Datasets:
Upload create_data_reranking.py
Browse files- create_data_reranking.py +82 -0
create_data_reranking.py
ADDED
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import datasets
|
2 |
+
from sentence_transformers import SentenceTransformer, util
|
3 |
+
import torch
|
4 |
+
from huggingface_hub import create_repo
|
5 |
+
from huggingface_hub.utils._errors import HfHubHTTPError
|
6 |
+
|
7 |
+
"""
|
8 |
+
To create a reranking dataset from the initial retrieval dataset,
|
9 |
+
we use a model (sentence-transformers/all-MiniLM-L6-v2) to embed the queries and the documents.
|
10 |
+
We then compute the cosine similarity for each query and document.
|
11 |
+
For each query we get the topk articles, as we would for a retrieval task.
|
12 |
+
Each couple query-document is labeled as relevant if it was labeled like so in the retrieval dataset,
|
13 |
+
or irrelevant if it was not
|
14 |
+
"""
|
15 |
+
# Download the documents (corpus)
|
16 |
+
corpus_raw = datasets.load_dataset("lyon-nlp/alloprof", "documents")
|
17 |
+
# Download the queries
|
18 |
+
queries_raw = datasets.load_dataset("lyon-nlp/alloprof", "queries")
|
19 |
+
# Get the model
|
20 |
+
model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
21 |
+
|
22 |
+
# Generate document text (title + content)
|
23 |
+
corpus = corpus_raw.map(lambda x: {"text": x["title"] + "\n\n" + x["text"]})
|
24 |
+
# Embed documents and queries
|
25 |
+
corpus = corpus.map(lambda x: {"embeddings": model.encode(x['text'])}, batched=True)
|
26 |
+
queries = queries_raw.map(lambda x: {"embeddings": model.encode(x["text"])}, batched=True)
|
27 |
+
|
28 |
+
# change document uuid with integer id
|
29 |
+
doc_name_id_mapping = {doc["uuid"]: i for i, doc in enumerate(corpus["documents"])}
|
30 |
+
corpus = corpus.map(lambda x: {"uuid" : doc_name_id_mapping[x["uuid"]]})
|
31 |
+
queries = queries.map(lambda x: {"relevant": [doc_name_id_mapping[r] for r in x["relevant"]]})
|
32 |
+
|
33 |
+
# Retrieve best documents by cosine similarity
|
34 |
+
def retrieve_documents(queries_embs, documents_embs, topk:int=10) -> torch.return_types.topk:
|
35 |
+
"""Finds the topk documents for each embed query among all the embed documents
|
36 |
+
|
37 |
+
Args:
|
38 |
+
queries_embs (_type_): the embedings of all queries of the dataset (dataset["queries"]["embeddings"])
|
39 |
+
documents_embs (_type_): the embedings of all coprus of the dataset (dataset["corpus"]["embeddings"])
|
40 |
+
topk (int, optional): The amount of top documents to retrieve. Defaults to 5.
|
41 |
+
|
42 |
+
Returns:
|
43 |
+
torch.return_types.topk : The topk object, with topk.values being the cosine similarities
|
44 |
+
and the topk.indices being the indices of best documents for each queries
|
45 |
+
"""
|
46 |
+
similarities = util.cos_sim(queries_embs, documents_embs)
|
47 |
+
tops = torch.topk(similarities, k=topk, axis=1)
|
48 |
+
|
49 |
+
return tops
|
50 |
+
|
51 |
+
top_docs_train = retrieve_documents(queries["train"]["embeddings"], corpus["documents"]["embeddings"])
|
52 |
+
top_docs_test = retrieve_documents(queries["test"]["embeddings"], corpus["documents"]["embeddings"])
|
53 |
+
queries["train"] = queries["train"].map(
|
54 |
+
lambda _, i: {"top_cosim_values": top_docs_train.values[i], "top_cosim_indexes": top_docs_train.indices[i]},
|
55 |
+
with_indices=True
|
56 |
+
)
|
57 |
+
queries["test"] = queries["test"].map(
|
58 |
+
lambda _, i: {"top_cosim_values": top_docs_test.values[i], "top_cosim_indexes": top_docs_test.indices[i]},
|
59 |
+
with_indices=True
|
60 |
+
)
|
61 |
+
|
62 |
+
# Remove id in best_indices if it corresponds to ground truth a relevant document
|
63 |
+
queries = queries.map(lambda x : {"top_cosim_indexes": [i for i in x["top_cosim_indexes"] if i not in x["relevant"]]})
|
64 |
+
# Convert document ids to document texts based on the corpus
|
65 |
+
queries = queries.map(lambda x: {"negative": [corpus["documents"][i]["text"] for i in x["top_cosim_indexes"]]})
|
66 |
+
queries = queries.map(lambda x: {"positive": [corpus["documents"][i]["text"] for i in x["relevant"]]})
|
67 |
+
|
68 |
+
# Format as the MTEB format
|
69 |
+
queries = queries.rename_column("text", "query")
|
70 |
+
dataset = queries.remove_columns(['embeddings', 'relevant', 'top_cosim_values', 'top_cosim_indexes', 'answer', 'subject', "id"])
|
71 |
+
# Rename the key of dataset key as "test"
|
72 |
+
# dataset["test"] = dataset.pop("queries")
|
73 |
+
|
74 |
+
# create HF repo
|
75 |
+
repo_id = "lyon-nlp/mteb-fr-reranking-alloprof-s2p"
|
76 |
+
try:
|
77 |
+
create_repo(repo_id, repo_type="dataset")
|
78 |
+
except HfHubHTTPError as e:
|
79 |
+
print("HF repo already exist")
|
80 |
+
|
81 |
+
# save dataset as json
|
82 |
+
dataset.push_to_hub(repo_id)
|