def sent_translation(self, to_log):
"""
Evaluation on sentence translation.
Only available on Europarl, for en - {de, es, fr, it} language pairs.
"""
lg1 = self.src_dico.lang
lg2 = self.tgt_dico.lang
# parameters
n_keys = 200000
n_queries = 2000
n_idf = 300000
# load europarl data
if not hasattr(self, 'europarl_data'):
self.europarl_data = load_europarl_data(
lg1, lg2, n_max=(n_keys + 2 * n_idf)
)
# if no Europarl data for this language pair
if not self.europarl_data:
return
# mapped word embeddings
src_emb = self.mapping(self.src_emb.weight).data
tgt_emb = self.tgt_emb.weight.data
# get idf weights
idf = get_idf(self.europarl_data, lg1, lg2, n_idf=n_idf)
for method in ['nn', 'csls_knn_10']:
# source <- target sentence translation
results = get_sent_translation_accuracy(
self.europarl_data,
self.src_dico.lang, self.src_dico.word2id, src_emb,
self.tgt_dico.lang, self.tgt_dico.word2id, tgt_emb,
n_keys=n_keys, n_queries=n_queries,
method=method, idf=idf
)
to_log.update([('tgt_to_src_%s-%s' % (k, method), v) for k, v in results])
# target <- source sentence translation
results = get_sent_translation_accuracy(
self.europarl_data,
self.tgt_dico.lang, self.tgt_dico.word2id, tgt_emb,
self.src_dico.lang, self.src_dico.word2id, src_emb,
n_keys=n_keys, n_queries=n_queries,
method=method, idf=idf
)
to_log.update([('src_to_tgt_%s-%s' % (k, method), v) for k, v in results])
def sent_translation(self, to_log):
"""
Evaluation on sentence translation.
Only available on Europarl, for en - {de, es, fr, it} language pairs.
"""
lg1 = self.src_dico.lang
lg2 = self.tgt_dico.lang
# parameters
n_keys = 200000
n_queries = 2000
n_idf = 300000
# load europarl data
if not hasattr(self, 'europarl_data'):
self.europarl_data = load_europarl_data(
lg1, lg2, n_max=(n_keys + 2 * n_idf)
)
# if no Europarl data for this language pair
if not self.europarl_data:
return
# mapped word embeddings
src_emb = self.mapping(self.src_emb.weight).data
tgt_emb = self.tgt_emb.weight.data
# get idf weights
idf = get_idf(self.europarl_data, lg1, lg2, n_idf=n_idf)
for method in ['nn', 'csls_knn_10']:
# source <- target sentence translation
results = get_sent_translation_accuracy(
self.europarl_data,
self.src_dico.lang, self.src_dico.word2id, src_emb,
self.tgt_dico.lang, self.tgt_dico.word2id, tgt_emb,
n_keys=n_keys, n_queries=n_queries,
method=method, idf=idf
)
to_log.update([('tgt_to_src_%s-%s' % (k, method), v) for k, v in results])
# target <- source sentence translation
results = get_sent_translation_accuracy(
self.europarl_data,
self.tgt_dico.lang, self.tgt_dico.word2id, tgt_emb,
self.src_dico.lang, self.src_dico.word2id, src_emb,
n_keys=n_keys, n_queries=n_queries,
method=method, idf=idf
)
to_log.update([('src_to_tgt_%s-%s' % (k, method), v) for k, v in results])
def sent_translation(self, to_log):
"""
Evaluation on sentence translation.
Only available on Europarl, for en - {de, es, fr, it} language pairs.
"""
lg1 = self.src_dico.lang
lg2 = self.tgt_dico.lang
# parameters
n_keys = 9076
n_queries = 2000
n_idf = 300000
# load europarl data
if not hasattr(self, 'europarl_data'):
self.europarl_data = load_bucc_data(lg1,
lg2,
self.params.split,
n_max=(n_keys + 2 * n_idf),
full=True)
self.gold = load_bucc_labels(lg1, lg2, self.params.split)
# if no Europarl data for this language pair
if not self.europarl_data:
return
# mapped word embeddings
src_emb = self.mapping(self.src_emb.weight).data
tgt_emb = self.tgt_emb.weight.data
# get idf weights
idf = get_idf(self.europarl_data, lg1, lg2, n_idf=n_idf)
for method in ['csls_knn_10']:
# source <- target sentence translation
pred_src, results = get_sent_translation_accuracy(
self.europarl_data,
self.gold, # swap cols
self.src_dico.lang,
self.src_dico.word2id,
src_emb,
self.tgt_dico.lang,
self.tgt_dico.word2id,
tgt_emb,
method=method,
idf=idf,
test=(self.params.split == 'test'),
device=self.params.device)
to_log.update([('tgt_to_src_%s-%s' % (k, method), v)
for k, v in results])
# target <- source sentence translation
pred_tgt, results = get_sent_translation_accuracy(
self.europarl_data,
self.gold[:, [1, 0]] if self.gold else None,
self.tgt_dico.lang,
self.tgt_dico.word2id,
tgt_emb,
self.src_dico.lang,
self.src_dico.word2id,
src_emb,
method=method,
idf=idf,
test=(self.params.split == 'test'),
device=self.params.device)
to_log.update([('src_to_tgt_%s-%s' % (k, method), v)
for k, v in results])
if self.params.split == 'test':
self.to_file(pred_src, lg1, lg2)
self.to_file(pred_tgt, lg2, lg1)
def sent_translation(self, to_log, src_lang=None, tgt_lang=None):
"""
Evaluation on sentence translation.
If src_lang and tgt_lang are not specified, evaluate all src_langs to tgt_lang
Only available on Europarl, for en - {de, es, fr, it} language pairs.
"""
# parameters
n_keys = 200000
n_queries = 2000
n_idf = 300000
# load europarl data
if not hasattr(self, 'europarl_data'):
self.europarl_data = {}
# evaluate all src langs to tgt_lang by default
if src_lang is None and tgt_lang is None:
tgt_lang = self.params.tgt_lang
for src_lang in self.params.src_langs:
lang_pair = (src_lang, tgt_lang)
# load europarl data
if lang_pair not in self.europarl_data:
self.europarl_data[lang_pair] = load_europarl_data(
src_lang, tgt_lang, n_max=(n_keys + 2 * n_idf))
# if no Europarl data for this language pair
if not self.europarl_data or lang_pair not in self.europarl_data \
or self.europarl_data[lang_pair] is None:
logger.info(
f'Europarl data not found for {src_lang}-{tgt_lang}.')
continue
# mapped word embeddings
src_emb = apply_mapping(self.mappings[src_lang],
self.embs[src_lang].weight)
tgt_emb = self.embs[tgt_lang].weight
# get idf weights
idf = get_idf(self.europarl_data[lang_pair],
src_lang,
tgt_lang,
n_idf=n_idf)
for method in ['nn', 'csls_knn_10']:
# source <- target sentence translation
results = get_sent_translation_accuracy(
self.europarl_data[lang_pair],
src_lang,
self.vocabs[src_lang].word2id,
src_emb,
tgt_lang,
self.vocabs[tgt_lang].word2id,
tgt_emb,
n_keys=n_keys,
n_queries=n_queries,
method=method,
idf=idf)
to_log.update([
('%s_to_%s_%s-%s' % (tgt_lang, src_lang, k, method), v)
for k, v in results
])
# target <- source sentence translation
results = get_sent_translation_accuracy(
self.europarl_data[lang_pair],
tgt_lang,
self.vocabs[tgt_lang].word2id,
tgt_emb,
src_lang,
self.vocabs[src_lang].word2id,
src_emb,
n_keys=n_keys,
n_queries=n_queries,
method=method,
idf=idf)
to_log.update([
('%s_to_%s_%s-%s' % (src_lang, tgt_lang, k, method), v)
for k, v in results
])
else:
# only evaluate src_lang to tgt_lang; bridge as necessary
assert src_lang is not None and tgt_lang is not None
lang_pair = (src_lang, tgt_lang)
# load europarl data
if lang_pair not in self.europarl_data:
self.europarl_data[lang_pair] = load_europarl_data(
src_lang, tgt_lang, n_max=(n_keys + 2 * n_idf))
# if no Europarl data for this language pair
if not self.europarl_data or lang_pair not in self.europarl_data \
or self.europarl_data[lang_pair] is None:
logger.info(
f'Europarl data not found for {src_lang}-{tgt_lang}.')
return
# encode src
src_emb = apply_mapping(self.mappings[src_lang],
self.embs[src_lang].weight)
# encode tgt
tgt_emb = apply_mapping(self.mappings[tgt_lang],
self.embs[tgt_lang].weight)
# get idf weights
idf = get_idf(self.europarl_data[lang_pair],
src_lang,
tgt_lang,
n_idf=n_idf)
for method in ['nn', 'csls_knn_10']:
# source <- target sentence translation
results = get_sent_translation_accuracy(
self.europarl_data[lang_pair],
src_lang,
self.vocabs[src_lang].word2id,
src_emb,
tgt_lang,
self.vocabs[tgt_lang].word2id,
tgt_emb,
n_keys=n_keys,
n_queries=n_queries,
method=method,
idf=idf)
to_log.update([
('%s_to_%s_%s-%s' % (tgt_lang, src_lang, k, method), v)
for k, v in results
])
# target <- source sentence translation
results = get_sent_translation_accuracy(
self.europarl_data[lang_pair],
tgt_lang,
self.vocabs[tgt_lang].word2id,
tgt_emb,
src_lang,
self.vocabs[src_lang].word2id,
src_emb,
n_keys=n_keys,
n_queries=n_queries,
method=method,
idf=idf)
to_log.update([
('%s_to_%s_%s-%s' % (src_lang, tgt_lang, k, method), v)
for k, v in results
])
