def make_differing_words_bold(orig_sent, sys_sent, make_bold):
'''Returns the two sentences with differing words in bold'''
def format_words(words, mutual_words):
'''Makes all words bold except the mutual ones'''
words_generator = iter(words)
formatted_string = ''
for mutual_word in mutual_words:
word = next(words_generator)
bold_text = ''
while word != mutual_word:
bold_text += ' ' + word
word = next(words_generator)
if bold_text != '':
formatted_string += ' ' + make_bold(bold_text)
formatted_string += ' ' + word
# Add remaining words
formatted_string += ' ' + make_bold(' '.join(words_generator))
return formatted_string.strip()
orig_words = to_words(orig_sent)
sys_words = to_words(sys_sent)
mutual_words = get_lcs(orig_words, sys_words)
return format_words(orig_words, mutual_words), format_words(sys_words, mutual_words)
def truncate(sentence):
# Take first 80% words
words = to_words(sentence)
return ' '.join(words[:int(len(words) * 0.8)]) + '.'
def get_qualitative_html_examples(orig_sents, sys_sents, refs_sents):
title_key_print = [
('Randomly sampled simplifications',
lambda c, s, refs: 0,
lambda value: ''),
('Best simplifications according to SARI',
lambda c, s, refs: -corpus_sari([c], [s], [refs]),
lambda value: f'SARI={-value:.2f}'),
('Worst simplifications according to SARI',
lambda c, s, refs: corpus_sari([c], [s], [refs]),
lambda value: f'SARI={value:.2f}'),
('Simplifications with only one differing word',
lambda c, s, refs: -(count_words(c) == count_words(s) == len(get_lcs(to_words(c), to_words(s))) + 1),
lambda value: ''),
('Simplifications with the most compression',
lambda c, s, refs: get_compression_ratio(c, s),
lambda value: f'compression_ratio={value:.2f}'),
('Simplifications that are longer than the source',
lambda c, s, refs: -get_compression_ratio(c, s),
lambda value: f'compression_ratio={-value:.2f}'),
('Simplifications that paraphrase the source',
lambda c, s, refs: get_levenshtein_similarity(c, s) / get_compression_ratio(c, s),
lambda value: f'levenshtein_similarity={value:.2f}'),
('Simplifications that are the most similar to the source (excluding exact matches)',
lambda c, s, refs: -get_levenshtein_similarity(c, s) * int(c != s),
lambda value: f'levenshtein_similarity={-value:.2f}'),
('Simplifications with the most sentence splits (if there are any)',
lambda c, s, refs: -count_sentence_splits(c, s),
lambda value: f'nb_sentences_ratio={-value:.2f}'),
]
def get_one_sample_html(orig_sent, sys_sent, ref_sents, sort_key, print_func):
doc = Doc()
with doc.tag('div', klass='mb-2 p-1'):
# Sort key
with doc.tag('div', klass='text-muted small'):
doc.asis(print_func(sort_key(orig_sent, sys_sent, ref_sents)))
with doc.tag('div', klass='ml-2'):
orig_sent_bold, sys_sent_bold = make_differing_words_bold(orig_sent, sys_sent, make_text_bold_html)
# Source
with doc.tag('div'):
doc.asis(orig_sent_bold)
# Prediction
with doc.tag('div'):
doc.asis(sys_sent_bold)
# References
collapse_id = get_random_html_id()
with doc.tag('div', klass='position-relative'):
with doc.tag('a', ('data-toggle', 'collapse'), ('href', f'#{collapse_id}'),
klass='stretched-link small'):
doc.text('References')
with doc.tag('div', klass='collapse', id=collapse_id):
for ref_sent in refs:
_, ref_sent_bold = make_differing_words_bold(orig_sent, ref_sent, make_text_bold_html)
with doc.tag('div', klass='text-muted'):
doc.asis(ref_sent_bold)
return doc.getvalue()
doc = Doc()
for title, sort_key, print_func in title_key_print:
# stretched-link needs position-relative
with doc.tag('div', klass='container-fluid mt-4 p-2 position-relative border'):
doc.line('h3', klass='m-2', text_content=title)
# Make whole div clickable to collapse / uncollapse examples
collapse_id = get_random_html_id()
with doc.tag('a', ('data-toggle', 'collapse'), ('href', f'#{collapse_id}'), klass='stretched-link'):
pass # doc.stag and doc.line don't seem to work with stretched-link
# Now lets print the examples
sample_generator = sorted(
zip(orig_sents, sys_sents, zip(*refs_sents)),
key=lambda args: sort_key(*args),
)
# Samples displayed by default
with doc.tag('div', klass='collapse show', id=collapse_id):
n_samples = 10
for i, (orig_sent, sys_sent, refs) in enumerate(sample_generator):
if i >= n_samples:
break
doc.asis(get_one_sample_html(orig_sent, sys_sent, refs, sort_key, print_func))
return doc.getvalue()