(tf.shape(E_LC)[0]/tf.cast(batch_size, tf.int32)), tf.int64), name='nr_batches') return iterator, next_batch, nr_batches def getAccuracy(self, pred, y): return tf.metrics.accuracy(labels=tf.math.argmax(y,1), predictions=tf.cast(tf.math.argmax(pred,1), tf.int64), name='accuracy') def getLoss(self, logits, y): L2 = tf.reduce_sum([tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'bias' not in v.name and 'b_' not in v.name]) * self.FLAGS.regularization return tf.identity(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=y) + L2, name='loss') def getLossOp(self, loss): return tf.metrics.mean(loss, name='loss_op') def getSummaries(self, logits, prediction, Y): with tf.name_scope('Summaries'): loss = self.getLoss(logits, Y) loss_log, loss_log_update_op = self.getLossOp(loss) loss_scalar = tf.summary.scalar('loss', loss_log) accuracy, accuracy_update_op = self.getAccuracy(prediction, Y) accuracy_scalar = tf.summary.scalar('accuracy', accuracy) return loss, loss_log, loss_log_update_op, loss_scalar, accuracy, accuracy_update_op, accuracy_scalar def initWeights(self): with tf.name_scope('weights'): weights = {\ ### Bidirectional GRU 'W_fw_L' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_fw_L'), 'W_bw_L' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_bw_L'), 'b_bi_L' : tf.get_variable(initializer=tf.zeros([1, self.d]),dtype=tf.float32, name='b_bi_L'), 'W_fw_R' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_fw_R'), 'W_bw_R' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_bw_R'), 'b_bi_R' : tf.get_variable(initializer=tf.zeros([1, self.d]), dtype=tf.float32, name='b_bi_R'), ### MLP Beta 'W_1' : tf.get_variable(initializer=tf.random_normal([1, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_1'), 'b_1' : tf.get_variable(initializer=tf.zeros([1,1]), dtype=tf.float32, name='b_1'), 'b_l' : tf.get_variable(initializer=0.5, trainable=False, dtype=tf.float32, name='b_l'), 'W_2' : tf.get_variable(initializer=tf.random_normal([1, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_2'), 'b_2' : tf.get_variable(initializer=tf.zeros([1,1]), dtype=tf.float32, name='b_2'), 'b_r' : tf.get_variable(initializer=0.5, trainable=False, dtype=tf.float32, name='b_r'), ### Sentence Level Content Attention Module 'W_3' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.m, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_3'), 'W_4' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.m, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_4'), 'W_5' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.m, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_5'), 'W_6' : tf.get_variable(initializer=tf.random_normal([1, self.FLAGS.m], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_6'), 'b_3' : tf.get_variable(initializer=tf.zeros([1, self.FLAGS.m]), dtype=tf.float32, name='b_3'), # # ### Classification Module 'W_7' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_7'), 'W_8' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_8'), 'b_4' : tf.get_variable(initializer=tf.zeros([1, self.d]), dtype=tf.float32, name='b_4'), 'W_9' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_9'), 'W_10' : tf.get_variable(initializer=tf.random_normal([self.d, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_10'), 'b_5' : tf.get_variable(initializer=tf.zeros([1, self.d]), dtype=tf.float32, name='b_5'), 'W_11' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.k, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_11'), 'W_12' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.k, self.d], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_12'), 'b_6' : tf.get_variable(initializer=tf.zeros([1, self.FLAGS.k]), dtype=tf.float32, name='b_6'), # # # ### CNN layer 'W_conv1' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.kernel_size, self.FLAGS.k, self.FLAGS.q], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_conv1'), 'b_conv1' : tf.get_variable(initializer=tf.zeros([1, self.FLAGS.q]), dtype=tf.float32, name='b_conv1'), 'W_conv2' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.kernel_size, self.FLAGS.q, self.FLAGS.q], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_conv2'), 'b_conv2' : tf.get_variable(initializer=tf.zeros([1, self.FLAGS.q]), dtype=tf.float32, name='b_conv2'), # # ### Linear Layer 'W_13' : tf.get_variable(initializer=tf.random_normal([self.FLAGS.nr_cat, self.FLAGS.q], mean=self.FLAGS.mean, stddev=self.FLAGS.stddev, seed=self.FLAGS.seed), dtype=tf.float32, name='W_13'), 'b_7' : tf.get_variable(initializer=tf.zeros([1, self.FLAGS.nr_cat]), dtype=tf.float32, name='b_7')} return weights def getBeta(self, H_LS_masked, H_RS_masked, mask_left, mask_aspect, mask_right, mask_left_with_aspect, mask_right_with_aspect, weights): with tf.name_scope('MLP_beta'): left = tf.reshape(mask_left[:,:,0], [-1, self.FLAGS.N, 1]) aspect = tf.reshape(mask_aspect[:,:,0], [-1, self.FLAGS.N, 1]) right = tf.reshape(mask_right[:,:,0], [-1, self.FLAGS.N, 1]) left_with_aspect = tf.reshape(mask_left_with_aspect[:,:,0], [-1, self.FLAGS.N, 1]) right_with_aspect = tf.reshape(mask_right_with_aspect[:,:,0], [-1, self.FLAGS.N, 1]) beta_LS = tf.identity(tf.nn.sigmoid(tf.where(left_with_aspect, tf.tensordot(H_LS_masked, weights['W_1'], [2, 1]) + weights['b_1'], -1e10*tf.ones_like(left_with_aspect, dtype=tf.float32))) + weights['b_l'], name='beta_LS') # (S x N x 1) beta_RS = tf.identity(tf.nn.sigmoid(tf.where(right_with_aspect, tf.tensordot(H_RS_masked, weights['W_2'], [2, 1]) + weights['b_2'], -1e10*tf.ones_like(right_with_aspect, dtype=tf.float32))) + weights['b_r'], name='beta_RS') # (S x N x 1) beta_LC = tf.where(left, beta_LS, tf.zeros_like(left, dtype=tf.float32), name='beta_LC') # (S x N x 1) beta_A = tf.where(aspect, (tf.where(aspect, beta_LS, tf.zeros_like(aspect, dtype=tf.float32)) + tf.where(aspect, beta_RS, tf.zeros_like(aspect, dtype=tf.float32))) / 2, tf.zeros_like(aspect, dtype=tf.float32), name='beta_A') # (S x N x 1) beta_RC = tf.where(right, beta_RS, tf.zeros_like(right, dtype=tf.float32), name='beta_RC') # (S x N x 1) beta = tf.identity(tf.reshape(beta_LC + beta_A + beta_RC, [-1, self.FLAGS.N]), name='beta') # (S x N) return beta def getGRUBi(self, side, E, length, sentence_length, mask_left_with_aspect, mask_right_with_aspect, dropout_prob, weights): with tf.name_scope('GRU_bi'): W_fw = 'W_fw_L' if side == 'left' else 'W_fw_R' W_bw = 'W_bw_L' if side == 'left' else 'W_bw_R' b_bi = 'b_bi_L' if side == 'left' else 'b_bi_R' with tf.variable_scope(side) as scope: cell = tf.nn.rnn_cell.GRUCell(self.d) cell = tf.nn.rnn_cell.DropoutWrapper(cell, output_keep_prob=1-dropout_prob, input_keep_prob=1-dropout_prob, state_keep_prob=1-dropout_prob, seed=self.FLAGS.seed, dtype=tf.float32) output, _ = tf.nn.bidirectional_dynamic_rnn(cell, cell, E, sequence_length=length, dtype=tf.float32, scope=scope) H_fw, H_bw = output # (S x N x d) H = tf.nn.tanh(tf.tensordot(H_fw, weights[W_fw], [2, 1]) + tf.tensordot(H_bw, weights[W_bw], [2, 1]) + weights[b_bi]) # (S x N x d) if side == 'left': H_masked = tf.where(mask_left_with_aspect, tf.reverse_sequence(H, length, seq_axis=1, batch_axis=0), tf.zeros_like(mask_left_with_aspect, dtype=tf.float32)) # (S x N x d) else: H_masked = tf.where(mask_right_with_aspect, tf.reverse_sequence(tf.reverse_sequence(H, length, seq_axis=1, batch_axis=0), sentence_length, seq_axis=1, batch_axis=0), tf.zeros_like(mask_right_with_aspect, dtype=tf.float32)) # (S x N x d) return H_masked def model(self, E, E_LC, E_A, E_RC, dropout_prob, weights): with tf.name_scope('Context_Attention_Module'): mask_left = tf.not_equal(E_LC, tf.constant(0.0)) mask_left_with_aspect = tf.not_equal(E_LC+E_A, tf.constant(0.0)) mask_aspect = tf.not_equal(E_A, tf.constant(0.0)) mask_right = tf.not_equal(E_RC, tf.constant(0.0)) mask_right_with_aspect = tf.not_equal(E_RC+E_A, tf.constant(0.0)) mask_sentence = tf.not_equal(E, tf.constant(0.0)) left_length = tf.reduce_sum(tf.cast(mask_left_with_aspect[:,:,0], tf.int32), 1) # (S) right_length = tf.reduce_sum(tf.cast(mask_right_with_aspect[:,:,0], tf.int32), 1) # (S) sentence_length = tf.reduce_sum(tf.cast(mask_sentence[:,:,0], tf.int32), 1) # (S) E_LS_r = tf.reverse_sequence(E_LC+E_A, left_length, seq_axis=1, batch_axis=0) # (S x N x d) E_RS_r = tf.reverse_sequence(tf.reverse_sequence(E_RC+E_A, sentence_length, seq_axis=1, batch_axis=0), right_length, seq_axis=1, batch_axis=0) # (S x N x d) H_LS = tf.identity(self.getGRUBi('left', E_LS_r, left_length, sentence_length, mask_left_with_aspect, mask_right_with_aspect, dropout_prob, weights), name='H_LS') # (S x N x d) H_RS = tf.identity(self.getGRUBi('right', E_RS_r, right_length, sentence_length, mask_left_with_aspect, mask_right_with_aspect, dropout_prob, weights), name='H_RS') # (S x N x d) beta = self.getBeta(H_LS, H_RS, mask_left, mask_aspect, mask_right, mask_left_with_aspect, mask_right_with_aspect, weights) # (S x N) with tf.name_scope('context_attention_weighted_memory'): M_w = tf.identity(tf.tile(tf.expand_dims(beta, 2),[1,1,self.d]) * E, name='M_w') # (S x N x d) with tf.name_scope('Sentence_Level_Content_Attention_Module'): with tf.name_scope('scores_of_words'): v_a = tf.identity(tf.reduce_sum(E_A, 1) / tf.math.count_nonzero(E_A, 1, dtype=tf.float32), name='v_a') # (S x d) v_s = tf.identity(tf.reduce_sum(E, 1) / tf.math.count_nonzero(E, 1, dtype=tf.float32), name='v_s') # (S x d) v_a_matrix = tf.tile(tf.expand_dims(v_a, 1), [1, self.FLAGS.N, 1], name='v_a_matrix') # (S x N x d) v_s_matrix = tf.tile(tf.expand_dims(v_s, 1), [1, self.FLAGS.N, 1], name='v_s_matrix') # (S x N x d) C = tf.where(tf.not_equal(E, tf.constant(0.0))[:,:,0], tf.reshape(tf.tensordot(tf.nn.tanh( tf.tensordot(tf.nn.dropout(M_w, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_3'], [2,1]) + tf.tensordot(tf.nn.dropout(v_a_matrix, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_4'], [2,1]) + tf.tensordot(tf.nn.dropout(v_s_matrix, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_5'], [2,1]) + tf.nn.dropout(weights['b_3'], rate=dropout_prob, seed=self.FLAGS.seed)), weights['W_6'], [2, 1]), [-1, self.FLAGS.N]), -1e10*tf.ones([tf.shape(E)[0], self.FLAGS.N]), name='C') # (S x N) with tf.name_scope('attention_weights_for_scores'): A = tf.nn.softmax(C, name='A') # (S x N) with tf.name_scope('weighted_embedding_vector'): v_we = tf.transpose(tf.matmul(tf.transpose(M_w, [0,2,1]), tf.expand_dims(A, 2)), [0,2,1]) # (S x 1 x d) with tf.name_scope('Classification_Module'): v_aw = tf.nn.dropout(tf.nn.tanh( tf.tensordot(tf.nn.dropout(tf.reshape(v_s, [-1, 1, self.d]), rate=dropout_prob, seed=self.FLAGS.seed), weights['W_7'], [2,1]) + tf.tensordot(tf.nn.dropout(v_we, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_8'], [2,1]) + weights['b_4']), rate=dropout_prob, seed=self.FLAGS.seed, name='v_aw') # (S x 1 x d) v_sw = tf.nn.dropout(tf.nn.tanh( tf.tensordot(tf.nn.dropout(tf.reshape(v_a, [-1, 1, self.d]), rate=dropout_prob, seed=self.FLAGS.seed), weights['W_9'], [2,1]) + tf.tensordot(tf.nn.dropout(v_we, rate=dropout_prob, seed=self.FLAGS.seed), weights['W_10'], [2,1]) + weights['b_5']), rate=dropout_prob, seed=self.FLAGS.seed, name='v_sw') # (S x 1 x d) v_o = tf.nn.dropout(tf.nn.tanh(tf.tensordot(v_aw, weights['W_11'], [2,1]) + tf.tensordot(v_sw, weights['W_12'], [2,1]) + weights['b_6']), rate=dropout_prob, seed=self.FLAGS.seed, name='v_o') # (S x 1 x k)) with tf.name_scope('CNN_layer'): conv1 = tf.nn.dropout(tf.nn.tanh(tf.nn.conv1d(v_o, weights['W_conv1'], stride=1, padding='SAME', use_cudnn_on_gpu=True) + weights['b_conv1']), rate=dropout_prob, seed=self.FLAGS.seed, name='conv1') # (S x 256) conv2 = tf.nn.dropout(tf.nn.tanh(tf.nn.conv1d(conv1, weights['W_conv2'], stride=1, padding='SAME', use_cudnn_on_gpu=True) + weights['b_conv2']), rate=dropout_prob, seed=self.FLAGS.seed, name='conv2') # (S x 256) with tf.name_scope('Linear_Layer'): v_L = tf.reshape(tf.tensordot(conv2, weights['W_13'], [2,1]) + weights['b_7'], [-1, self.FLAGS.nr_cat], name='v_L') # (S x nr_cat) return v_L def trainModel(self, data, data_test=None, save_to_files=False): predictions_ont = np.zeros([1,3]) data_DBGRU = [] data_ont = [] for sentence, aspect, polarity in data: prediction = self.predictSentiment(sentence, aspect) if prediction is not None: predictions_ont = np.concatenate([predictions_ont, prediction]) data_ont.append([sentence, aspect, polarity]) else: data_DBGRU.append([sentence, aspect, polarity]) predictions_ont = predictions_ont[1:,:] y_ont = np.array(list(map(self.transformPolarity, [y for _, _ , y in data_ont]))) e_lc, e_a, e_rc, y = self.getEmbeddings(data) e_lc_DBGRU, e_a_DBGRU, e_rc_DBGRU, y_DBGRU = self.getEmbeddings(data_DBGRU) if data_test is not None: predictions_ont_test = np.zeros([1,3]) data_DBGRU_test = [] data_ont_test = [] for sentence, aspect, polarity in data_test: prediction_test = self.predictSentiment(sentence, aspect) if prediction_test is not None: predictions_ont_test = np.concatenate([predictions_ont_test, prediction_test]) data_ont_test.append([sentence, aspect, polarity]) else: data_DBGRU_test.append([sentence, aspect, polarity]) predictions_ont_test = predictions_ont_test[1:,:] y_ont_test = np.array(list(map(self.transformPolarity, [y for _, _ , y in data_ont_test]))) e_lc_DBGRU_test, e_a_DBGRU_test, e_rc_DBGRU_test, y_DBGRU_test = self.getEmbeddings(data_DBGRU_test) if len(e_lc_DBGRU) > 0 or len(e_lc_DBGRU_test[0]) > 0: tf.reset_default_graph() Y = tf.placeholder(tf.int64, [None, 3], name='Y') Y_ont = tf.placeholder(tf.int64, [None, 3], name='Y_ont') prediction_ont = tf.placeholder(tf.float32, [None, 3], name='prediction_ont') batch_size = tf.placeholder(tf.int64, name='batch_size') dropout_prob = tf.placeholder(tf.float32, name='dropout_prob') with tf.name_scope('Embeddings'): E_LC = tf.placeholder(tf.float32, [None, self.FLAGS.N, self.d], name='E_LC') E_A = tf.placeholder(tf.float32, [None, self.FLAGS.N, self.d], name='E_A') E_RC = tf.placeholder(tf.float32, [None, self.FLAGS.N, self.d], name='E_RC') E = tf.identity(E_LC + E_A + E_RC, name='E') iterator, next_batch, nr_batches = self.getMinibatches(batch_size, E_LC, E_A, E_RC, Y) weights = self.initWeights() logits = self.model(E, E_LC, E_A, E_RC, dropout_prob, weights) prediction = tf.nn.softmax(logits, name='prediction') loss = self.getLoss(logits, Y) with tf.name_scope('TrainOp'): global_step = tf.Variable(0, trainable=False, name='global_step') optimizer = tf.train.AdamOptimizer(self.FLAGS.learning_rate, self.FLAGS.beta1, self.FLAGS.beta2) train_op = optimizer.minimize(loss, global_step=global_step) # Checking whether Ont has classified anything logits_stacked = tf.identity(tf.cond(tf.not_equal(tf.reduce_sum(Y_ont), tf.constant(-3, dtype=tf.int64)), lambda: tf.concat([prediction_ont, logits], axis=0), lambda: logits), name='logits_stacked') prediction_stacked = tf.identity(tf.cond(tf.not_equal(tf.reduce_sum(Y_ont), tf.constant(-3, dtype=tf.int64)), lambda: tf.concat([prediction_ont, prediction], axis=0), lambda: prediction), name='prediction_stacked') Y_stacked = tf.identity(tf.cond(tf.not_equal(tf.reduce_sum(Y_ont), tf.constant(-3, dtype=tf.int64)), lambda: tf.concat([Y_ont, Y], axis=0), lambda: Y), name='Y_stacked') _, loss_log, loss_log_update_op, loss_scalar, accuracy, \ accuracy_update_op, accuracy_scalar = self.getSummaries(logits_stacked, prediction_stacked, Y_stacked) with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess: if save_to_files: writer_train = tf.summary.FileWriter('./Results/logs/{0}/train/{1}'.format(self.__class__.__name__, datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S')), sess.graph) writer_test = tf.summary.FileWriter('./Results/logs/{0}/test/{1}'.format(self.__class__.__name__, datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S')), sess.graph) summary_op = tf.summary.merge_all() saver = tf.train.Saver(max_to_keep = 3, save_relative_paths=True) sess.run(tf.global_variables_initializer()) batches, _ = sess.run([nr_batches, iterator.initializer], feed_dict={'Embeddings/E_LC:0':e_lc, 'Embeddings/E_A:0':e_a, 'Embeddings/E_RC:0':e_rc, 'Y:0':y, 'batch_size:0':self.FLAGS.batch_size}) acc_test_best = 0 loss_test_best = np.inf epoch_best = -1 for epoch in range(self.FLAGS.num_epochs): sess.run(tf.local_variables_initializer()) for _ in range(batches): e_lc_batch, e_a_batch, e_rc_batch, y_batch = sess.run(next_batch) sess.run(train_op, feed_dict={'Embeddings/E_LC:0':e_lc_batch, 'Embeddings/E_A:0':e_a_batch, 'Embeddings/E_RC:0':e_rc_batch, 'Y:0':y_batch, 'dropout_prob:0':self.FLAGS.dropout_prob}) step = sess.run(global_step) if len(e_lc_DBGRU) > 0: sess.run([accuracy_update_op, loss_log_update_op], feed_dict={'Embeddings/E_LC:0':e_lc_DBGRU, 'Embeddings/E_A:0':e_a_DBGRU, 'Embeddings/E_RC:0':e_rc_DBGRU, 'Y:0':y_DBGRU, 'Y_ont:0':y_ont if len(y_ont)>0 else np.array([[-1,-1,-1]]), 'prediction_ont:0':predictions_ont if len(predictions_ont)>0 else np.array([[-1,-1,-1]]), 'dropout_prob:0':self.FLAGS.dropout_prob}) acc, cost = sess.run([accuracy, loss_log]) if save_to_files: writer_train.add_summary(sess.run(summary_op), step) writer_train.flush() else: acc = sklearn.metrics.accuracy_score(np.argmax(predictions_ont, 1), np.argmax(y_ont, 1)) cost = sklearn.metrics.log_loss(y_ont, predictions_ont) if len(e_lc_DBGRU_test[0] > 0): sess.run(tf.local_variables_initializer()) sess.run([accuracy_update_op, loss_log_update_op], feed_dict={'Embeddings/E_LC:0':e_lc_DBGRU_test, 'Embeddings/E_A:0':e_a_DBGRU_test, 'Embeddings/E_RC:0':e_rc_DBGRU_test, 'Y:0':y_DBGRU_test, 'Y_ont:0':y_ont_test if len(y_ont_test)>0 else np.array([[-1,-1,-1]]), 'prediction_ont:0':predictions_ont_test if len(predictions_ont_test)>0 else np.array([[-1,-1,-1]]), 'dropout_prob:0':0.0}) acc_test, cost_test = sess.run([accuracy, loss_log]) if save_to_files: writer_test.add_summary(sess.run(loss_scalar), step) writer_test.add_summary(sess.run(accuracy_scalar), step) writer_test.flush() else: acc_test = sklearn.metrics.accuracy_score(np.argmax(predictions_ont_test, 1), np.argmax(y_ont_test, 1)) cost_test = sklearn.metrics.log_loss(y_ont_test, predictions_ont_test) if acc_test > acc_test_best: acc_test_best = acc_test loss_test_best = cost_test epoch_best = epoch if save_to_files: if not os.path.exists('./Results/ckpt/{}'.format(self.__class__.__name__)): os.makedirs('./Results/ckpt/{}'.format(self.__class__.__name__)) saver.save(sess, './Results/ckpt/{}/model.ckpt'.format(self.__class__.__name__), step) improved = '*' else: improved = '' print('Step: {0:>6}, Train-Batch Accuracy: {1:>6.1%}, Test Accuracy: {2:>6.1%} Train-Batch Loss: {3} Test Loss: {4} {5}'. format(step, acc, acc_test, np.round(cost,5), np.round(cost_test,5), improved)) if epoch_best + 1000 <= epoch: print('Stopping in step: {}, the best result in epoch: {}'.format(step, epoch_best)) break if save_to_files: writer_train.close() writer_test.close() else: acc = sklearn.metrics.accuracy_score(np.argmax(predictions_ont, 1), np.argmax(y_ont, 1)) cost = sklearn.metrics.log_loss(y_ont, predictions_ont) if data_test is not None: acc_test_best = sklearn.metrics.accuracy_score(np.argmax(predictions_ont_test, 1), np.argmax(y_ont_test, 1)) cost_test_best = sklearn.metrics.log_loss(y_ont_test, predictions_ont_test) print('Train Accuracy: {0:>6.1%}, Test Accuracy: {1:>6.1%} Train Loss: {2} Test Loss: {3}'. format(acc, acc_test_best, np.round(cost,5), np.round(cost_test_best,5))) else: print('Train Accuracy: {0:>6.1%}, Train Loss: {1}'.format(acc, np.round(cost,5))) return acc_test_best, loss_test_best def predict(self, data): predictions_ont = np.zeros([1,3]) data_DBGRU = [] data_ont = [] for sentence, aspect, polarity in data: prediction = self.predictSentiment(sentence, aspect) if prediction is not None: predictions_ont = np.concatenate([predictions_ont, prediction]) data_ont.append([sentence, aspect, polarity]) else: data_DBGRU.append([sentence, aspect, polarity]) predictions_ont = predictions_ont[1:,:] y_ont = np.array(list(map(self.transformPolarity, [y for _, _ , y in data_ont]))) e_lc, e_a, e_rc, y = self.getEmbeddings(data_DBGRU) if len(e_lc) > 0: tf.reset_default_graph() with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess: ckpt = tf.train.latest_checkpoint('./Results/ckpt/{}/'.format(self.__class__.__name__)) print('./Results/ckpt/{}/'.format(self.__class__.__name__)) print(ckpt) saver = tf.train.import_meta_graph(ckpt+'.meta') sess.run(tf.global_variables_initializer()) saver.restore(sess, ckpt) graph = tf.get_default_graph() global_step = graph.get_tensor_by_name('TrainOp/global_step:0') Y = graph.get_tensor_by_name('Y_stacked:0') logits = graph.get_tensor_by_name('logits_stacked:0') prediction = graph.get_tensor_by_name('prediction_stacked:0') _, accuracy = self.getAccuracy(prediction, Y) _, loss = self.getLossOp(self.getLoss(logits, Y)) sess.run(tf.local_variables_initializer()) A = graph.get_tensor_by_name('Sentence_Level_Content_Attention_Module/attention_weights_for_scores/A:0') A_scores, predictions, acc, cost, step = sess.run([A, prediction, accuracy, loss, global_step], feed_dict={'Embeddings/E_LC:0':e_lc, 'Embeddings/E_A:0':e_a, 'Embeddings/E_RC:0':e_rc, 'Y:0':y, 'Y_ont:0':y_ont if len(y_ont)>0 else np.array([[-1,-1,-1]]), 'prediction_ont:0':predictions_ont if len(predictions_ont)>0 else np.array([[-1,-1,-1]]), 'dropout_prob:0':0.0}) print('Step: {0:>6}, Accuracy: {1:>6.1%}, Loss: {2}'. format(step, acc, np.round(cost,5))) return A_scores, predictions else: acc = sklearn.metrics.accuracy_score(np.argmax(predictions_ont, 1), np.argmax(y_ont, 1)) cost = sklearn.metrics.log_loss(y_ont, predictions_ont) print('Accuracy: {0:>6.1%}, Loss: {1}'.format(acc, np.round(cost,5))) return predictions_ont if __name__ == '__main__': np.set_printoptions(threshold=np.inf, edgeitems=3, linewidth=120) FLAGS = tf.app.flags.FLAGS for name in list(FLAGS): if name not in ('showprefixforinfo',): delattr(FLAGS, name) tf.app.flags.DEFINE_string('f', '', 'kernel') tf.app.flags.DEFINE_boolean('logtostderr', True, 'tensorboard') tf.app.flags.DEFINE_float('train_proportion', 1.0, 'Train proportion for train/validation split') #0.75 tf.app.flags.DEFINE_boolean('shuffle', True, 'Shuffle datasets') tf.app.flags.DEFINE_integer('num_epochs', 600, 'Number of iterations') tf.app.flags.DEFINE_integer('seed', None, 'Random seed') tf.app.flags.DEFINE_float('mean', 0.0, 'Mean of normally initialized variables') tf.app.flags.DEFINE_float('stddev', 0.15, 'Standard deviation of normally initialized variables') tf.app.flags.DEFINE_float('dropout_prob', 0.3, 'Dropout probability') tf.app.flags.DEFINE_float('learning_rate', 0.001, 'Adam optimzier learning rate') tf.app.flags.DEFINE_float('beta1', 0.9, 'Adam optimzer beta1') tf.app.flags.DEFINE_float('beta2', 0.999, 'Adam optimizer beta2') tf.app.flags.DEFINE_integer('batch_size', 128, 'Batch size') tf.app.flags.DEFINE_integer('nr_cat', 3, 'Number of classification categories') tf.app.flags.DEFINE_float('regularization', 0.001, 'L2 regularization term') tf.app.flags.DEFINE_integer('m', 300, 'm') tf.app.flags.DEFINE_integer('q', 256, 'q') tf.app.flags.DEFINE_integer('kernel_size', 2, 'kernel_size') tf.app.flags.DEFINE_integer('k', 128, 'k') tf.app.flags.DEFINE_string('model_type', 'bert-large-uncased', 'BERT LARGE model') tf.app.flags.DEFINE_string('agg_type', 'sum_last_four', 'Aggregation type of BERT layers') tf.app.flags.DEFINE_boolean('plot_attention', True, 'Plot sentence attention') tf.app.flags.DEFINE_boolean('train_model', True, 'Train ALDONAr') tf.app.flags.DEFINE_boolean('save_to_files', True, 'Whether to save checkpoints') tf.app.flags.DEFINE_boolean('predict_values', True, 'Load and predict with ALDONAr') tf.app.flags.DEFINE_string('ontology_path', './Ontology/Ontology_restaurants.owl', 'Ontology path') tf.app.flags.DEFINE_string('train_data_path', './Data/ABSA-16_SB1_Restaurants_Train_Data.xml', 'Train data path') tf.app.flags.DEFINE_string('test_data_path', './Data/ABSA-16_SB1_Restaurants_Test_Gold.xml', 'Test data path') reader = Reader(FLAGS) data_train, longest_sentence_train = reader.readData(FLAGS.train_data_path) reader.aspectCategories('Train 2015') reader.aspectCategories('Train 2016') data_test, longest_sentence_test = reader.readData(FLAGS.test_data_path) reader.aspectCategories('Test 2015') reader.aspectCategories('Test 2016') reader.plotPolarityDistribution(data_train, data_test, print_results=True, plot_results=False) longest_sentence = max(longest_sentence_train, longest_sentence_test) tf.app.flags.DEFINE_integer('N', longest_sentence, 'Length of the longest sentence') model = ALDONAr(FLAGS) model.longest_sentence = longest_sentence if FLAGS.train_proportion != 1: data_train, data_test = reader.splitTrainData(data_train) ## Train and evaluate models if FLAGS.train_model: print('Training...') a = datetime.datetime.now() model.trainModel(data_train, data_test, save_to_files=FLAGS.save_to_files) b = datetime.datetime.now() print('Time:', b-a) ## Predict using trained models if FLAGS.predict_values: print('Prediction...') attention, pred = model.predict(data_test) print('') ## Visualize attention if FLAGS.plot_attention: sentence = 'do not get me started on the margaritas either' aspect = 'margaritas' polarity = 'negative' data_attention = np.array([[sentence, aspect, polarity]]) att, _ = model.predict(data_attention) reader.plotAttention(att, sentence.split(), './Results/attention.png')