Sophia Koehler
commited on
Commit
·
e39c176
1
Parent(s):
e8df6fa
fix2
Browse files
app.py
CHANGED
@@ -1,58 +1,49 @@
|
|
1 |
# -*- coding: utf-8 -*-
|
2 |
-
|
3 |
-
|
4 |
-
"""## Pre-requisite code
|
5 |
-
|
6 |
-
The code within this section will be used in the tasks. Please do not change these code lines.
|
7 |
-
|
8 |
-
### SciQ loading and counting
|
9 |
-
"""
|
10 |
-
|
11 |
from dataclasses import dataclass
|
12 |
-
import pickle
|
13 |
import os
|
14 |
-
|
15 |
-
from
|
16 |
-
from collections import Counter
|
17 |
-
import tqdm
|
18 |
import re
|
|
|
|
|
|
|
19 |
import nltk
|
20 |
-
|
|
|
|
|
21 |
from nltk.corpus import stopwords as nltk_stopwords
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
LANGUAGE = "english"
|
24 |
-
word_splitter = re.compile(r"(?u)\b\w\w+\b").findall
|
25 |
stopwords = set(nltk_stopwords.words(LANGUAGE))
|
26 |
-
|
27 |
-
|
28 |
-
def word_splitting(text: str) -> List[str]:
|
29 |
-
return word_splitter(text.lower())
|
30 |
-
|
31 |
-
def lemmatization(words: List[str]) -> List[str]:
|
32 |
-
return words # We ignore lemmatization here for simplicity
|
33 |
|
34 |
def simple_tokenize(text: str) -> List[str]:
|
35 |
-
words =
|
36 |
-
tokenized =
|
37 |
-
tokenized = lemmatization(tokenized)
|
38 |
return tokenized
|
39 |
|
40 |
-
T = TypeVar("T", bound="InvertedIndex")
|
41 |
-
|
42 |
@dataclass
|
43 |
class PostingList:
|
44 |
-
term: str
|
45 |
-
docid_postings: List[int]
|
46 |
-
tweight_postings: List[float]
|
47 |
|
|
|
48 |
|
49 |
@dataclass
|
50 |
class InvertedIndex:
|
51 |
-
posting_lists: List[PostingList]
|
52 |
vocab: Dict[str, int]
|
53 |
-
cid2docid: Dict[str, int]
|
54 |
-
collection_ids: List[str]
|
55 |
-
doc_texts: Optional[List[str]] = None
|
56 |
|
57 |
def save(self, output_dir: str) -> None:
|
58 |
os.makedirs(output_dir, exist_ok=True)
|
@@ -61,138 +52,28 @@ class InvertedIndex:
|
|
61 |
|
62 |
@classmethod
|
63 |
def from_saved(cls: Type[T], saved_dir: str) -> T:
|
64 |
-
index = cls(
|
65 |
-
posting_lists=[], vocab={}, cid2docid={}, collection_ids=[], doc_texts=None
|
66 |
-
)
|
67 |
with open(os.path.join(saved_dir, "index.pkl"), "rb") as f:
|
68 |
-
|
69 |
-
return index
|
70 |
-
|
71 |
-
|
72 |
-
# The output of the counting function:
|
73 |
-
@dataclass
|
74 |
-
class Counting:
|
75 |
-
posting_lists: List[PostingList]
|
76 |
-
vocab: Dict[str, int]
|
77 |
-
cid2docid: Dict[str, int]
|
78 |
-
collection_ids: List[str]
|
79 |
-
dfs: List[int] # tid -> df
|
80 |
-
dls: List[int] # docid -> doc length
|
81 |
-
avgdl: float
|
82 |
-
nterms: int
|
83 |
-
doc_texts: Optional[List[str]] = None
|
84 |
-
|
85 |
-
def run_counting(
|
86 |
-
documents: Iterable[Document],
|
87 |
-
tokenize_fn: Callable[[str], List[str]] = simple_tokenize,
|
88 |
-
store_raw: bool = True, # store the document text in doc_texts
|
89 |
-
ndocs: Optional[int] = None,
|
90 |
-
show_progress_bar: bool = True,
|
91 |
-
) -> Counting:
|
92 |
-
"""Counting TFs, DFs, doc_lengths, etc."""
|
93 |
-
posting_lists: List[PostingList] = []
|
94 |
-
vocab: Dict[str, int] = {}
|
95 |
-
cid2docid: Dict[str, int] = {}
|
96 |
-
collection_ids: List[str] = []
|
97 |
-
dfs: List[int] = [] # tid -> df
|
98 |
-
dls: List[int] = [] # docid -> doc length
|
99 |
-
nterms: int = 0
|
100 |
-
doc_texts: Optional[List[str]] = []
|
101 |
-
for doc in tqdm.tqdm(
|
102 |
-
documents,
|
103 |
-
desc="Counting",
|
104 |
-
total=ndocs,
|
105 |
-
disable=not show_progress_bar,
|
106 |
-
):
|
107 |
-
if doc.collection_id in cid2docid:
|
108 |
-
continue
|
109 |
-
collection_ids.append(doc.collection_id)
|
110 |
-
docid = cid2docid.setdefault(doc.collection_id, len(cid2docid))
|
111 |
-
toks = tokenize_fn(doc.text)
|
112 |
-
tok2tf = Counter(toks)
|
113 |
-
dls.append(sum(tok2tf.values()))
|
114 |
-
for tok, tf in tok2tf.items():
|
115 |
-
nterms += tf
|
116 |
-
tid = vocab.get(tok, None)
|
117 |
-
if tid is None:
|
118 |
-
posting_lists.append(
|
119 |
-
PostingList(term=tok, docid_postings=[], tweight_postings=[])
|
120 |
-
)
|
121 |
-
tid = vocab.setdefault(tok, len(vocab))
|
122 |
-
posting_lists[tid].docid_postings.append(docid)
|
123 |
-
posting_lists[tid].tweight_postings.append(tf)
|
124 |
-
if tid < len(dfs):
|
125 |
-
dfs[tid] += 1
|
126 |
-
else:
|
127 |
-
dfs.append(0)
|
128 |
-
if store_raw:
|
129 |
-
doc_texts.append(doc.text)
|
130 |
-
else:
|
131 |
-
doc_texts = None
|
132 |
-
return Counting(
|
133 |
-
posting_lists=posting_lists,
|
134 |
-
vocab=vocab,
|
135 |
-
cid2docid=cid2docid,
|
136 |
-
collection_ids=collection_ids,
|
137 |
-
dfs=dfs,
|
138 |
-
dls=dls,
|
139 |
-
avgdl=sum(dls) / len(dls),
|
140 |
-
nterms=nterms,
|
141 |
-
doc_texts=doc_texts,
|
142 |
-
)
|
143 |
-
|
144 |
-
from nlp4web_codebase.ir.data_loaders.sciq import load_sciq
|
145 |
-
sciq = load_sciq()
|
146 |
-
counting = run_counting(documents=iter(sciq.corpus), ndocs=len(sciq.corpus))
|
147 |
-
|
148 |
-
"""### BM25 Index"""
|
149 |
-
|
150 |
-
from __future__ import annotations
|
151 |
-
from dataclasses import asdict, dataclass
|
152 |
-
import math
|
153 |
-
import os
|
154 |
-
from typing import Iterable, List, Optional, Type
|
155 |
-
import tqdm
|
156 |
-
from nlp4web_codebase.ir.data_loaders.dm import Document
|
157 |
-
|
158 |
|
159 |
@dataclass
|
160 |
class BM25Index(InvertedIndex):
|
161 |
|
162 |
-
@staticmethod
|
163 |
-
def tokenize(text: str) -> List[str]:
|
164 |
-
return simple_tokenize(text)
|
165 |
-
|
166 |
@staticmethod
|
167 |
def cache_term_weights(
|
168 |
-
posting_lists: List[PostingList],
|
169 |
-
total_docs: int,
|
170 |
-
avgdl: float,
|
171 |
-
dfs: List[int],
|
172 |
-
dls: List[int],
|
173 |
-
k1: float,
|
174 |
-
b: float,
|
175 |
) -> None:
|
176 |
-
"""Compute term weights and caching"""
|
177 |
-
|
178 |
N = total_docs
|
179 |
-
for tid, posting_list in enumerate(
|
180 |
-
tqdm.tqdm(posting_lists, desc="Regularizing TFs")
|
181 |
-
):
|
182 |
idf = BM25Index.calc_idf(df=dfs[tid], N=N)
|
183 |
-
for i in
|
184 |
-
docid = posting_list.docid_postings[i]
|
185 |
tf = posting_list.tweight_postings[i]
|
186 |
dl = dls[docid]
|
187 |
-
|
188 |
tf=tf, dl=dl, avgdl=avgdl, k1=k1, b=b
|
189 |
-
)
|
190 |
-
posting_list.tweight_postings[i] = regularized_tf * idf
|
191 |
|
192 |
@staticmethod
|
193 |
-
def calc_regularized_tf(
|
194 |
-
tf: int, dl: float, avgdl: float, k1: float, b: float
|
195 |
-
) -> float:
|
196 |
return tf / (tf + k1 * (1 - b + b * dl / avgdl))
|
197 |
|
198 |
@staticmethod
|
@@ -201,305 +82,54 @@ class BM25Index(InvertedIndex):
|
|
201 |
|
202 |
@classmethod
|
203 |
def build_from_documents(
|
204 |
-
cls: Type[BM25Index],
|
205 |
-
documents: Iterable[Document],
|
206 |
-
store_raw: bool = True,
|
207 |
-
output_dir: Optional[str] = None,
|
208 |
-
ndocs: Optional[int] = None,
|
209 |
-
show_progress_bar: bool = True,
|
210 |
-
k1: float = 0.9,
|
211 |
-
b: float = 0.4,
|
212 |
) -> BM25Index:
|
213 |
-
#
|
214 |
-
counting = run_counting(
|
215 |
-
|
216 |
-
|
217 |
-
store_raw=store_raw,
|
218 |
-
ndocs=ndocs,
|
219 |
-
show_progress_bar=show_progress_bar,
|
220 |
-
)
|
221 |
-
|
222 |
-
# Compute term weights and caching:
|
223 |
-
posting_lists = counting.posting_lists
|
224 |
-
total_docs = len(counting.cid2docid)
|
225 |
-
BM25Index.cache_term_weights(
|
226 |
-
posting_lists=posting_lists,
|
227 |
-
total_docs=total_docs,
|
228 |
-
avgdl=counting.avgdl,
|
229 |
-
dfs=counting.dfs,
|
230 |
-
dls=counting.dls,
|
231 |
-
k1=k1,
|
232 |
-
b=b,
|
233 |
-
)
|
234 |
-
|
235 |
-
# Assembly and save:
|
236 |
-
index = BM25Index(
|
237 |
-
posting_lists=posting_lists,
|
238 |
-
vocab=counting.vocab,
|
239 |
-
cid2docid=counting.cid2docid,
|
240 |
-
collection_ids=counting.collection_ids,
|
241 |
-
doc_texts=counting.doc_texts,
|
242 |
-
)
|
243 |
-
return index
|
244 |
-
|
245 |
-
bm25_index = BM25Index.build_from_documents(
|
246 |
-
documents=iter(sciq.corpus),
|
247 |
-
ndocs=12160,
|
248 |
-
show_progress_bar=True,
|
249 |
-
)
|
250 |
-
bm25_index.save("output/bm25_index")
|
251 |
-
!ls
|
252 |
-
|
253 |
-
"""### BM25 Retriever"""
|
254 |
-
|
255 |
-
from nlp4web_codebase.ir.models import BaseRetriever
|
256 |
-
from typing import Type
|
257 |
-
from abc import abstractmethod
|
258 |
-
|
259 |
-
|
260 |
-
class BaseInvertedIndexRetriever(BaseRetriever):
|
261 |
-
|
262 |
-
@property
|
263 |
-
@abstractmethod
|
264 |
-
def index_class(self) -> Type[InvertedIndex]:
|
265 |
-
pass
|
266 |
|
|
|
267 |
def __init__(self, index_dir: str) -> None:
|
268 |
-
self.index =
|
269 |
-
|
270 |
-
def get_term_weights(self, query: str, cid: str) -> Dict[str, float]:
|
271 |
-
toks = self.index.tokenize(query)
|
272 |
-
target_docid = self.index.cid2docid[cid]
|
273 |
-
term_weights = {}
|
274 |
-
for tok in toks:
|
275 |
-
if tok not in self.index.vocab:
|
276 |
-
continue
|
277 |
-
tid = self.index.vocab[tok]
|
278 |
-
posting_list = self.index.posting_lists[tid]
|
279 |
-
for docid, tweight in zip(
|
280 |
-
posting_list.docid_postings, posting_list.tweight_postings
|
281 |
-
):
|
282 |
-
if docid == target_docid:
|
283 |
-
term_weights[tok] = tweight
|
284 |
-
break
|
285 |
-
return term_weights
|
286 |
-
|
287 |
-
def score(self, query: str, cid: str) -> float:
|
288 |
-
return sum(self.get_term_weights(query=query, cid=cid).values())
|
289 |
|
290 |
def retrieve(self, query: str, topk: int = 10) -> Dict[str, float]:
|
291 |
-
toks =
|
292 |
-
docid2score
|
293 |
for tok in toks:
|
294 |
-
if tok
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
299 |
-
posting_list.docid_postings, posting_list.tweight_postings
|
300 |
-
):
|
301 |
-
docid2score.setdefault(docid, 0)
|
302 |
-
docid2score[docid] += tweight
|
303 |
-
docid2score = dict(
|
304 |
-
sorted(docid2score.items(), key=lambda pair: pair[1], reverse=True)[:topk]
|
305 |
-
)
|
306 |
return {
|
307 |
-
self.index.collection_ids[docid]: score
|
308 |
-
for docid, score in docid2score.items()
|
309 |
}
|
310 |
|
311 |
-
|
312 |
-
class BM25Retriever(BaseInvertedIndexRetriever):
|
313 |
-
|
314 |
-
@property
|
315 |
-
def index_class(self) -> Type[BM25Index]:
|
316 |
-
return BM25Index
|
317 |
-
|
318 |
-
bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
|
319 |
-
bm25_retriever.retrieve("What type of diseases occur when the immune system attacks normal body cells?")
|
320 |
-
|
321 |
-
"""# TASK1: tune b and k1 (4 points)
|
322 |
-
|
323 |
-
Tune b and k1 on the **dev** split of SciQ using the metric MAP@10. The evaluation function (`evalaute_map`) is provided. Record the values in `plots_k1` and `plots_b`. Do it in a greedy manner: as the influence from b is larger, please first tune b (with k1 fixed to the default value 0.9) and use the best value of b to further tune k1.
|
324 |
-
|
325 |
-
$${\displaystyle {\text{score}}(D,Q)=\sum _{i=1}^{n}{\text{IDF}}(q_{i})\cdot {\frac {f(q_{i},D)\cdot (k_{1}+1)}{f(q_{i},D)+k_{1}\cdot \left(1-b+b\cdot {\frac {|D|}{\text{avgdl}}}\right)}}}$$
|
326 |
-
"""
|
327 |
-
|
328 |
-
from nlp4web_codebase.ir.data_loaders import Split
|
329 |
-
import pytrec_eval
|
330 |
-
|
331 |
-
|
332 |
-
def evaluate_map(rankings: Dict[str, Dict[str, float]], split=Split.dev) -> float:
|
333 |
-
metric = "map_cut_10"
|
334 |
-
qrels = sciq.get_qrels_dict(split)
|
335 |
-
evaluator = pytrec_eval.RelevanceEvaluator(sciq.get_qrels_dict(split), (metric,))
|
336 |
-
qps = evaluator.evaluate(rankings)
|
337 |
-
return float(np.mean([qp[metric] for qp in qps.values()]))
|
338 |
-
|
339 |
-
"""Example of using the pre-requisite code:"""
|
340 |
-
|
341 |
-
# Loading dataset:
|
342 |
-
from nlp4web_codebase.ir.data_loaders.sciq import load_sciq
|
343 |
-
sciq = load_sciq()
|
344 |
-
counting = run_counting(documents=iter(sciq.corpus), ndocs=len(sciq.corpus))
|
345 |
-
|
346 |
-
# Building BM25 index and save:
|
347 |
-
bm25_index = BM25Index.build_from_documents(
|
348 |
-
documents=iter(sciq.corpus),
|
349 |
-
ndocs=12160,
|
350 |
-
show_progress_bar=True
|
351 |
-
)
|
352 |
-
bm25_index.save("output/bm25_index")
|
353 |
-
|
354 |
-
# Loading index and use BM25 retriever to retrieve:
|
355 |
-
bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
|
356 |
-
print(bm25_retriever.retrieve("What type of diseases occur when the immune system attacks normal body cells?")) # the ranking
|
357 |
-
|
358 |
-
plots_b: Dict[str, List[float]] = {
|
359 |
-
"X": [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0],
|
360 |
-
"Y": []
|
361 |
-
}
|
362 |
-
plots_k1: Dict[str, List[float]] = {
|
363 |
-
"X": [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0],
|
364 |
-
"Y": []
|
365 |
-
}
|
366 |
-
|
367 |
-
## YOUR_CODE_STARTS_HERE
|
368 |
-
class MyBMIndex(BM25Index):
|
369 |
-
|
370 |
-
@staticmethod
|
371 |
-
def calc_regularized_tf(
|
372 |
-
tf: int, dl: float, avgdl: float, k1: float, b: float
|
373 |
-
) -> float:
|
374 |
-
return tf * (k1 + 1) / (tf + k1 * (1 - b + b * (dl / avgdl)**1.5))
|
375 |
-
|
376 |
-
@staticmethod
|
377 |
-
def calc_idf(df: int, N: int):
|
378 |
-
return math.log((N + 1) / (df + 0.5)) + 1
|
379 |
-
import numpy as np
|
380 |
-
# Two steps should be involved:
|
381 |
-
# Step 1. Fix k1 value to the default one 0.9,
|
382 |
-
# go through all the candidate b values (0, 0.1, ..., 1.0),
|
383 |
-
# and record in plots_b["Y"] the corresponding performances obtained via evaluate_map;
|
384 |
-
# Step 2. Fix b to the best one in step 1. and do the same for k1.
|
385 |
-
|
386 |
-
# Hint (on using the pre-requisite code):
|
387 |
-
# - One can use the loaded sciq dataset directly (loaded in the pre-requisite code);
|
388 |
-
# - One can build bm25_index with `BM25Index.build_from_documents`;
|
389 |
-
# - One can use BM25Retriever to load the index and perform retrieval on the dev queries
|
390 |
-
# (dev queries can be obtained via sciq.get_split_queries(Split.dev))
|
391 |
-
|
392 |
-
counting = run_counting(documents=iter(sciq.corpus), ndocs=len(sciq.corpus))
|
393 |
-
|
394 |
-
def get_ranking(k1, b, counting) -> Dict[str, Dict[str, float]]:
|
395 |
-
# Building BM25 index and save:
|
396 |
-
bm25_index = MyBMIndex.build_from_documents(
|
397 |
-
documents=iter(sciq.corpus),
|
398 |
-
ndocs=12160,
|
399 |
-
show_progress_bar=True,
|
400 |
-
k1=k1,
|
401 |
-
b=b
|
402 |
-
)
|
403 |
-
bm25_index.save("output/bm25_index")
|
404 |
-
|
405 |
-
# Loading index and use BM25 retriever to retrieve:
|
406 |
-
bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
|
407 |
-
query_terms = sciq.get_split_queries(split= Split.dev)
|
408 |
-
rankings = {}
|
409 |
-
for query in query_terms:
|
410 |
-
ranking = bm25_retriever.retrieve(query=query.text)
|
411 |
-
rankings[query.query_id] = ranking
|
412 |
-
return rankings
|
413 |
-
for b in plots_b["X"]:
|
414 |
-
ranking = get_ranking(0.9, b, counting)
|
415 |
-
plots_b["Y"].append(evaluate_map(rankings=ranking))
|
416 |
-
|
417 |
-
max_b = np.max(plots_b["Y"])
|
418 |
-
for k1 in plots_k1["X"]:
|
419 |
-
ranking = get_ranking(k1, max_b, counting)
|
420 |
-
plots_k1["Y"].append(evaluate_map(rankings=ranking))
|
421 |
-
## YOU_CODE_ENDS_HERE
|
422 |
-
|
423 |
-
## TEST_CASES (should be close to 0.8135637188208616 and 0.7512916099773244)
|
424 |
-
print(plots_k1["Y"][9])
|
425 |
-
print(plots_b["Y"][1])
|
426 |
-
|
427 |
-
## RESULT_CHECKING_POINT
|
428 |
-
print(plots_k1)
|
429 |
-
print(plots_b)
|
430 |
-
|
431 |
-
from matplotlib import pyplot as plt
|
432 |
-
plt.plot(plots_b["X"], plots_b["Y"], label="b")
|
433 |
-
plt.plot(plots_k1["X"], plots_k1["Y"], label="k1")
|
434 |
-
plt.ylabel("MAP")
|
435 |
-
plt.legend()
|
436 |
-
plt.grid()
|
437 |
-
plt.show()
|
438 |
-
|
439 |
-
"""Let's check the effectiveness gain on test after this tuning on dev"""
|
440 |
-
|
441 |
-
default_map = 0.7849
|
442 |
-
best_b = plots_b["X"][np.argmax(plots_b["Y"])]
|
443 |
-
best_k1 = plots_k1["X"][np.argmax(plots_k1["Y"])]
|
444 |
-
bm25_index = BM25Index.build_from_documents(
|
445 |
-
documents=iter(sciq.corpus),
|
446 |
-
ndocs=12160,
|
447 |
-
show_progress_bar=True,
|
448 |
-
k1=best_k1,
|
449 |
-
b=best_b
|
450 |
-
)
|
451 |
-
bm25_index.save("output/bm25_index")
|
452 |
-
bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
|
453 |
-
rankings = {}
|
454 |
-
for query in sciq.get_split_queries(Split.test): # note this is now on test
|
455 |
-
ranking = bm25_retriever.retrieve(query=query.text)
|
456 |
-
rankings[query.query_id] = ranking
|
457 |
-
optimized_map = evaluate_map(rankings, split=Split.test) # note this is now on test
|
458 |
-
print(default_map, optimized_map)
|
459 |
-
|
460 |
-
"""# TASK3: a search-engine demo based on Huggingface space (4 points)
|
461 |
-
|
462 |
-
## TASK3.1: create the gradio app (2 point)
|
463 |
-
|
464 |
-
Create a gradio app to demo the BM25 search engine index on SciQ. The app should have a single input variable for the query (of type `str`) and a single output variable for the returned ranking (of type `List[Hit]` in the code below). Please use the BM25 system with default k1 and b values.
|
465 |
-
|
466 |
-
Hint: it should use a "search" function of signature:
|
467 |
-
|
468 |
-
```python
|
469 |
-
def search(query: str) -> List[Hit]:
|
470 |
-
...
|
471 |
-
```
|
472 |
-
"""
|
473 |
-
|
474 |
-
import gradio as gr
|
475 |
-
from typing import TypedDict
|
476 |
-
|
477 |
class Hit(TypedDict):
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
demo: Optional[gr.Interface] = None # Assign your gradio demo to this variable
|
483 |
-
return_type = List[Hit]
|
484 |
|
485 |
-
## YOUR_CODE_STARTS_HERE
|
486 |
def search_sciq(query: str) -> List[Hit]:
|
487 |
results = bm25_retriever.retrieve(query)
|
488 |
-
|
489 |
for cid, score in results.items():
|
490 |
-
|
491 |
-
text = bm25_retriever.index.doc_texts[
|
492 |
-
|
|
|
493 |
|
494 |
-
|
495 |
|
496 |
demo = gr.Interface(
|
497 |
fn=search_sciq,
|
498 |
inputs="textbox",
|
499 |
-
outputs="
|
500 |
description="BM25 Search Engine Demo on SciQ Dataset"
|
501 |
)
|
502 |
-
## YOUR_CODE_ENDS_HERE
|
503 |
-
demo.launch()
|
504 |
|
|
|
|
|
505 |
|
|
|
1 |
# -*- coding: utf-8 -*-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
from dataclasses import dataclass
|
|
|
3 |
import os
|
4 |
+
import pickle
|
5 |
+
from typing import List, Dict, Optional, Type, TypeVar, TypedDict
|
|
|
|
|
6 |
import re
|
7 |
+
import math
|
8 |
+
from collections import Counter
|
9 |
+
import gradio as gr
|
10 |
import nltk
|
11 |
+
from nlp4web_codebase.ir.data_loaders.dm import Document
|
12 |
+
from nlp4web_codebase.ir.data_loaders.sciq import load_sciq
|
13 |
+
from nlp4web_codebase.ir.models import BaseRetriever
|
14 |
from nltk.corpus import stopwords as nltk_stopwords
|
15 |
|
16 |
+
# Check nltk stopwords data
|
17 |
+
try:
|
18 |
+
nltk.data.find("corpora/stopwords")
|
19 |
+
except LookupError:
|
20 |
+
nltk.download("stopwords", quiet=True)
|
21 |
+
|
22 |
+
# Tokenization and helper functions
|
23 |
LANGUAGE = "english"
|
|
|
24 |
stopwords = set(nltk_stopwords.words(LANGUAGE))
|
25 |
+
word_splitter = re.compile(r"(?u)\b\w\w+\b").findall
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
def simple_tokenize(text: str) -> List[str]:
|
28 |
+
words = word_splitter(text.lower())
|
29 |
+
tokenized = [word for word in words if word not in stopwords]
|
|
|
30 |
return tokenized
|
31 |
|
|
|
|
|
32 |
@dataclass
|
33 |
class PostingList:
|
34 |
+
term: str
|
35 |
+
docid_postings: List[int]
|
36 |
+
tweight_postings: List[float]
|
37 |
|
38 |
+
T = TypeVar("T", bound="InvertedIndex")
|
39 |
|
40 |
@dataclass
|
41 |
class InvertedIndex:
|
42 |
+
posting_lists: List[PostingList]
|
43 |
vocab: Dict[str, int]
|
44 |
+
cid2docid: Dict[str, int]
|
45 |
+
collection_ids: List[str]
|
46 |
+
doc_texts: Optional[List[str]] = None
|
47 |
|
48 |
def save(self, output_dir: str) -> None:
|
49 |
os.makedirs(output_dir, exist_ok=True)
|
|
|
52 |
|
53 |
@classmethod
|
54 |
def from_saved(cls: Type[T], saved_dir: str) -> T:
|
|
|
|
|
|
|
55 |
with open(os.path.join(saved_dir, "index.pkl"), "rb") as f:
|
56 |
+
return pickle.load(f)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
|
58 |
@dataclass
|
59 |
class BM25Index(InvertedIndex):
|
60 |
|
|
|
|
|
|
|
|
|
61 |
@staticmethod
|
62 |
def cache_term_weights(
|
63 |
+
posting_lists: List[PostingList], total_docs: int, avgdl: float, dfs: List[int], dls: List[int], k1: float, b: float,
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
) -> None:
|
|
|
|
|
65 |
N = total_docs
|
66 |
+
for tid, posting_list in enumerate(posting_lists):
|
|
|
|
|
67 |
idf = BM25Index.calc_idf(df=dfs[tid], N=N)
|
68 |
+
for i, docid in enumerate(posting_list.docid_postings):
|
|
|
69 |
tf = posting_list.tweight_postings[i]
|
70 |
dl = dls[docid]
|
71 |
+
posting_list.tweight_postings[i] = BM25Index.calc_regularized_tf(
|
72 |
tf=tf, dl=dl, avgdl=avgdl, k1=k1, b=b
|
73 |
+
) * idf
|
|
|
74 |
|
75 |
@staticmethod
|
76 |
+
def calc_regularized_tf(tf: int, dl: float, avgdl: float, k1: float, b: float) -> float:
|
|
|
|
|
77 |
return tf / (tf + k1 * (1 - b + b * dl / avgdl))
|
78 |
|
79 |
@staticmethod
|
|
|
82 |
|
83 |
@classmethod
|
84 |
def build_from_documents(
|
85 |
+
cls: Type[BM25Index], documents: List[Document], avgdl: float, total_docs: int, k1: float = 0.9, b: float = 0.4
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
) -> BM25Index:
|
87 |
+
# Assume run_counting() is defined to return counting object with relevant data
|
88 |
+
counting = run_counting(documents, simple_tokenize)
|
89 |
+
BM25Index.cache_term_weights(counting.posting_lists, total_docs, avgdl, counting.dfs, counting.dls, k1, b)
|
90 |
+
return cls(counting.posting_lists, counting.vocab, counting.cid2docid, counting.collection_ids, counting.doc_texts)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
91 |
|
92 |
+
class BM25Retriever(BaseRetriever):
|
93 |
def __init__(self, index_dir: str) -> None:
|
94 |
+
self.index = BM25Index.from_saved(index_dir)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
|
96 |
def retrieve(self, query: str, topk: int = 10) -> Dict[str, float]:
|
97 |
+
toks = simple_tokenize(query)
|
98 |
+
docid2score = Counter()
|
99 |
for tok in toks:
|
100 |
+
if tok in self.index.vocab:
|
101 |
+
tid = self.index.vocab[tok]
|
102 |
+
posting_list = self.index.posting_lists[tid]
|
103 |
+
for docid, weight in zip(posting_list.docid_postings, posting_list.tweight_postings):
|
104 |
+
docid2score[docid] += weight
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
105 |
return {
|
106 |
+
self.index.collection_ids[docid]: score for docid, score in docid2score.most_common(topk)
|
|
|
107 |
}
|
108 |
|
109 |
+
# Gradio app setup
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
110 |
class Hit(TypedDict):
|
111 |
+
cid: str
|
112 |
+
score: float
|
113 |
+
text: str
|
|
|
|
|
|
|
114 |
|
|
|
115 |
def search_sciq(query: str) -> List[Hit]:
|
116 |
results = bm25_retriever.retrieve(query)
|
117 |
+
hits = []
|
118 |
for cid, score in results.items():
|
119 |
+
docid = bm25_retriever.index.cid2docid[cid]
|
120 |
+
text = bm25_retriever.index.doc_texts[docid]
|
121 |
+
hits.append(Hit(cid=cid, score=score, text=text))
|
122 |
+
return hits
|
123 |
|
124 |
+
bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
|
125 |
|
126 |
demo = gr.Interface(
|
127 |
fn=search_sciq,
|
128 |
inputs="textbox",
|
129 |
+
outputs="json",
|
130 |
description="BM25 Search Engine Demo on SciQ Dataset"
|
131 |
)
|
|
|
|
|
132 |
|
133 |
+
if __name__ == "__main__":
|
134 |
+
demo.launch()
|
135 |
|