computational-SCM/reproduce.py at main · katiefraser/computational-SCM · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
import pandas as pd
import numpy as np
import random
import os.path
import json
from sentence_transformers import SentenceTransformer
from sklearn.cross_decomposition import PLSRegression
from sklearn.decomposition import PCA
import pickle
from scipy.spatial.transform import Rotation as R

from train import get_embeddings, do_PLS, do_PCA, get_centroid, normalize, compute_rotation_matrix
from test import rotate_data, compute_warmth_competence, compute_1d_accuracy, compute_2d_accuracy, compute_accuracy


if __name__ == "__main__":

    #optimal sentence embedding model
    model_name = 'roberta-large-nli-mean-tokens' # can be any model here: https://www.sbert.net/docs/pretrained_models.html

    #comparison models
    #model_name = 'stsb-roberta-large'
    #model_name = 'paraphrase-distilroberta-base-v2'
    #model_name = 'average_word_embeddings_glove.840B.300d'
    #model_name = 'paraphrase-mpnet-base-v2'

    model = SentenceTransformer(model_name)

    # these files contain the training sentences for all folds
    train_df_one_adj = pd.read_csv('data/cross_validation/training_one_adjective.csv')
    train_df_two_adj = pd.read_csv('data/cross_validation/training_two_adjectives.csv')

    test_results = {}

    # four functionalities:
    functional_testing = ['basic_functionality', 'negation', 'semantic_composition', 'syntax']

    # six model configurations:
    for test_case in functional_testing:
        test_results[test_case] = {'original': {'none': [], 'PCA': [], 'PLS': []}, 'axis_rotated': {'none': [], 'PCA': [], 'PLS': []}}


    for fold in range(5):

        print('=== FOLD ' + str(fold) + ' ===')

        # use different random seed in each fold
        rseed = fold

        np.random.seed(seed=rseed)
        random.seed(rseed)

        ############
        # TRAINING #
        ############


        # get sentence embeddings for single-adjective training data - can skip if embeddings already exist
        if not os.path.isfile('embeddings/cross_validation/training_one_adjective_' + model_name + '_fold' + str(fold) + '.csv'):
            train_df = train_df_one_adj[train_df_one_adj['Fold'] == fold]
            train_df = get_embeddings(train_df, model)
            train_df.to_csv('embeddings/cross_validation/training_one_adjective_' + model_name + '_fold' + str(fold) + '.csv')
        else:
            print('Embeddings appear to exist, moving on ...')

        # get sentence embeddings for double-adjective training data - can skip if embeddings already exist
        if not os.path.isfile('embeddings/cross_validation/training_two_adjectives_' + model_name + '_fold' + str(fold) + '.csv'):
            train_df = train_df_two_adj[train_df_two_adj['Fold'] == fold]
            train_df = get_embeddings(train_df, model)
            train_df.to_csv('embeddings/cross_validation/training_two_adjectives_' + model_name + '_fold' + str(fold) + '.csv')
        else:
            print('Embeddings appear to exist, moving on ...')

        # read in embeddings for this fold
        df_single_adj = pd.read_csv('embeddings/cross_validation/training_one_adjective_' + model_name + '_fold' + str(fold) + '.csv')
        df_double_adj = pd.read_csv('embeddings/cross_validation/training_two_adjectives_' + model_name + '_fold' + str(fold) + '.csv')

        tot_df = pd.concat([df_single_adj, df_double_adj], ignore_index=True)
        for i, row in tot_df.iterrows():
            tot_df.at[i, 'Embeddings'] = json.loads(row['Embeddings'])

        # do PLS dimensionality reduction
        tot_df = do_PLS(tot_df, model_name)

        # do PCA dimensionality reduction
        tot_df = do_PCA(tot_df, model_name)

        # compute rotation matrices for each column in Table 2
        compute_rotation_matrix(tot_df, model_name, polar_model='original', PLS=False, PCA=False)
        compute_rotation_matrix(tot_df, model_name, polar_model='original', PLS=False, PCA=True)
        compute_rotation_matrix(tot_df, model_name, polar_model='original', PLS=True, PCA=False)

        compute_rotation_matrix(tot_df, model_name, polar_model='axis_rotated', PLS=False, PCA=False)
        compute_rotation_matrix(tot_df, model_name, polar_model='axis_rotated', PLS=False, PCA=True)
        compute_rotation_matrix(tot_df, model_name, polar_model='axis_rotated', PLS=True, PCA=False)


        #############
        #  TESTING  #
        #############

        for test_case in functional_testing:


            # check if the embeddings already exist; if not, generate them
            if not os.path.isfile('embeddings/cross_validation/testing_' + test_case + '_' + model_name + '_fold' + str(fold) + '.csv'):
                test_df = pd.read_csv('data/cross_validation/testing_' + test_case + '.csv')
                test_df = test_df[test_df['Fold'] == fold]
                test_df = get_embeddings(test_df, model)
                test_df.to_csv('embeddings/cross_validation/testing_' + test_case + '_' + model_name + '_fold' + str(fold) + '.csv')
            else:
                print('Embeddings appear to exist, moving on ...')

            # load embeddings for test data
            test_df = pd.read_csv('embeddings/cross_validation/testing_' + test_case + '_' + model_name + '_fold' + str(fold) + '.csv')
            for i, row in test_df.iterrows():
                test_df.at[i, 'Embeddings'] = json.loads(row['Embeddings'])

            ## Compute accuracy for each column in Table 2 ##

            test_df = compute_warmth_competence(test_df, model_name, polar_model='original', PLS=False, PCA=False)
            test_results[test_case]['original']['none'].append(compute_accuracy(test_df))

            test_df = compute_warmth_competence(test_df, model_name, polar_model='original', PLS=False, PCA=True)
            test_results[test_case]['original']['PCA'].append(compute_accuracy(test_df))

            test_df = compute_warmth_competence(test_df, model_name, polar_model='original', PLS=True, PCA=False)
            test_results[test_case]['original']['PLS'].append(compute_accuracy(test_df))

            test_df = compute_warmth_competence(test_df, model_name, polar_model='axis_rotated', PLS=False, PCA=False)
            test_results[test_case]['axis_rotated']['none'].append(compute_accuracy(test_df))

            test_df = compute_warmth_competence(test_df, model_name, polar_model='axis_rotated', PLS=False, PCA=True)
            test_results[test_case]['axis_rotated']['PCA'].append(compute_accuracy(test_df))

            test_df = compute_warmth_competence(test_df, model_name, polar_model='axis_rotated', PLS=True, PCA=False)
            test_results[test_case]['axis_rotated']['PLS'].append(compute_accuracy(test_df))


    ### RESULTS SUMMARY ###

    print()
    print('== Cross-validation for embedding model:', model_name, '==')
    print()

    for test_case in functional_testing:
        print('--', test_case, '--')

        for polar_model in ['original', 'axis_rotated']:
            for dim_reduction in ['none', 'PCA', 'PLS']:
                print('Model:', polar_model, '+', dim_reduction, ':', round(100*np.mean(test_results[test_case][polar_model][dim_reduction]), 1), '(' + str(round(100*np.std(test_results[test_case][polar_model][dim_reduction]), 1)) +')')