bm25-pt

A minimal implementation of Okapi BM25 using PyTorch. (Also uses HuggingFace tokenizers behind the scenes to tokenize text.)

pip install bm25_pt

Usage

from bm25_pt import BM25

bm25 = BM25()
corpus = [
    "A high weight in tf–idf is reached by a high term frequency",
    "(in the given document) and a low document frequency of the term",
    "in the whole collection of documents; the weights hence tend to filter",
    "out common terms. Since the ratio inside the idf's log function is always",
    "greater than or equal to 1, the value of idf (and tf–idf) is greater than or equal",
    "to 0. As a term appears in more documents, the ratio inside the logarithm approaches",
    "1, bringing the idf and tf–idf closer to 0.",
]
bm25.index(corpus)

queries = ["weights", "ratio logarithm"]
doc_scores = bm25.score_batch(queries)
print(doc_scores)
>> tensor([[0.0000, 0.0000, 1.4238, 0.0000, 0.0000, 0.0000, 0.0000],
        [0.0000, 0.0000, 0.0000, 1.5317, 0.0000, 2.0203, 0.0000]])

can also call score() with a single example:

doc_scores = bm25.score(queries[0])
print(doc_scores)
>> tensor([0.0000, 0.0000, 1.4238, 0.0000, 0.0000, 0.0000, 0.0000])

Use your own tokenizer

You can use your own tokenizer if you want. Simply provide your tokenizer to the BM25 constructor:

from bm25_pt import BM25
import transformers

tokenizer = transformers.AutoTokenizer.from_pretrained("t5-base")
bm25 = BM25(tokenizer=tokenizer)

Use a GPU

To perform operations on a GPU, it's easy, just pass a device to initialization:

from bm25_pt import BM25

bm25 = BM25(device='cuda')

Benchmarking on LoCO

I benchmarked this library on the Stanford LoCo benchmark. I'm not using any special tokenizer, so it's just using bert-base-uncased and matching based on subword tokens. We could probably get better performance by using a word-based model, perhaps by using a cased model, and certainly by tuning hyperparameters, especially b and k1.

Retrieval Encoders	Seq Len	Tau Scrolls – Summ.	Tau Scrolls – Gov.	Tau Scrolls QMSUM	QASPER - Title	QASPER - Abstract	Average
bm25_pt	infinity	89.5	96.1	56.2	94.3	99.4	87.1
M2-BERT-2048 (80M)	2048	81.8	94.7	58.5	87.3	95.5	83.6
M2-BERT-8192 (80M)	8192	94.7	96.5	64.1	86.8	97.5	85.9
M2-BERT-32768 (80M)	32768	98.6	98.5	69.5	97.4	98.7	92.5

As you can see below, this library is very fast, even on CPU:

Dataset	Size	Index time (s)	Scoring time (s)
Tau Scrolls – Summ.	3673	43.2	14.5
Tau Scrolls – Gov.	17,457	622.7	328.2
Tau Scrolls QMSUM	272	5.5	0.2
QASPER - Title	416	6.1	0.2
QASPER - Abstract	416	5.9	0.2

I think that most of the time is spent tokenizing long documents. It could likely be sped up with a faster tokenizer and/or multiprocessing.