def generate_src_word_accuracy_report(ref, outs, src, ref_align_file=None, out_align_files=None,
acc_type='fmeas', bucket_type='freq', bucket_cutoffs=None,
freq_count_file=None, freq_corpus_file=None,
label_set=None,
src_labels=None,
title=None,
case_insensitive=False):
"""
Generate a report for source word analysis in both plain text and graphs.
Args:
ref: Tokens from the reference
outs: Tokens from the output file(s)
src: Tokens from the source
ref_align_file: Alignment file for the reference
out_align_files: Alignment file for the output file
acc_type: The type of accuracy to show (prec/rec/fmeas). Can also have multiple separated by '+'.
bucket_type: A string specifying the way to bucket words together to calculate F-measure (freq/tag)
bucket_cutoffs: The boundaries between buckets, specified as a colon-separated string.
freq_corpus_file: When using "freq" as a bucketer, which corpus to use to calculate frequency.
By default this uses the frequency in the reference test set, but it's often more informative
se the frequency in the training set, in which case you specify the path of the target side
he training corpus.
freq_count_file: An alternative to freq_corpus that uses a count file in "word\tfreq" format.
src_labels: either a filename of a file full of source labels, or a list of strings corresponding to `ref`.
title: A string specifying the caption of the printed table
case_insensitive: A boolean specifying whether to turn on the case insensitive option
"""
case_insensitive = True if case_insensitive == 'True' else False
if not src or not ref_align_file or not out_align_files:
raise ValueError("Must specify the source and the alignment files when performing source analysis.")
ref_align = corpus_utils.load_tokens(ref_align_file)
out_aligns = [corpus_utils.load_tokens(x) for x in arg_utils.parse_files(out_align_files)]
if len(out_aligns) != len(outs):
raise ValueError(f'The number of output files should be equal to the number of output alignment files.')
bucketer = bucketers.create_word_bucketer_from_profile(bucket_type,
bucket_cutoffs=bucket_cutoffs,
freq_count_file=freq_count_file,
freq_corpus_file=freq_corpus_file,
freq_data=src,
label_set=label_set,
case_insensitive=case_insensitive)
src_labels = corpus_utils.load_tokens(src_labels) if type(src_labels) == str else src_labels
matches = [bucketer.calc_source_bucketed_matches(src, ref, out, ref_align, out_align, src_labels=src_labels) for out, out_align in zip(outs, out_aligns)]
reporter = reporters.WordReport(bucketer=bucketer, matches=matches,
acc_type=acc_type, header="Source Word Accuracy Analysis",
title=title)
reporter.generate_report(output_fig_file=f'src-word-acc',
output_fig_format='pdf',
output_directory='outputs')
return reporter
def generate_word_accuracy_report(ref, outs,
acc_type='fmeas', bucket_type='freq', bucket_cutoffs=None,
freq_count_file=None, freq_corpus_file=None,
label_set=None,
ref_labels=None, out_labels=None,
title=None,
case_insensitive=False):
"""
Generate a report comparing the word accuracy in both plain text and graphs.
Args:
ref: Tokens from the reference
outs: Tokens from the output file(s)
acc_type: The type of accuracy to show (prec/rec/fmeas). Can also have multiple separated by '+'.
bucket_type: A string specifying the way to bucket words together to calculate F-measure (freq/tag)
bucket_cutoffs: The boundaries between buckets, specified as a colon-separated string.
freq_corpus_file: When using "freq" as a bucketer, which corpus to use to calculate frequency.
By default this uses the frequency in the reference test set, but it's often more informative
to use the frequency in the training set, in which case you specify the path of the
training corpus.
freq_count_file: An alternative to freq_corpus that uses a count file in "word\tfreq" format.
ref_labels: either a filename of a file full of reference labels, or a list of strings corresponding to `ref`.
out_labels: output labels. must be specified if ref_labels is specified.
title: A string specifying the caption of the printed table
case_insensitive: A boolean specifying whether to turn on the case insensitive option
"""
case_insensitive = True if case_insensitive == 'True' else False
if out_labels is not None:
out_labels = arg_utils.parse_files(out_labels)
if len(out_labels) != len(outs):
raise ValueError(f'The number of output files should be equal to the number of output labels.')
bucketer = bucketers.create_word_bucketer_from_profile(bucket_type,
bucket_cutoffs=bucket_cutoffs,
freq_count_file=freq_count_file,
freq_corpus_file=freq_corpus_file,
freq_data=ref,
label_set=label_set,
case_insensitive=case_insensitive)
ref_labels = corpus_utils.load_tokens(ref_labels) if type(ref_labels) == str else ref_labels
out_labels = [corpus_utils.load_tokens(out_labels[i]) if not out_labels is None else None for i in range(len(outs))]
matches = [bucketer.calc_bucketed_matches(ref, out, ref_labels=ref_labels, out_labels=out_label) for out, out_label in zip(outs, out_labels)]
reporter = reporters.WordReport(bucketer=bucketer, matches=matches,
acc_type=acc_type, header="Word Accuracy Analysis",
title=title)
reporter.generate_report(output_fig_file=f'word-acc',
output_fig_format='pdf',
output_directory='outputs')
return reporter
def generate_ngram_report(ref, outs,
min_ngram_length=1, max_ngram_length=4,
report_length=50, alpha=1.0, compare_type='match',
ref_labels=None, out_labels=None,
compare_directions='0-1',
case_insensitive=False):
"""
Generate a report comparing aggregate n-gram statistics in both plain text and graphs
Args:
ref: Tokens from the reference
outs: Tokens from the output file(s)
min_ngram_length: minimum n-gram length
max_ngram_length: maximum n-gram length
report_length: the number of n-grams to report
alpha: when sorting n-grams for salient features, the smoothing coefficient. A higher smoothing coefficient
will result in more frequent phenomena (sometimes this is good).
compare_type: what type of statistic to compare
(match: n-grams that match the reference, over: over-produced ngrams, under: under-produced ngrams)
ref_labels: either a filename of a file full of reference labels, or a list of strings corresponding to `ref`.
If specified, will aggregate statistics over labels instead of n-grams.
out_labels: output labels. must be specified if ref_labels is specified.
compare_directions: A string specifying which systems to compare
case_insensitive: A boolean specifying whether to turn on the case insensitive option
"""
min_ngram_length, max_ngram_length, report_length = int(min_ngram_length), int(max_ngram_length), int(report_length)
alpha = float(alpha)
case_insensitive = True if case_insensitive == 'True' else False
if out_labels is not None:
out_labels = arg_utils.parse_files(out_labels)
if len(out_labels) != len(outs):
raise ValueError(f'The number of output files should be equal to the number of output labels.')
if type(ref_labels) == str:
label_files_str = f' ref_labels={ref_labels},'
for i, out_label in enumerate(out_labels):
label_files_str += f' out{i}_labels={out_label},'
label_files = (label_files_str)
else:
label_files = None
if type(alpha) == str:
alpha = float(alpha)
if not type(ref_labels) == str and case_insensitive:
ref = corpus_utils.lower(ref)
outs = [corpus_utils.lower(out) for out in outs]
ref_labels = corpus_utils.load_tokens(ref_labels) if type(ref_labels) == str else ref_labels
out_labels = [corpus_utils.load_tokens(out_labels[i]) if not out_labels is None else None for i in range(len(outs))]
totals, matches, overs, unders = zip(*[ngram_utils.compare_ngrams(ref, out, ref_labels=ref_labels, out_labels=out_label,
min_length=min_ngram_length, max_length=max_ngram_length) for out, out_label in zip(outs, out_labels)])
direcs = arg_utils.parse_compare_directions(compare_directions)
scores = []
for (left, right) in direcs:
if compare_type == 'match':
scores.append(stat_utils.extract_salient_features(matches[left], matches[right], alpha=alpha))
elif compare_type == 'over':
scores.append(stat_utils.extract_salient_features(overs[left], overs[right], alpha=alpha))
elif compare_type == 'under':
scores.append(stat_utils.extract_salient_features(unders[left], unders[right], alpha=alpha))
else:
raise ValueError(f'Illegal compare_type "{compare_type}"')
scorelist = [sorted(score.items(), key=operator.itemgetter(1), reverse=True) for score in scores]
reporter = reporters.NgramReport(scorelist=scorelist, report_length=report_length,
min_ngram_length=min_ngram_length,
max_ngram_length=max_ngram_length,
matches=matches,
compare_type=compare_type, alpha=alpha,
compare_directions=direcs,
label_files=label_files)
reporter.generate_report(output_fig_file=f'ngram-min{min_ngram_length}-max{max_ngram_length}-{compare_type}',
output_fig_format='pdf',
output_directory='outputs')
return reporter
def generate_sentence_bucketed_report(ref, outs,
bucket_type='score', bucket_cutoffs=None,
statistic_type='count',
score_measure='bleu',
label_set=None,
ref_labels=None, out_labels=None,
title=None,
case_insensitive=False):
"""
Generate a report of sentences by bucket in both plain text and graphs
Args:
ref: Tokens from the reference
outs: Tokens from the output file(s)
bucket_type: The type of bucketing method to use
score_measure: If using 'score' as either bucket_type or statistic_type, which scorer to use
ref_labels: either a filename of a file full of reference labels, or a list of strings corresponding to `ref`. Would overwrite out_labels if specified.
out_labels: output labels.
title: A string specifying the caption of the printed table
case_insensitive: A boolean specifying whether to turn on the case insensitive option
"""
case_insensitive = True if case_insensitive == 'True' else False
if ref_labels is not None:
ref_labels = corpus_utils.load_tokens(ref_labels) if type(ref_labels) == str else ref_labels
if len(ref_labels) != len(ref):
raise ValueError(f'The number of labels should be equal to the number of sentences.')
elif out_labels is not None:
out_labels = arg_utils.parse_files(out_labels)
if len(out_labels) != len(outs):
raise ValueError(f'The number of output files should be equal to the number of output labels.')
out_labels = [corpus_utils.load_tokens(out_label) if type(out_label) == str else out_label for out_label in out_labels]
for out, out_label in zip(outs, out_labels):
if len(out_label) != len(out):
raise ValueError(f'The number of labels should be equal to the number of sentences.')
bucketer = bucketers.create_sentence_bucketer_from_profile(bucket_type, bucket_cutoffs=bucket_cutoffs,
score_type=score_measure, label_set=label_set, case_insensitive=case_insensitive)
bcs = [bucketer.create_bucketed_corpus(out, ref=ref, ref_labels=ref_labels if ref_labels else None, out_labels=out_labels[i] if out_labels else None) for i, out in enumerate(outs)]
if statistic_type == 'count':
scorer = None
aggregator = lambda out,ref: len(out)
elif statistic_type == 'score':
scorer = scorers.create_scorer_from_profile(score_measure, case_insensitive=case_insensitive)
aggregator = lambda out,ref: scorer.score_corpus(ref,out)[0]
else:
raise ValueError(f'Illegal statistic_type {statistic_type}')
stats = [[aggregator(out,ref) for (out,ref) in bc] for bc in bcs]
reporter = reporters.SentenceReport(bucketer=bucketer,
sys_stats=stats,
statistic_type=statistic_type, scorer=scorer,
title=title)
reporter.generate_report(output_fig_file=f'sentence-{statistic_type}-{score_measure}',
output_fig_format='pdf',
output_directory='outputs')
return reporter
def generate_sentence_bucketed_report(ref,
outs,
src=None,
bucket_type='score',
bucket_cutoffs=None,
statistic_type='count',
score_measure='sentbleu',
label_set=None,
ref_labels=None,
out_labels=None,
title=None,
case_insensitive=False,
output_bucket_details=False,
to_cache=False,
cache_dicts=None):
"""
Generate a report of sentences by bucket in both plain text and graphs
Args:
ref: Tokens from the reference
outs: Tokens from the output file(s)
bucket_type: The type of bucketing method to use
score_measure: If using 'score' as either bucket_type or statistic_type, which scorer to use
ref_labels: either a filename of a file full of reference labels, or a list of strings corresponding to `ref`. Would overwrite out_labels if specified.
out_labels: output labels.
title: A string specifying the caption of the printed table
case_insensitive: A boolean specifying whether to turn on the case insensitive option
output_bucket_details: A boolean specifying whether to output the number of words in each bucket
to_cache: Return a list of computed statistics if True
cache_dicts: A list of dictionaries that store cached statistics for each output
"""
# check and set parameters
if type(case_insensitive) == str:
case_insensitive = True if case_insensitive == 'True' else False
if type(output_bucket_details) == str:
output_bucket_details = True if output_bucket_details == 'True' else False
if ref_labels is not None:
ref_labels = corpus_utils.load_tokens(ref_labels) if type(
ref_labels) == str else ref_labels
if len(ref_labels) != len(ref):
raise ValueError(
f'The number of labels should be equal to the number of sentences.'
)
elif out_labels is not None:
out_labels = arg_utils.parse_files(out_labels)
if len(out_labels) != len(outs):
raise ValueError(
f'The number of output files should be equal to the number of output labels.'
)
out_labels = [
corpus_utils.load_tokens(out_label)
if type(out_label) == str else out_label
for out_label in out_labels
]
for out, out_label in zip(outs, out_labels):
if len(out_label) != len(out):
raise ValueError(
f'The number of labels should be equal to the number of sentences.'
)
# compute statistics
bucketer = bucketers.create_sentence_bucketer_from_profile(
bucket_type,
bucket_cutoffs=bucket_cutoffs,
score_type=score_measure,
label_set=label_set,
case_insensitive=case_insensitive)
src = [None for _ in ref] if src is None else src
if statistic_type == 'count':
scorer = None
if bucket_type != 'score' and bucket_type != 'lengthdiff':
ref = ref_label = None
aggregator = lambda out, refs, src: len(out)
elif statistic_type == 'score':
scorer = scorers.create_scorer_from_profile(
score_measure, case_insensitive=case_insensitive)
aggregator = lambda out, ref, src: scorer.score_corpus(ref, out, src)[0
]
else:
raise ValueError(f'Illegal statistic_type {statistic_type}')
cache_key_list = ['stats']
stats = cache_utils.extract_cache_dicts(cache_dicts, cache_key_list,
len(outs))
if cache_dicts is None:
bcs = [
bucketer.create_bucketed_corpus(
out,
ref=ref,
src=src,
ref_labels=ref_labels if ref_labels else None,
out_labels=out_labels[i] if out_labels else None)
for i, out in enumerate(outs)
]
stats = [[aggregator(out, ref, src) for (out, ref, src) in bc]
for bc in bcs]
if output_bucket_details and statistic_type == 'score':
bucket_cnt_calculator = lambda out, ref, src: len(out)
bucket_interval_calculator = lambda out, ref: sign_utils.eval_with_paired_bootstrap(
ref, [out], src, scorer, None)[1][0]
if cache_dicts is not None: # we don't cache bcs
bcs = [
bucketer.create_bucketed_corpus(
out,
ref=ref,
src=src,
ref_labels=ref_labels if ref_labels else None,
out_labels=out_labels[i] if out_labels else None)
for i, out in enumerate(outs)
]
bucket_cnts = [
bucket_cnt_calculator(out, ref, src) for (out, ref, src) in bcs[0]
]
bucket_intervals = [[
bucket_interval_calculator(out, ref, src) for (out, ref, src) in bc
] for bc in bcs]
else:
bucket_cnts = bucket_intervals = None
if to_cache:
cache_dict = cache_utils.return_cache_dict(cache_key_list, [stats])
return cache_dict
# generate reports
reporter = reporters.SentenceReport(bucketer=bucketer,
sys_stats=stats,
statistic_type=statistic_type,
scorer=scorer,
bucket_cnts=bucket_cnts,
bucket_intervals=bucket_intervals,
title=title)
reporter.generate_report(
output_fig_file=f'sentence-{statistic_type}-{score_measure}',
output_fig_format='pdf',
output_directory='outputs')
return reporter
def generate_word_accuracy_report(ref,
outs,
src=None,
acc_type='fmeas',
bucket_type='freq',
bucket_cutoffs=None,
freq_count_file=None,
freq_corpus_file=None,
label_set=None,
ref_labels=None,
out_labels=None,
title=None,
case_insensitive=False,
output_bucket_details=False,
to_cache=False,
cache_dicts=None):
"""
Generate a report comparing the word accuracy in both plain text and graphs.
Args:
ref: Tokens from the reference
outs: Tokens from the output file(s)
src: Tokens from the source
acc_type: The type of accuracy to show (prec/rec/fmeas). Can also have multiple separated by '+'.
bucket_type: A string specifying the way to bucket words together to calculate F-measure (freq/tag)
bucket_cutoffs: The boundaries between buckets, specified as a colon-separated string.
freq_corpus_file: When using "freq" as a bucketer, which corpus to use to calculate frequency.
By default this uses the frequency in the reference test set, but it's often more informative
to use the frequency in the training set, in which case you specify the path of the
training corpus.
freq_count_file: An alternative to freq_corpus that uses a count file in "word\tfreq" format.
ref_labels: either a filename of a file full of reference labels, or a list of strings corresponding to `ref`.
out_labels: output labels. must be specified if ref_labels is specified.
title: A string specifying the caption of the printed table
case_insensitive: A boolean specifying whether to turn on the case insensitive option
output_bucket_details: A boolean specifying whether to output the number of words in each bucket
to_cache: Return a list of computed statistics if True
cache_dicts: A list of dictionaries that store cached statistics for each output
"""
# check and set parameters
if type(case_insensitive) == str:
case_insensitive = True if case_insensitive == 'True' else False
if type(output_bucket_details) == str:
output_bucket_details = True if output_bucket_details == 'True' else False
if type(ref_labels) == str:
ref_labels = corpus_utils.load_tokens(ref_labels)
if out_labels is not None:
out_label_files = arg_utils.parse_files(out_labels)
out_labels = [corpus_utils.load_tokens(x) for x in out_label_files]
if len(out_labels) != len(outs):
raise ValueError(
f'The number of output files should be equal to the number of output labels.'
)
for i, (o, ol) in enumerate(zip(outs, out_labels)):
if len(o) != len(ol):
raise ValueError(
f'The labels in {out_label_files[i]} do not match the length of the output file {outs[i]}.'
)
# compute statistics
bucketer = bucketers.create_word_bucketer_from_profile(
bucket_type,
bucket_cutoffs=bucket_cutoffs,
freq_count_file=freq_count_file,
freq_corpus_file=freq_corpus_file,
freq_data=ref,
label_set=label_set,
case_insensitive=case_insensitive)
cache_key_list = [
'statistics', 'my_ref_total_list', 'my_out_totals_list',
'my_out_matches_list'
]
statistics, my_ref_total_list, my_out_totals_list, my_out_matches_list = cache_utils.extract_cache_dicts(
cache_dicts, cache_key_list, len(outs))
if cache_dicts is None:
statistics, my_ref_total_list, my_out_totals_list, my_out_matches_list = bucketer.calc_statistics(
ref, outs, ref_labels=ref_labels, out_labels=out_labels)
else:
my_ref_total_list = my_ref_total_list[0]
my_out_totals_list = list(np.concatenate(my_out_totals_list, 1))
my_out_matches_list = list(np.concatenate(my_out_matches_list, 1))
examples = bucketer.calc_examples(len(ref), len(outs), statistics,
my_ref_total_list, my_out_matches_list)
bucket_cnts, bucket_intervals = bucketer.calc_bucket_details(
my_ref_total_list, my_out_totals_list,
my_out_matches_list) if output_bucket_details else (None, None)
if to_cache:
cache_dict = cache_utils.return_cache_dict(cache_key_list, [
statistics, [my_ref_total_list], [my_out_totals_list],
[my_out_matches_list]
])
return cache_dict
# generate reports
reporter = reporters.WordReport(bucketer=bucketer,
statistics=statistics,
examples=examples,
bucket_cnts=bucket_cnts,
bucket_intervals=bucket_intervals,
src_sents=src,
ref_sents=ref,
ref_labels=ref_labels,
out_sents=outs,
out_labels=out_labels,
acc_type=acc_type,
header="Word Accuracy Analysis",
title=title)
reporter.generate_report(output_fig_file=f'word-acc',
output_fig_format='pdf',
output_directory='outputs')
return reporter
