forked from stanford-oval/storm
-
Notifications
You must be signed in to change notification settings - Fork 0
/
citation_quality.py
382 lines (307 loc) · 14.4 KB
/
citation_quality.py
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
"""The code is adapted from https://github.com/princeton-nlp/ALCE/blob/main/eval.py"""
import copy
import json
import logging
import random
import re
from argparse import ArgumentParser
import numpy as np
import pandas as pd
import torch
from nltk import sent_tokenize
from tqdm import tqdm
from transformers import (
AutoTokenizer, AutoModelForCausalLM
)
random.seed(0)
logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S')
logger = logging.getLogger(__name__)
AUTOAIS_MODEL = "google/t5_xxl_true_nli_mixture"
global autoais_model, autoais_tokenizer, mistral_7b_instruct, mistral_7b_tokenizer
autoais_model, autoais_tokenizer, mistral_7b_instruct, mistral_7b_tokenizer = None, None, None, None
def get_max_memory():
"""Get the maximum memory available for the current GPU for loading models."""
free_in_GB = int(torch.cuda.mem_get_info()[0] / 1024 ** 3)
max_memory = f'{free_in_GB - 6}GB'
n_gpus = torch.cuda.device_count()
max_memory = {i: max_memory for i in range(n_gpus)}
return max_memory
def remove_citations(sent):
return re.sub(r"\[\d+", "", re.sub(r" \[\d+", "", sent)).replace(" |", "").replace("]", "")
def truncate_paragraph(paragraph, max_words):
# Tokenize paragraph into sentences
sentences = re.split(r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?)\s', paragraph)
# Tokenize each sentence into words and form trunks
trunks = []
current_trunk = []
current_word_count = 0
for sentence in sentences:
sentence_words = sentence.split() # Tokenize sentence into words
sentence_word_count = len(sentence_words)
if current_word_count + sentence_word_count <= max_words:
current_trunk.append(sentence)
current_word_count += sentence_word_count
else:
trunks.append(' '.join(current_trunk))
current_trunk = [sentence]
current_word_count = sentence_word_count
if current_trunk:
trunks.append(' '.join(current_trunk))
return trunks
def _run_nli_autoais(passage, claim, partial):
"""
Run inference for assessing AIS between a premise and hypothesis.
Adapted from https://github.com/google-research-datasets/Attributed-QA/blob/main/evaluation.py
"""
global mistral_7b_instruct, mistral_7b_tokenizer
passage_trunks = truncate_paragraph(passage, 500)
inference = 0
for trunk in passage_trunks:
if partial:
s = f"Can the source at least partially support the claim? Start your answer with 'Yes' or 'No'.\nSource: {trunk}\nClaim: {claim}"
else:
s = f"Is the claim faithful to the source? A claim is faithful to the source if the core part in the claim can be supported by the source.\nStart your answer with 'Yes' or 'No'.\nSource: {trunk}\nClaim: {claim}"
messages = [{'role': 'user', 'content': s}]
encodeds = mistral_7b_tokenizer.apply_chat_template(messages, return_tensors="pt")
model_inputs = encodeds.to('cuda')
generated_ids = mistral_7b_instruct.generate(model_inputs, max_new_tokens=200, do_sample=False)
decoded = mistral_7b_tokenizer.batch_decode(generated_ids, temperature=0)[0]
res = decoded[decoded.find('[/INST]') + len('[/INST]'):].strip()
if res.startswith('Yes'):
inference = 1
break
return inference
def compute_autoais(data,
decontext=False,
concat=False,
qampari=False,
at_most_citations=None):
"""
Compute AutoAIS score.
Args:
data: requires field `output` and `docs`
- docs should be a list of items with fields `title` and `text` (or `phrase` and `sent` for QA-extracted docs)
citation: check citations and use the corresponding references.
decontext: decontextualize the output
"""
global mistral_7b_instruct, mistral_7b_tokenizer
def _format_document(doc):
"""Format document for AutoAIS."""
if "sent" in doc:
# QA-extracted docs
return "Title: %s\n%s" % (doc['title'], doc['sent'])
else:
return "Title: %s\n%s" % (doc['title'], doc['text'])
ais_scores = []
ais_scores_prec = []
sent_total = 0
sent_mcite = 0
sent_mcite_support = 0
sent_mcite_overcite = 0
eval_log = []
for item in tqdm(data):
# Get sentences by using NLTK
if qampari:
sents = [item['question'] + " " + x.strip() for x in
item['output'].rstrip().rstrip(".").rstrip(",").split(",")]
else:
sents = sent_tokenize(item['output'])
if len(sents) == 0:
continue
target_sents = [remove_citations(sent).strip() for sent in sents]
entail = 0
entail_prec = 0
total_citations = 0
for sent_id, sent in enumerate(sents):
target_sent = target_sents[sent_id] # Citation removed and (if opted for) decontextualized
joint_entail = -1 # Undecided
# Find references
ref = [int(r[1:]) - 1 for r in re.findall(r"\[\d+", sent)] # In text citation id starts from 1
logger.info(f"For `{sent}`, find citations {ref}")
if len(ref) == 0:
# No citations
joint_entail = 0
elif any([ref_id >= len(item['docs']) for ref_id in ref]):
# Citations out of range
joint_entail = 0
else:
if at_most_citations is not None:
ref = ref[:at_most_citations]
total_citations += len(ref)
joint_passage = '\n'.join([_format_document(item['docs'][psgs_id]) for psgs_id in ref])
# If not directly rejected by citation format error, calculate the recall score
if joint_entail == -1:
joint_entail = _run_nli_autoais(joint_passage, target_sent, partial=False)
entail += joint_entail
if joint_entail == 0:
logger.info(f'[Unsupported sentence] {sent}')
if len(ref) > 1:
sent_mcite += 1
unnecessary_citations = []
# calculate the precision score if applicable
if joint_entail and len(ref) > 1:
sent_mcite_support += 1
# Precision check: did the model cite any unnecessary documents?
for psgs_id in ref:
# condition A
passage = _format_document(item['docs'][psgs_id])
nli_result = _run_nli_autoais(passage, target_sent, partial=True)
# condition B
if not nli_result:
subset_exclude = copy.deepcopy(ref)
subset_exclude.remove(psgs_id)
passage = '\n'.join([_format_document(item['docs'][pid]) for pid in subset_exclude])
nli_result = _run_nli_autoais(passage, target_sent, partial=False)
if nli_result: # psgs_id is not necessary
flag = 0
sent_mcite_overcite += 1
logger.info(f'[Unnecessary citation] sent: {sent} citation: [{psgs_id}]')
unnecessary_citations.append(psgs_id)
else:
entail_prec += 1
else:
entail_prec += 1
else:
entail_prec += joint_entail
eval_log.append({
"sent": sent,
"target_sent": target_sent,
"ref": ref,
"joint_entail": joint_entail,
"unnecessary_citations": unnecessary_citations,
})
sent_total += len(sents)
ais_scores.append(entail / len(sents))
ais_scores_prec.append(entail_prec / total_citations if total_citations > 0 else 0) # len(sents))
if sent_mcite > 0 and sent_mcite_support > 0:
print(
"Among all sentences, %.2f%% have multiple citations, among which %.2f%% are supported by the joint set, among which %.2f%% overcite." % (
100 * sent_mcite / sent_total,
100 * sent_mcite_support / sent_mcite,
100 * sent_mcite_overcite / sent_mcite_support
))
citation_rec = 100 * np.mean(ais_scores)
citation_prec = 100 * np.mean(ais_scores_prec)
return {
"evaluation_logs": eval_log,
"citation_rec": citation_rec,
"citation_prec": citation_prec,
}
def load_str(path):
with open(path, 'r') as f:
return '\n'.join(f.readlines())
def load_json(path):
with open(path, 'r') as f:
return json.load(f)
def extract_url(text):
# This pattern matches the format [number]: text (url)
pattern = r'\[\d+\]: .* \((https?://[^\)]+)\)'
# Find the first match of the pattern in the text
match = re.search(pattern, text)
# Return the URL if a match is found
if match:
return match.group(1)
else:
return None
def expand_citaions(output):
"""
Expand citations by following rule:
1. convert "<sentence 1>. <sentence 2> [2][3]" into "<sentence 1> [2][3]"
2. "<sentence 1>[1]. <last paragraph senetence>" will be changed to "<sentence 1>[1]. <last paragraph senetence>[1]. "
"""
def find_citations(sentence):
return re.findall(r'\[(\d+)\]', sentence)
modified_pargraphs = []
for paragraph_idx, paragraph in enumerate(output.split("\n")):
if len(paragraph) == 0:
continue
sentence_endings = r"(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?|\!)\s"
sentences = re.split(sentence_endings, paragraph)
sentences = [sentence.strip() for sentence in sentences if sentence]
modified_sentences = []
for sentence_idx, sentence in enumerate(sentences):
if len(sentence) == 0:
continue
citations = find_citations(sentence)
added_citations = ''
if len(citations) == 0:
if sentence_idx == len(sentences) - 1 and sentence_idx - 1 >= 0:
for citation in find_citations(sentences[sentence_idx - 1]):
added_citations += f"[{citation}]"
elif sentence_idx + 1 < len(sentences):
for citation in find_citations(sentences[sentence_idx + 1]):
added_citations += f"[{citation}]"
modified_sentences.append(sentence[:-1] + added_citations + sentence[-1])
modified_pargraph = " ".join(modified_sentences)
modified_pargraphs.append(modified_pargraph)
modified_output = '\n'.join(modified_pargraphs).strip()
return modified_output
def format_data(root_dir, file_name_suffix, do_citation_expansion=False):
final_page = load_str(f'{root_dir}{file_name_suffix}.txt')
search_results = load_json(f'{root_dir}_search_results.json')
if 'url_to_info' in search_results:
url_to_info = search_results['url_to_info']
assert list(search_results['url_to_unified_index'].keys()) == list(url_to_info.keys())
else:
url_to_info = {d['url']: {'title': d['title'], 'snippets': d['snippets']} for d in search_results}
output = []
for line in final_page.split('\n'):
if len(line) == 0 or line[0] == '#':
continue
output.append(line)
output = '\n'.join(output).strip()
if do_citation_expansion:
output = expand_citaions(output)
docs = []
for url in url_to_info:
docs.append({
'title': url_to_info[url]['title'],
'text': '\n'.join(set(url_to_info[url]['snippets'])),
})
return output, docs
def main(args):
global mistral_7b_instruct, mistral_7b_tokenizer
mistral_7b_instruct = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1")
mistral_7b_tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1")
mistral_7b_instruct = mistral_7b_instruct.to('cuda')
if args.mode == 'single':
output, docs = format_data(args.dir, args.file_name_suffix, args.do_citation_expansion)
data = [{'output': output, 'docs': docs}]
result = compute_autoais(data=data)
print('===== Citation Quality =====')
print(f'recall: {result["citation_rec"]}, precision: {result["citation_prec"]}')
elif args.mode == 'batch':
df = pd.read_csv(args.batch_topic_path)
results = {'topic': [], 'recall': [], 'precision': [], 'eval_log': []}
for i, row in tqdm(df.iterrows(), total=len(df)):
file_name = row['topic'].replace(' ', '_').replace('/', '_')
output, docs = format_data(f'{args.dir}/{file_name}', args.file_name_suffix, args.do_citation_expansion)
data = [{'output': output, 'docs': docs}]
result = compute_autoais(data=data)
results['topic'].append(row['topic'])
results['recall'].append(result['citation_rec'])
results['precision'].append(result['citation_prec'])
results['eval_log'].append(result['evaluation_logs'])
with open(f'{args.dir}/citation_quality.json', 'w') as f:
json.dump(results, f, indent=2)
avg_recall = sum(results['recall']) / len(results['recall'])
avg_precision = sum(results['precision']) / len(results['precision'])
print('===== Citation Quality =====')
print(f'Average recall: {avg_recall}, average precision: {avg_precision}')
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument('--mode', type=str, choices=['single', 'batch'],
help='Whether to calculate the precision recall/acc on a single doc or a batch of docs.')
parser.add_argument('--disable_log', action='store_true',
help='Whether to disable log on consoles.')
parser.add_argument('--dir', type=str, help='Directory of the saved results.')
parser.add_argument('--file_name_suffix', default='', type=str, help='Suffix of the file name.')
parser.add_argument('--batch_topic_path', type=str, help='Path of the file storing batch topics.')
parser.add_argument("-p", '--do_citation_expansion', action='store_true', help="whether expand citations")
args = parser.parse_args()
if args.disable_log:
logger.setLevel(logging.ERROR)
else:
logger.setLevel(logging.INFO)
main(args)