cothec / Corneferencer

Commit db32de75dc92a99c150c294da15ac46f9b0f850b

Authored by Bartłomiej Nitoń
1 parent 0d8e647d

Slightly different neural network model.

Inline Side-by-side
Showing 3 changed files with 12 additions and 10 deletions
... ... @@ -6,7 +6,6 @@ from gensim.models.word2vec import Word2Vec
6 6  
7 7  
8 8 CONTEXT = 5
9   -# THRESHOLD = 0.001
10 9 RANDOM_WORD_VECTORS = True
11 10 W2V_SIZE = 50
12 11 W2V_MODEL_NAME = 'w2v_allwiki_nkjpfull_50.model'
... ... @@ -14,7 +13,7 @@ W2V_MODEL_NAME = 'w2v_allwiki_nkjpfull_50.model'
14 13 # simple or siamese
15 14 NEURAL_MODEL_ARCHITECTURE = 'simple'
16 15 NUMBER_OF_FEATURES = 1190
17   -NEURAL_MODEL_NAME = 'model_1190_features.h5'
  16 +NEURAL_MODEL_NAME = 'model_with_singletons.h5'
18 17  
19 18 FREQ_LIST_NAME = 'base.lst'
20 19 LEMMA2SYNONYMS_NAME = 'lemma2synonyms.map'
... ...
corneferencer/resolvers/features.py
  View file @db32de7
... ... @@ -174,14 +174,14 @@ def distances_vec(ante, ana):
174 174 words_dist = [0] * 11
175 175 words_bucket = 0
176 176 if mnts_intersect != 1:
177   - words_bucket = get_distance_bucket(ana.start_in_words - ante.end_in_words - 1)
  177 + words_bucket = get_distance_bucket(ana.start_in_words - ante.end_in_words)
178 178 words_dist[words_bucket] = 1
179 179 vec.extend(words_dist)
180 180  
181 181 mentions_dist = [0] * 11
182 182 mentions_bucket = 0
183 183 if mnts_intersect != 1:
184   - mentions_bucket = get_distance_bucket(ana.position_in_mentions - ante.position_in_mentions - 1)
  184 + mentions_bucket = get_distance_bucket(ana.position_in_mentions - ante.position_in_mentions)
185 185 if words_bucket == 10:
186 186 mentions_bucket = 10
187 187 mentions_dist[mentions_bucket] = 1
... ...
corneferencer/utils.py
  View file @db32de7
... ... @@ -27,17 +27,20 @@ def initialize_neural_model(architecture, number_of_features, path_to_model):
27 27 def initialize_simple_model(number_of_features, path_to_model):
28 28 inputs = Input(shape=(number_of_features,))
29 29  
30   - output_from_1st_layer = Dense(1000, activation='relu')(inputs)
31   - output_from_1st_layer = Dropout(0.2)(output_from_1st_layer)
  30 + output_from_1st_layer = Dense(500)(inputs)
32 31 output_from_1st_layer = BatchNormalization()(output_from_1st_layer)
  32 + output_from_1st_layer = Activation('relu')(output_from_1st_layer)
  33 + output_from_1st_layer = Dropout(0.2)(output_from_1st_layer)
33 34  
34   - output_from_2nd_layer = Dense(500, activation='relu')(output_from_1st_layer)
35   - output_from_2nd_layer = Dropout(0.2)(output_from_2nd_layer)
  35 + output_from_2nd_layer = Dense(200)(output_from_1st_layer)
36 36 output_from_2nd_layer = BatchNormalization()(output_from_2nd_layer)
  37 + output_from_2nd_layer = Activation('relu')(output_from_2nd_layer)
  38 + output_from_2nd_layer = Dropout(0.2)(output_from_2nd_layer)
37 39  
38   - output_from_3rd_layer = Dense(300, activation='relu')(output_from_2nd_layer)
39   - output_from_3rd_layer = Dropout(0.2)(output_from_3rd_layer)
  40 + output_from_3rd_layer = Dense(100)(output_from_2nd_layer)
40 41 output_from_3rd_layer = BatchNormalization()(output_from_3rd_layer)
  42 + output_from_3rd_layer = Activation('relu')(output_from_3rd_layer)
  43 + output_from_3rd_layer = Dropout(0.2)(output_from_3rd_layer)
41 44  
42 45 output = Dense(1, activation='sigmoid')(output_from_3rd_layer)
43 46  
... ...