Experimentation with further improving feature and relation extraction

ADA-X/.gitignore

@@ -5,6 +5,7 @@ server/agent/amazon_data/
 server/agent/SA/data/
 server/agent/target_extraction/data/
 server/agent/target_extraction/BERT/data/
+server/agent/target_extraction/eval/qa/
 .DS_Store
 *.pickle
 *.wv
\ No newline at end of file

ADA-X/server/agent/target_extraction/BERT/entity_extractor/bert_entity_extractor.py

@@ -10,7 +10,7 @@ from sklearn import metrics
 import statistics
 from transformers import get_linear_schedule_with_warmup
 from agent.target_extraction.BERT.entity_extractor.entity_dataset import EntityDataset, generate_batch, generate_production_batch
-from agent.target_extraction.BERT.entity_extractor.entitybertnet import NUM_CLASSES, EntityBertNet
+from agent.target_extraction.BERT.entity_extractor.entitybertnet import NUM_CLASSES, EntityBertNet, BATCH_SIZE
 
 device = torch.device('cuda')
 
@@ -21,7 +21,6 @@ MAX_GRAD_NORM = 1.0
 
 # training
 N_EPOCHS = 3
-BATCH_SIZE = 32
 WARM_UP_FRAC = 0.05
 
 # loss
@@ -61,8 +60,7 @@ class BertEntityExtractor:
         else:
             train_size = int(size * (1 - valid_frac)) if size is not None else None
             train_data, _ = EntityDataset.from_file(file_path, size=train_size)
-            valid_size = int(size * valid_frac) if size is not None else int(len(train_data) * valid_frac)
-            valid_data, _ = EntityDataset.from_file(valid_file_path, size=valid_size)
+            valid_data, _ = EntityDataset.from_file(valid_file_path)
         train_loader = DataLoader(train_data, batch_size=BATCH_SIZE, shuffle=True, num_workers=4,
                                   collate_fn=generate_batch)
 
@@ -119,11 +117,11 @@ class BertEntityExtractor:
 
             print('epoch done')
 
+            torch.save(self.net.state_dict(), '{}_epoch_{}.pt'.format(save_file, epoch_idx + 1))
+
             if valid_data is not None:
                 self.evaluate(data=valid_data)
 
-        torch.save(self.net.state_dict(), '{}.pt'.format(save_file))
-
         end = time.time()
         print('Training took', end - start, 'seconds')
 
@@ -207,3 +205,7 @@ class BertEntityExtractor:
                     probs[ins.entity].append(score)
 
         return {t: statistics.mean(t_probs) if len(t_probs) > 0 else None for t, t_probs in probs.items()}
+
+
+BertEntityExtractor.train_and_validate('all_reviews_features.tsv', 'feature_extractor',
+                                       valid_file_path='annotated_watch_review_features.tsv')

ADA-X/server/agent/target_extraction/BERT/entity_extractor/entity_dataset.py

@@ -8,58 +8,22 @@ import os.path
 from agent.target_extraction.BERT.relation_extractor.pairbertnet import TRAINED_WEIGHTS, HIDDEN_OUTPUT_FEATURES
 
 MAX_SEQ_LEN = 128
-LABELS = ['ASPECT', 'NAN']
-LABEL_MAP = {'ASPECT': 1, 'NAN': 0, None: None}
 MASK_TOKEN = '[MASK]'
 tokenizer = BertTokenizer.from_pretrained(TRAINED_WEIGHTS)
 
 
-def generate_batch(batch):
-    encoded = tokenizer.batch_encode_plus([instance.tokens for instance in batch], add_special_tokens=True,
-                                          max_length=MAX_SEQ_LEN, pad_to_max_length=True, is_pretokenized=True,
-                                          return_tensors='pt')
-    input_ids = encoded['input_ids']
-    attn_mask = encoded['attention_mask']
-    labels = torch.tensor([instance.label for instance in batch])
-
-    entity_indices = indices_for_entity_ranges([instance.entity_range for instance in batch])
-
-    return input_ids, attn_mask, entity_indices, labels
-
-
-def generate_production_batch(batch):
-    encoded = tokenizer.batch_encode_plus([instance.tokens for instance in batch], add_special_tokens=True,
-                                          max_length=MAX_SEQ_LEN, pad_to_max_length=True, is_pretokenized=True,
-                                          return_tensors='pt')
-    input_ids = encoded['input_ids']
-    attn_mask = encoded['attention_mask']
-
-    entity_indices = indices_for_entity_ranges([instance.entity_range for instance in batch])
-
-    return input_ids, attn_mask, entity_indices, batch
-
-
-def indices_for_entity_ranges(ranges):
-    max_e_len = max(end - start for start, end in ranges)
-    indices = torch.tensor([[[min(t, end)] * HIDDEN_OUTPUT_FEATURES
-                             for t in range(start, start + max_e_len + 1)]
-                            for start, end in ranges])
-    return indices
-
-
 class EntityDataset(Dataset):
 
-    def __init__(self, df, size=None):
-        # filter inapplicable rows
-        self.df = df[df.apply(lambda x: EntityDataset.instance_from_row(x) is not None, axis=1)]
-
+    def __init__(self, df, training=True, size=None):
+        self.df = df
+        self.training = training
         # sample data if a size is specified
         if size is not None and size < len(self):
             self.df = self.df.sample(size, replace=False)
 
     @staticmethod
-    def from_df(df, size=None):
-        dataset = EntityDataset(df, size=size)
+    def for_extraction(df):
+        dataset = EntityDataset(df, training=False)
         print('Obtained dataset of size', len(dataset))
         return dataset
 
@@ -83,80 +47,60 @@ class EntityDataset(Dataset):
             print('Obtained train set of size', len(dataset), 'and validation set of size', len(validset))
             return dataset, validset
 
-    @staticmethod
-    def instance_from_row(row):
-        unpacked_arr = literal_eval(row['entityMentions']) if type(row['entityMentions']) is str else row['entityMentions']
-        rms = [rm for rm in unpacked_arr if 'label' not in rm or rm['label'] in LABELS]
-        if len(rms) == 1:
-            entity, label = rms[0]['text'], (rms[0]['label'] if 'label' in rms[0] else None)
-        else:
-            return None  # raise AttributeError('Instances must have exactly one relation')
-
-        text = row['sentText']
-        return EntityDataset.get_instance(text, entity, label=label)
-
-    @staticmethod
-    def get_instance(text, entity, label=None):
-        tokens = tokenizer.tokenize(text)
-
-        i = 0
-        found_entity = False
-        entity_range = None
-        while i < len(tokens):
-            match_length = EntityDataset.token_entity_match(i, entity.lower(), tokens)
-            if match_length is not None:
-                if found_entity:
-                    return None  # raise AttributeError('Entity {} appears twice in text {}'.format(entity, text))
-                found_entity = True
-                tokens[i:i + match_length] = [MASK_TOKEN] * match_length
-                entity_range = (i + 1, i + match_length)  # + 1 taking into account the [CLS] token
-                i += match_length
-            else:
-                i += 1
-
-        if found_entity:
-            return PairRelInstance(tokens, entity, entity_range, LABEL_MAP[label], text)
+    def instance_from_row(self, row):
+        if self.training:
+            return EntityInstance(literal_eval(row['tokens']),
+                                  row['entity_idx'],
+                                  label=row['label'])
         else:
-            return None
-
-    @staticmethod
-    def token_entity_match(first_token_idx, entity, tokens):
-        token_idx = first_token_idx
-        remaining_entity = entity
-        while remaining_entity:
-            if remaining_entity == entity or remaining_entity.lstrip() != remaining_entity:
-                # start of new word
-                remaining_entity = remaining_entity.lstrip()
-                if token_idx < len(tokens) and tokens[token_idx] == remaining_entity[:len(tokens[token_idx])]:
-                    remaining_entity = remaining_entity[len(tokens[token_idx]):]
-                    token_idx += 1
-                else:
-                    break
-            else:
-                # continuing same word
-                if (token_idx < len(tokens) and tokens[token_idx].startswith('##')
-                        and tokens[token_idx][2:] == remaining_entity[:len(tokens[token_idx][2:])]):
-                    remaining_entity = remaining_entity[len(tokens[token_idx][2:]):]
-                    token_idx += 1
-                else:
-                    break
-        if remaining_entity:
-            return None
-        else:
-            return token_idx - first_token_idx
+            return EntityInstance(row['tokens'],
+                                  row['entity_idx'],
+                                  entity=row['entity'])
 
     def __len__(self):
         return len(self.df.index)
 
     def __getitem__(self, idx):
-        return EntityDataset.instance_from_row(self.df.iloc[idx])
+        return self.instance_from_row(self.df.iloc[idx])
 
 
-class PairRelInstance:
+class EntityInstance:
 
-    def __init__(self, tokens, entity, entity_range, label, text):
+    def __init__(self, tokens, entity_idx, label=None, entity=None):
         self.tokens = tokens
-        self.entity = entity
-        self.entity_range = entity_range
+        self.entity_idx = entity_idx
         self.label = label
-        self.text = text
+        self.entity = entity
+
+
+def generate_batch(instances: [EntityInstance]):
+    encoded = tokenizer.batch_encode_plus([instance.tokens for instance in instances], add_special_tokens=True,
+                                          max_length=MAX_SEQ_LEN, pad_to_max_length=True, is_pretokenized=True,
+                                          return_tensors='pt')
+    input_ids = encoded['input_ids']
+    attn_mask = encoded['attention_mask']
+
+    entity_indices = torch.tensor([instance.entity_idx for instance in instances])
+    labels = torch.tensor([instance.label for instance in instances])
+
+    return input_ids, attn_mask, entity_indices, labels
+
+
+def generate_production_batch(instances: [EntityInstance]):
+    encoded = tokenizer.batch_encode_plus([instance.tokens for instance in instances], add_special_tokens=True,
+                                          max_length=MAX_SEQ_LEN, pad_to_max_length=True, is_pretokenized=True,
+                                          return_tensors='pt')
+    input_ids = encoded['input_ids']
+    attn_mask = encoded['attention_mask']
+
+    entity_indices = torch.tensor([instance.entity_idx for instance in instances])
+
+    return input_ids, attn_mask, entity_indices, instances
+
+
+# def indices_for_entity_ranges(ranges):
+#     max_e_len = max(end - start for start, end in ranges)
+#     indices = torch.tensor([[[min(t, end)] * HIDDEN_OUTPUT_FEATURES
+#                              for t in range(start, start + max_e_len + 1)]
+#                             for start, end in ranges])
+#     return indices

ADA-X/server/agent/target_extraction/BERT/entity_extractor/entitybertnet.py

@@ -5,6 +5,7 @@ from transformers import *
 HIDDEN_OUTPUT_FEATURES = 768
 TRAINED_WEIGHTS = 'bert-base-uncased'
 NUM_CLASSES = 2  # entity, not entity
+BATCH_SIZE = 32
 
 
 class EntityBertNet(nn.Module):
@@ -20,14 +21,9 @@ class EntityBertNet(nn.Module):
         bert_output, _ = self.bert_base(input_ids=input_ids, attention_mask=attn_mask)
 
         # max pooling at entity locations
-        entity_pooled_output = EntityBertNet.pooled_output(bert_output, entity_indices)
+        entity_pooled_output = bert_output[torch.arange(0, bert_output.shape[0]), entity_indices]
 
         # fc layer (softmax activation done in loss function)
         x = self.fc(entity_pooled_output)
         return x
 
-    @staticmethod
-    def pooled_output(bert_output, indices):
-        outputs = torch.gather(bert_output, dim=1, index=indices)
-        pooled_output, _ = torch.max(outputs, dim=1)
-        return pooled_output

ADA-X/server/agent/target_extraction/BERT/relation_extractor/bert_rel_extractor.py

@@ -8,8 +8,10 @@ import time
 import numpy as np
 from sklearn import metrics
 from transformers import get_linear_schedule_with_warmup
-from agent.target_extraction.BERT.relation_extractor.pair_rel_dataset import PairRelDataset, generate_batch, generate_production_batch
-from agent.target_extraction.BERT.relation_extractor.pairbertnet import NUM_CLASSES, PairBertNet
+# from agent.target_extraction.BERT.relation_extractor.pair_rel_dataset import PairRelDataset, generate_batch, generate_production_batch
+from agent.target_extraction.BERT.relation_extractor.rel_dataset import PairRelDataset, generate_batch, generate_production_batch, RelInstance
+# from agent.target_extraction.BERT.relation_extractor.pairbertnet import NUM_CLASSES, PairBertNet
+from agent.target_extraction.BERT.relation_extractor.relbertnet import NUM_CLASSES, RelBertNet
 
 device = torch.device('cuda')
 
@@ -30,12 +32,12 @@ loss_criterion = CrossEntropyLoss()
 class BertRelExtractor:
 
     def __init__(self):
-        self.net = PairBertNet()
+        self.net = RelBertNet()
 
     @staticmethod
     def load_saved(path):
         extr = BertRelExtractor()
-        extr.net = PairBertNet()
+        extr.net = RelBertNet()
         extr.net.load_state_dict(torch.load(path))
         extr.net.eval()
         return extr
@@ -60,8 +62,7 @@ class BertRelExtractor:
         else:
             train_size = int(size * (1 - valid_frac)) if size is not None else None
             train_data, _ = PairRelDataset.from_file(file_path, size=train_size)
-            valid_size = int(size * valid_frac) if size is not None else int(len(train_data) * valid_frac)
-            valid_data, _ = PairRelDataset.from_file(valid_file_path, size=valid_size)
+            valid_data, _ = PairRelDataset.from_file(valid_file_path)
         train_loader = DataLoader(train_data, batch_size=BATCH_SIZE, shuffle=True, num_workers=4,
                                   collate_fn=generate_batch)
 
@@ -87,16 +88,16 @@ class BertRelExtractor:
 
             for batch_idx, batch in enumerate(train_loader):
                 # send batch to gpu
-                input_ids, attn_mask, masked_indices, fst_indices, snd_indices, target_labels = tuple(i.to(device) for i in batch)
+                input_ids, attn_mask, entity_indices, entity_mask, labels = tuple(i.to(device) for i in batch)
 
                 # zero param gradients
                 optimiser.zero_grad()
 
                 # forward pass
-                output_scores = self.net(input_ids, attn_mask, masked_indices, fst_indices, snd_indices)
+                output_scores = self.net(input_ids, attn_mask, entity_indices, entity_mask)
 
                 # backward pass
-                loss = loss_criterion(output_scores, target_labels)
+                loss = loss_criterion(output_scores, labels)
                 loss.backward()
 
                 # clip gradient norm
@@ -117,12 +118,11 @@ class BertRelExtractor:
                     batch_loss = 0.0
 
             print('epoch done')
+            torch.save(self.net.state_dict(), '{}_epoch_{}.pt'.format(save_file, epoch_idx + 1))
 
             if valid_data is not None:
                 self.evaluate(data=valid_data)
 
-        torch.save(self.net.state_dict(), '{}.pt'.format(save_file))
-
         end = time.time()
         print('Training took', end - start, 'seconds')
 
@@ -147,15 +147,14 @@ class BertRelExtractor:
         with torch.no_grad():
             for batch in test_loader:
                 # send batch to gpu
-                input_ids, attn_mask, masked_indices, fst_indices, snd_indices, target_labels = tuple(i.to(device)
-                                                                                                      for i in batch)
+                input_ids, attn_mask, entity_indices, entity_mask, labels = tuple(i.to(device) for i in batch)
 
                 # forward pass
-                output_scores = self.net(input_ids, attn_mask, masked_indices, fst_indices, snd_indices)
+                output_scores = self.net(input_ids, attn_mask, entity_indices, entity_mask)
                 _, output_labels = torch.max(output_scores.data, 1)
 
                 outputs += output_labels.tolist()
-                targets += target_labels.tolist()
+                targets += labels.tolist()
 
         assert len(outputs) == len(targets)
 
@@ -176,25 +175,24 @@ class BertRelExtractor:
         recall = metrics.recall_score(targets, outputs, average=None)
         print('recall:', recall)
 
-    def extract_single_relation(self, text, e1, e2):
-        ins = PairRelDataset.get_instance(text, e1, e2)
-        input_ids, attn_mask, masked_indices, prod_indices, feat_indices, instances = generate_production_batch([ins])
+    def extract_single_relation(self, text, entities):
+        ins = RelInstance.from_sentence(text, entities)
+        input_ids, attn_mask, entity_indices, entity_mask, _ = generate_production_batch([ins])
 
         self.net.cuda()
         self.net.eval()
 
         with torch.no_grad():
             # send batch to gpu
-            input_ids, attn_mask, masked_indices, prod_indices, feat_indices = tuple(i.to(device) for i in
-                                                                                     [input_ids, attn_mask,
-                                                                                      masked_indices, prod_indices,
-                                                                                      feat_indices])
+            input_ids, attn_mask, entity_indices, entity_mask = tuple(i.to(device) for i in [input_ids, attn_mask,
+                                                                                             entity_indices,
+                                                                                             entity_mask])
 
             # forward pass
-            output_scores = softmax(self.net(input_ids, attn_mask, masked_indices, prod_indices, feat_indices), dim=1)
+            output_scores = softmax(self.net(input_ids, attn_mask, entity_indices, entity_mask), dim=1)
             _, output_labels = torch.max(output_scores.data, 1)
 
-            print(instances[0].get_relation_for_label(output_labels[0]))
+            ins.print_results_for_labels(output_labels)
 
     def extract_relations(self, n_aspects, aspect_index_map, aspect_counts, file_path=None, dataset=None, size=None):
         # load data
@@ -215,15 +213,14 @@ class BertRelExtractor:
         count_matrix = np.zeros((n_aspects, n_aspects))
 
         with torch.no_grad():
-            for input_ids, attn_mask, masked_indices, prod_indices, feat_indices, instances in loader:
+            for input_ids, attn_mask, prod_indices, feat_indices, instances in loader:
                 # send batch to gpu
-                input_ids, attn_mask, masked_indices, prod_indices, feat_indices = tuple(i.to(device) for i in
-                                                                                         [input_ids, attn_mask,
-                                                                                          masked_indices, prod_indices,
-                                                                                          feat_indices])
+                input_ids, attn_mask, prod_indices, feat_indices = tuple(i.to(device) for i in [input_ids, attn_mask,
+                                                                                                prod_indices,
+                                                                                                feat_indices])
 
                 # forward pass
-                output_scores = softmax(self.net(input_ids, attn_mask, masked_indices, prod_indices, feat_indices), dim=1)
+                output_scores = softmax(self.net(input_ids, attn_mask, prod_indices, feat_indices), dim=1)
                 rel_scores = output_scores.narrow(1, 1, 2)
 
                 for ins, scores in zip(instances, rel_scores.tolist()):
@@ -236,4 +233,38 @@ class BertRelExtractor:
 
         return prob_matrix, count_matrix
 
+    def extract_relations2(self, n_aspects, dataset):
+        loader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=4,
+                            collate_fn=generate_production_batch)
+
+        self.net.cuda()
+        self.net.eval()
+
+        prob_matrix = np.zeros((n_aspects, n_aspects))
+        count_matrix = np.zeros((n_aspects, n_aspects))
+
+        with torch.no_grad():
+            for input_ids, attn_mask, entity_indices, combination_indices, instances in loader:
+                # send batch to gpu
+                input_ids, attn_mask, entity_indices, combination_indices = tuple(i.to(device) for i in
+                                                                                  [input_ids, attn_mask,
+                                                                                   entity_indices, combination_indices])
+