moreo
/
dante-verification
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

some cleaning

This commit is contained in:
Alejandro Moreo Fernandez 2020-04-01 10:40:34 +02:00
parent c0de604df7
commit b4796f4882
2 changed files with 30 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

50
src/author_identification.py
View File

@ -2,15 +2,6 @@ from sklearn.linear_model import LogisticRegression
from data.dante_loader import load_texts
from data.features import *
from model import AuthorshipVerificator, f1_from_counters
from sklearn.svm import LinearSVC, SVC
from util.color_visualization import color
# DONE: ngrams should contain punctuation marks according to Sapkota et al. [39] in the PAN 2015 overview
# (More recently, it was shown that character
# n-grams corresponding to word affixes and including punctuation marks are the most
# significant features in cross-topic authorship attribution [57].) #we have cancelled the
# TODO: inspect the impact of chi-squared correlations against positive-only (or positive and negative) correlations for feature selection
# TODO: sentence length (Mendenhall-style) ?
for epistola in [1]:
@ -37,7 +28,9 @@ for epistola in [1]:
if epistola==2:
path+='_interaEpistola'
positive, negative, pos_files, neg_files, ep_text = load_texts(path, positive_author=author, unknown_target='EpistolaXIII_{}.txt'.format(epistola))
positive, negative, pos_files, neg_files, ep_text = load_texts(
path, positive_author=author, unknown_target='EpistolaXIII_{}.txt'.format(epistola)
)
files = np.asarray(pos_files + neg_files)
if len(positive) < 2:
discarded+=1
@ -45,16 +38,20 @@ for epistola in [1]:
n_full_docs = len(positive) + len(negative)
feature_extractor = FeatureExtractor(function_words_freq='latin',
conjugations_freq='latin',
features_Mendenhall=True,
features_sentenceLengths=True,
tfidf_feat_selection_ratio=0.1,
wordngrams=True, n_wordngrams=(1, 2),
charngrams=True, n_charngrams=(3, 4, 5),
preserve_punctuation=False,
split_documents=True, split_policy=split_by_sentences, window_size=3,
normalize_features=True)
feature_extractor = FeatureExtractor(
function_words_freq='latin',
conjugations_freq='latin',
features_Mendenhall=True,
features_sentenceLengths=True,
tfidf_feat_selection_ratio=0.1,
wordngrams=True, n_wordngrams=(1, 2),
charngrams=True, n_charngrams=(3, 4, 5),
preserve_punctuation=False,
split_documents=True,
split_policy=split_by_sentences,
window_size=3,
normalize_features=True
)
Xtr,ytr,groups = feature_extractor.fit_transform(positive, negative)
@ -66,22 +63,23 @@ for epistola in [1]:
av = AuthorshipVerificator(nfolds=10, estimator=LogisticRegression)
av.fit(Xtr,ytr,groups)
score_ave, score_std, tp, fp, fn, tn = av.leave_one_out(Xtr, ytr, files, groups, test_lowest_index_only=True, counters=True)
# print('LOO[full-docs]={:.3f} +-{:.5f}'.format(score_ave, score_std))
score_ave, score_std, tp, fp, fn, tn = av.leave_one_out(
Xtr, ytr, files, groups, test_lowest_index_only=True, counters=True
)
f1_scores.append(f1_from_counters(tp, fp, fn, tn))
counters.append((tp, fp, fn, tn))
print('F1 for {} = {:.3f}'.format(author,f1_scores[-1]))
print('F1 for {author} = {f1_scores[-1]:.3f}')
print('Computing macro- and micro-averages (discarded {}/{})'.format(discarded,len(authors)))
print(f'Computing macro- and micro-averages (discarded {discarded}/{len(authors)})')
f1_scores = np.array(f1_scores)
counters = np.array(counters)
macro_f1 = f1_scores.mean()
micro_f1 = f1_from_counters(*counters.sum(axis=0).tolist())
print('Macro-F1 = {:.3f}'.format(macro_f1))
print('Micro-F1 = {:.3f}'.format(micro_f1))
print(f'Macro-F1 = {macro_f1:.3f}')
print(f'Micro-F1 = {micro_f1:.3f}')
print()

7
src/data/features.py
View File

@ -229,6 +229,11 @@ def _features_tfidf(documents, tfidf_vectorizer=None, min_df = 1, ngrams=(1,1)):
return features, tfidf_vectorizer
# We have implemented ngrams extration generically, following Sapkota et al. (ref [39] in the PAN 2015 overview), i.e.,
# containing punctuation marks. However, this does not apply to this study since punctuation marks are filtered-out in
# editions of Latin texts.
# More recently, it was shown that character n-grams corresponding to word affixes and including punctuation
# marks are the most significant features in cross-topic authorship attribution [57].
def _features_ngrams(documents, ns=[4, 5], ngrams_vectorizer=None, min_df = 10, preserve_punctuation=True):
doc_ngrams = ngrams_extractor(documents, ns, preserve_punctuation)
return _features_tfidf(doc_ngrams, tfidf_vectorizer=ngrams_vectorizer, min_df = min_df)
@ -507,6 +512,7 @@ class FeatureExtractor:
if __name__=='__main__':
from collections import Counter
@ -594,4 +600,3 @@ if __name__=='__main__':
dots=False
print(counter)
print('rows',rows)