Datasets:

Modalities:
Tabular
Text
Formats:
parquet
ArXiv:
Libraries:
Datasets
Dask
kwang2049 commited on
Commit
df23a7a
1 Parent(s): 66b5d3b

readme: retrieval and evaluation

Browse files
Files changed (3) hide show
  1. .gitignore +2 -1
  2. README.md +81 -0
  3. retrieval.py +74 -0
.gitignore CHANGED
@@ -1 +1,2 @@
1
- ./loading.py
 
 
1
+ ./loading.py
2
+ ./retrieval.py
README.md CHANGED
@@ -244,5 +244,86 @@ for qrel in qrels:
244
  qrel["url"] # url to the document in Wikipedia
245
  ```
246
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
247
  ## Note
248
  This dataset was created with `datasets==2.15.0`. Make sure to use this or a newer version of the datasets library.
 
244
  qrel["url"] # url to the document in Wikipedia
245
  ```
246
 
247
+ ## Retrieval and Evaluation
248
+ The following shows an example, how the dataset can be used to build a semantic search application.
249
+ > This example is based on [clddp](https://github.com/kwang2049/clddp/tree/main) (`pip install -U cldpp`). One can further explore this [example](https://github.com/kwang2049/clddp/blob/main/examples/search_fiqa.sh) for convenient multi-GPU exact search.
250
+
251
+ ```python
252
+ # Please install cldpp with `pip install -U cldpp`
253
+ from clddp.retriever import Retriever, RetrieverConfig, Pooling, SimilarityFunction
254
+ from clddp.dm import Separator
255
+ from typing import Dict
256
+ from clddp.dm import Query, Passage
257
+ import torch
258
+ import pytrec_eval
259
+ import numpy as np
260
+ from datasets import load_dataset
261
+
262
+
263
+ # Define the retriever (DRAGON+ from https://arxiv.org/abs/2302.07452)
264
+ class DRAGONPlus(Retriever):
265
+ def __init__(self) -> None:
266
+ config = RetrieverConfig(
267
+ query_model_name_or_path="facebook/dragon-plus-query-encoder",
268
+ passage_model_name_or_path="facebook/dragon-plus-context-encoder",
269
+ shared_encoder=False,
270
+ sep=Separator.blank,
271
+ pooling=Pooling.cls,
272
+ similarity_function=SimilarityFunction.dot_product,
273
+ query_max_length=512,
274
+ passage_max_length=512,
275
+ )
276
+ super().__init__(config)
277
+
278
+
279
+ # Load data:
280
+ passages = load_dataset("kwang2049/dapr", "ConditionalQA-corpus", split="test")
281
+ queries = load_dataset("kwang2049/dapr", "ConditionalQA-queries", split="test")
282
+ qrels_rows = load_dataset("kwang2049/dapr", "ConditionalQA-qrels", split="test")
283
+ qrels: Dict[str, Dict[str, float]] = {}
284
+ for qrel_row in qrels_rows:
285
+ qid = qrel_row["query_id"]
286
+ pid = qrel_row["corpus_id"]
287
+ rel = qrel_row["score"]
288
+ qrels.setdefault(qid, {})
289
+ qrels[qid][pid] = rel
290
+
291
+ # Encode queries and passages: (refer to https://github.com/kwang2049/clddp/blob/main/examples/search_fiqa.sh for multi-GPU exact search)
292
+ retriever = DRAGONPlus()
293
+ retriever.eval()
294
+ queries = [Query(query_id=query["_id"], text=query["text"]) for query in queries]
295
+ passages = [
296
+ Passage(passage_id=passage["_id"], text=passage["text"]) for passage in passages
297
+ ]
298
+ query_embeddings = retriever.encode_queries(queries)
299
+ with torch.no_grad(): # Takes around a minute on a V100 GPU
300
+ passage_embeddings, passage_mask = retriever.encode_passages(passages)
301
+
302
+ # Calculate the similarities and keep top-K:
303
+ similarity_scores = torch.matmul(
304
+ query_embeddings, passage_embeddings.t()
305
+ ) # (query_num, passage_num)
306
+ topk = torch.topk(similarity_scores, k=10)
307
+ topk_values: torch.Tensor = topk[0]
308
+ topk_indices: torch.LongTensor = topk[1]
309
+ topk_value_lists = topk_values.tolist()
310
+ topk_index_lists = topk_indices.tolist()
311
+
312
+ # Run evaluation with pytrec_eval:
313
+ retrieval_scores: Dict[str, Dict[str, float]] = {}
314
+ for query_i, (values, indices) in enumerate(zip(topk_value_lists, topk_index_lists)):
315
+ query_id = queries[query_i].query_id
316
+ retrieval_scores.setdefault(query_id, {})
317
+ for value, passage_i in zip(values, indices):
318
+ passage_id = passages[passage_i].passage_id
319
+ retrieval_scores[query_id][passage_id] = value
320
+ evaluator = pytrec_eval.RelevanceEvaluator(
321
+ query_relevance=qrels, measures=["ndcg_cut_10"]
322
+ )
323
+ query_performances: Dict[str, Dict[str, float]] = evaluator.evaluate(retrieval_scores)
324
+ ndcg = np.mean([score["ndcg_cut_10"] for score in query_performances.values()])
325
+ print(ndcg) # 0.21796083196880855
326
+ ```
327
+
328
  ## Note
329
  This dataset was created with `datasets==2.15.0`. Make sure to use this or a newer version of the datasets library.
retrieval.py ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Please install cldpp with `pip install -U cldpp`
2
+ from clddp.retriever import Retriever, RetrieverConfig, Pooling, SimilarityFunction
3
+ from clddp.dm import Separator
4
+ from typing import Dict
5
+ from clddp.dm import Query, Passage
6
+ import torch
7
+ import pytrec_eval
8
+ import numpy as np
9
+ from datasets import load_dataset
10
+
11
+
12
+ # Define the retriever (DRAGON+ from https://arxiv.org/abs/2302.07452)
13
+ class DRAGONPlus(Retriever):
14
+ def __init__(self) -> None:
15
+ config = RetrieverConfig(
16
+ query_model_name_or_path="facebook/dragon-plus-query-encoder",
17
+ passage_model_name_or_path="facebook/dragon-plus-context-encoder",
18
+ shared_encoder=False,
19
+ sep=Separator.blank,
20
+ pooling=Pooling.cls,
21
+ similarity_function=SimilarityFunction.dot_product,
22
+ query_max_length=512,
23
+ passage_max_length=512,
24
+ )
25
+ super().__init__(config)
26
+
27
+
28
+ # Load data:
29
+ passages = load_dataset("kwang2049/dapr", "ConditionalQA-corpus", split="test")
30
+ queries = load_dataset("kwang2049/dapr", "ConditionalQA-queries", split="test")
31
+ qrels_rows = load_dataset("kwang2049/dapr", "ConditionalQA-qrels", split="test")
32
+ qrels: Dict[str, Dict[str, float]] = {}
33
+ for qrel_row in qrels_rows:
34
+ qid = qrel_row["query_id"]
35
+ pid = qrel_row["corpus_id"]
36
+ rel = qrel_row["score"]
37
+ qrels.setdefault(qid, {})
38
+ qrels[qid][pid] = rel
39
+
40
+ # Encode queries and passages: (refer to https://github.com/kwang2049/clddp/blob/main/examples/search_fiqa.sh for multi-GPU exact search)
41
+ retriever = DRAGONPlus()
42
+ retriever.eval()
43
+ queries = [Query(query_id=query["_id"], text=query["text"]) for query in queries]
44
+ passages = [
45
+ Passage(passage_id=passage["_id"], text=passage["text"]) for passage in passages
46
+ ]
47
+ query_embeddings = retriever.encode_queries(queries)
48
+ with torch.no_grad(): # Takes around a minute on a V100 GPU
49
+ passage_embeddings, passage_mask = retriever.encode_passages(passages)
50
+
51
+ # Calculate the similarities and keep top-K:
52
+ similarity_scores = torch.matmul(
53
+ query_embeddings, passage_embeddings.t()
54
+ ) # (query_num, passage_num)
55
+ topk = torch.topk(similarity_scores, k=10)
56
+ topk_values: torch.Tensor = topk[0]
57
+ topk_indices: torch.LongTensor = topk[1]
58
+ topk_value_lists = topk_values.tolist()
59
+ topk_index_lists = topk_indices.tolist()
60
+
61
+ # Run evaluation with pytrec_eval:
62
+ retrieval_scores: Dict[str, Dict[str, float]] = {}
63
+ for query_i, (values, indices) in enumerate(zip(topk_value_lists, topk_index_lists)):
64
+ query_id = queries[query_i].query_id
65
+ retrieval_scores.setdefault(query_id, {})
66
+ for value, passage_i in zip(values, indices):
67
+ passage_id = passages[passage_i].passage_id
68
+ retrieval_scores[query_id][passage_id] = value
69
+ evaluator = pytrec_eval.RelevanceEvaluator(
70
+ query_relevance=qrels, measures=["ndcg_cut_10"]
71
+ )
72
+ query_performances: Dict[str, Dict[str, float]] = evaluator.evaluate(retrieval_scores)
73
+ ndcg = np.mean([score["ndcg_cut_10"] for score in query_performances.values()])
74
+ print(ndcg) # 0.21796083196880855