cothec / Corneferencer

Commit ba27568ff1602427739fbebaf45ce90e08b504ff

Authored by Bartłomiej Nitoń
1 parent 4a097c4f
master

Minor fixes.

Inline Side-by-side
Showing 5 changed files with 24 additions and 26 deletions
... ... @@ -20,4 +20,4 @@ W2V_MODEL_PATH = os.path.join(os.path.dirname(__file__), 'models', W2V_MODEL_NAM
20 20 W2V_MODEL = Word2Vec.load(W2V_MODEL_PATH)
21 21  
22 22 NEURAL_MODEL_PATH = os.path.join(os.path.dirname(__file__), 'models', NEURAL_MODEL_NAME)
23   -NEURAL_MODEL = initialize_neural_model(NUMBER_OF_FEATURES)
  23 +NEURAL_MODEL = initialize_neural_model(NUMBER_OF_FEATURES, NEURAL_MODEL_PATH)
... ...
corneferencer/entities.py
  View file @ba27568
... ... @@ -22,6 +22,7 @@ class Mention:
22 22 start_in_words, end_in_words):
23 23 self.id = mnt_id
24 24 self.set = ''
  25 + self.old_set = ''
25 26 self.text = text
26 27 self.lemmatized_text = lemmatized_text
27 28 self.words = words
... ...
corneferencer/resolvers/resolve.py
  View file @ba27568
  1 +import numpy
  2 +
1 3 from conf import NEURAL_MODEL, THRESHOLD
2   -from corneferencer.resolvers.vectors import create_pair_vector
  4 +from corneferencer.resolvers.vectors import get_pair_vector
3 5  
4 6  
5 7 # incremental resolve algorithm
6 8 def incremental(text):
7   - last_set_id = 1
  9 + last_set_id = 0
8 10 for i, ana in enumerate(text.mentions):
9 11 if i > 0:
10 12 best_prediction = 0.0
11 13 best_ante = None
12   - for ante in text.mentions[:i:-1]:
13   - pair_vec = create_pair_vector(ante, ana)
14   - prediction = NEURAL_MODEL.predict(pair_vec)
15   - accuracy = prediction[0]
16   - if accuracy > THRESHOLD and accuracy > best_prediction:
17   - best_prediction = accuracy
  14 + for ante in text.mentions[:i]:
  15 + pair_vec = get_pair_vector(ante, ana)
  16 + sample = numpy.asarray([pair_vec], dtype=numpy.float32)
  17 + prediction = NEURAL_MODEL.predict(sample)[0]
  18 + if prediction > THRESHOLD and prediction >= best_prediction:
  19 + best_prediction = prediction
18 20 best_ante = ante
19 21 if best_ante is not None:
20 22 if best_ante.set:
... ... @@ -29,7 +31,7 @@ def incremental(text):
29 31 # entity based resolve algorithm
30 32 def entity_based(text):
31 33 sets = []
32   - last_set_id = 1
  34 + last_set_id = 0
33 35 for i, ana in enumerate(text.mentions):
34 36 if i > 0:
35 37 best_fit = get_best_set(sets, ana)
... ... @@ -64,13 +66,13 @@ def get_best_set(sets, ana):
64 66  
65 67  
66 68 def predict_set(mentions, ana):
67   - accuracy_sum = 0.0
  69 + prediction_sum = 0.0
68 70 for mnt in mentions:
69   - pair_vec = create_pair_vector(mnt, ana)
70   - prediction = NEURAL_MODEL.predict(pair_vec)
71   - accuracy = prediction[0]
72   - accuracy_sum += accuracy
73   - return accuracy_sum / float(len(mentions))
  71 + pair_vec = get_pair_vector(mnt, ana)
  72 + sample = numpy.asarray([pair_vec], dtype=numpy.float32)
  73 + prediction = NEURAL_MODEL.predict(sample)[0]
  74 + prediction_sum += prediction
  75 + return prediction_sum / float(len(mentions))
74 76  
75 77  
76 78 def remove_singletons(sets):
... ...
corneferencer/resolvers/vectors.py
  View file @ba27568
1   -import numpy
2   -
3 1 from corneferencer.resolvers import features
4 2  
5   -# input_1 to have shape (None, 1126) but got array with shape (1126, 1)
6   -def create_pair_vector(ante, ana):
  3 +
  4 +def get_pair_vector(ante, ana):
7 5 vec = []
8   - # ante_features = get_mention_features(ante)
9   - # vec.extend(ante_features)
10   - # ana_features = get_mention_features(ana)
11   - # vec.extend(ana_features)
12 6 vec.extend(ante.features)
13 7 vec.extend(ana.features)
14 8 pair_features = get_pair_features(ante, ana)
15 9 vec.extend(pair_features)
16   - return numpy.asarray([vec], dtype=numpy.float32)
  10 + return vec
17 11  
18 12  
19 13 def get_mention_features(mention):
... ...
corneferencer/utils.py
  View file @ba27568
... ... @@ -10,7 +10,7 @@ def eprint(*args, **kwargs):
10 10 print(*args, file=sys.stderr, **kwargs)
11 11  
12 12  
13   -def initialize_neural_model(number_of_features):
  13 +def initialize_neural_model(number_of_features, path_to_model):
14 14 inputs = Input(shape=(number_of_features,))
15 15 output_from_1st_layer = Dense(1000, activation='relu')(inputs)
16 16 output_from_1st_layer = Dropout(0.5)(output_from_1st_layer)
... ... @@ -22,4 +22,5 @@ def initialize_neural_model(number_of_features):
22 22  
23 23 model = Model(inputs, output)
24 24 model.compile(optimizer='Adam', loss='binary_crossentropy', metrics=['accuracy'])
  25 + model.load_weights(path_to_model)
25 26 return model
... ...