bm25s-j 0.2.0

Japanese version of bm25s.

bm25s-j

本プロジェクトはbm25sの日本語対応版です。 標準でjanomeを使用し、日本語のストップワードを適当にいれたものです。 またBM25におけるテキストの長さによるスコアへの影響を緩和するため、対数正規化を入れ、影響を緩和するオプションを追加しました。(use_log_normalization=True,デフォルトでON) Stemmerは特に対応していません。 オリジナルがMITライセンスですので、こちらもMITライセンスです。

時間があればトークナイザもより高速なものを使用するかもしれませんが、まずは環境構築が簡単なようにPythonだけで動くようにしています。

オリジナルプロジェクト

https://github.com/xhluca/bm25s

インストール

pip install bm25s-j

# オリジナルのbm25sの使用方法と同じです。
pip install bm25s-j[core]

# top-k選択のプロセスの高速化で`jax`をインストールします。
pip install jax[cpu]

使用方法

オリジナルのbm25sの使用方法と同じです。

# サンプルコード
import bm25s

def main():
    # https://ja.wikipedia.org/wiki/%E6%9A%81%E7%BE%8E%E3%81%BB%E3%82%80%E3%82%89
    # コーパスは日本語のテキストです。
    corpus = [
        "暁美 ほむら（あけみ ほむら）は、テレビアニメ『魔法少女まどか☆マギカ』に登場する架空の人物。まどか☆マギカの外伝漫画・『魔法少女おりこ☆マギカ』、『魔法少女まどか☆マギカ 〜The different story〜』『魔法少女まどか☆マギカ [魔獣編]』にも登場する。",
        "「時間操作」の魔法を操る魔法少女として設定されており、劇中では人間社会から持ち出した銃や爆弾の数々を時間操作能力と組み合わせて戦っている。",
        "劇中で直接そのように呼ばれる場面はないが[2][注 1]、ファンからは「ほむほむ」という愛称で呼ばれている[2][5]。",
        "一人称は「私」。",
        "まどかは「まどか」、さやかは「美樹さやか」、マミは「巴マミ」、杏子は「杏子」と呼び、まどかと杏子以外の魔法少女はフルネームで呼び捨てにしている。",
        "声優は各作品共通で斎藤千和（英語版はクリスティーナ・ヴィー）が担当する。『マギアレコード 魔法少女まどか☆マギカ外伝』の舞台版では河田陽菜（けやき坂46（現・日向坂46））が演じる[6]。",
    ]
    
    # コーパスを日本語でトークン化します。
    corpus_tokens = bm25s.tokenize(corpus, stopwords="japanese")
    print(corpus_tokens)

    # BM25を作成します。
    retriever = bm25s.BM25() # 文書長の対数正規化補正をしない場合、use_log_normalization=False
    retriever.index(corpus_tokens)

    # クエリを日本語でトークン化します。
    query = "ほむらは誰？"
    query_tokens = bm25s.tokenize(query, stopwords="japanese")

    # クエリで取得した結果を表示します。
    results, scores = retriever.retrieve(query_tokens, k=2)
    for i in range(results.shape[1]):
        doc, score = results[0, i], scores[0, i]
        print(f"Rank {i+1} (score: {score:.2f}): {doc}")

    # 保存します
    retriever.save("retriever_magical_girl.pkl")

    # ロードします
    retriever = bm25s.BM25.load("retriever_magical_girl.pkl")

    # ロード後に変更ないのを確認
    query = "ほむらは誰？"
    query_tokens = bm25s.tokenize(query, stopwords="japanese")
    print(query_tokens)

    results, scores = retriever.retrieve(query_tokens, k=2)
    for i in range(results.shape[1]):
        doc, score = results[0, i], scores[0, i]
        print(f"Rank {i+1} (score: {score:.2f}): {doc}")

if __name__ == "__main__":
    main()

🚀 変更履歴 (Changelog)

バージョン 0.2.0 (2024-12-20)

🎉 新機能

• 高度なメタデータフィルタリング機能: 検索時に文書のメタデータを利用した高度なフィルタリングが可能になりました
  • 基本的な equality フィルタリング ({"category": "tech"})
  • 複数条件 AND フィルタリング ({"category": "tech", "language": "ja"})
  • 比較演算子サポート ({"score": {"$gte": 0.8}, "date": {"$lt": "2024-01-01"}})
  • 配列操作 ({"tags": {"$in": ["AI", "ML"]}, "category": {"$nin": ["draft"]}})
  • 存在チェック ({"author": {"$exists": True}})
  • 正規表現フィルタリング ({"title": {"$regex": "Python.*Tutorial"}})
  • 論理演算子 ({"$or": [{"category": "tech"}, {"difficulty": "beginner"}]})
  • 複雑なネストした条件の完全サポート

💾 使用例

import bm25s

# メタデータ付きでインデックス作成
corpus = ["文書1", "文書2", "文書3"]
metadata = [
    {"category": "tech", "difficulty": "beginner", "score": 0.9},
    {"category": "science", "difficulty": "advanced", "score": 0.8},
    {"category": "tech", "difficulty": "intermediate", "score": 0.95}
]

retriever = bm25s.BM25()
corpus_tokens = bm25s.tokenize(corpus, stopwords="japanese")
retriever.index(corpus_tokens, metadata=metadata)

# 高度なフィルタリング検索
query_tokens = bm25s.tokenize("検索クエリ", stopwords="japanese")
results = retriever.retrieve(
    query_tokens, 
    k=5,
    filter={
        "category": "tech",
        "score": {"$gte": 0.9},
        "difficulty": {"$in": ["beginner", "intermediate"]}
    }
)

🔧 技術的改善

• BM25スコアリングシステムの統合問題を修正
• フィルタリング適用時のスコア計算バグを解決
• Weight mask適用の問題を修正
• Numbaバックエンドでのweight_mask パラメータサポートを追加
• $exists: False オペレーターの動作を修正
• 動的k調整による境界エラーの防止
• 空の結果に対する適切なハンドリングを実装

🧪 テスト強化

• メタデータフィルタリングの包括的なテストスイートを追加
• 高度な演算子のテストケースを充実
• 複雑なネスト条件のテスト検証を実装

📋 互換性

• 既存APIとの完全な後方互換性を維持
• メタデータなしでの従来通りの動作を保証

---

以下、オリジナルのbm25sのREADME

BM25-Sparse⚡

BM25S is an ultrafast implementation of BM25 in pure Python, powered by Scipy sparse matrices

💻 GitHub

🏠 Homepage

📝 Technical Report

🤗 Blog Post

🛠️ Installation

Welcome to bm25s, a library that implements BM25 in Python, allowing you to rank documents based on a query. BM25 is a widely used ranking function used for text retrieval tasks, and is a core component of search services like Elasticsearch.

It is designed to be:

• Fast: bm25s is implemented in pure Python and leverage Scipy sparse matrices to store eagerly computed scores for all document tokens. This allows extremely fast scoring at query time, improving performance over popular libraries by orders of magnitude (see benchmarks below).
• Simple: bm25s is designed to be easy to use and understand. You can install it with pip and start using it in minutes. There is no dependencies on Java or Pytorch - all you need is Scipy and Numpy, and optional lightweight dependencies for stemming.

Below, we compare bm25s with Elasticsearch in terms of speedup over rank-bm25, the most popular Python implementation of BM25. We measure the throughput in queries per second (QPS) on a few popular datasets from BEIR in a single-threaded setting.

[!IMPORTANT] New in version 0.2.0: We are rolling out support for a numba backend, which gives around 2x speedup for larger datasets! Learn more about it and share your thoughts in the version 0.2.0 release thread.

Show/Hide citation

@misc{bm25s,
      title={BM25S: Orders of magnitude faster lexical search via eager sparse scoring}, 
      author={Xing Han Lù},
      year={2024},
      eprint={2407.03618},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2407.03618},
}

Installation

You can install bm25s with pip:

pip install bm25s

If you want to use stemming for better results, you can install the recommended (but optional) dependencies:

# Install all extra dependencies
pip install bm25s[full]

# If you want to use stemming for better results, you can install a stemmer
pip install PyStemmer

# To speed up the top-k selection process, you can install `jax`
pip install jax[cpu]

Quickstart

Here is a simple example of how to use bm25s:

import bm25s
import Stemmer  # optional: for stemming

# Create your corpus here
corpus = [
    "a cat is a feline and likes to purr",
    "a dog is the human's best friend and loves to play",
    "a bird is a beautiful animal that can fly",
    "a fish is a creature that lives in water and swims",
]

# optional: create a stemmer
stemmer = Stemmer.Stemmer("english")

# Tokenize the corpus and only keep the ids (faster and saves memory)
corpus_tokens = bm25s.tokenize(corpus, stopwords="en", stemmer=stemmer)

# Create the BM25 model and index the corpus
retriever = bm25s.BM25()
retriever.index(corpus_tokens)

# Query the corpus
query = "does the fish purr like a cat?"
query_tokens = bm25s.tokenize(query, stemmer=stemmer)

# Get top-k results as a tuple of (doc ids, scores). Both are arrays of shape (n_queries, k).
# To return docs instead of IDs, set the `corpus=corpus` parameter.
results, scores = retriever.retrieve(query_tokens, k=2)

for i in range(results.shape[1]):
    doc, score = results[0, i], scores[0, i]
    print(f"Rank {i+1} (score: {score:.2f}): {doc}")

# You can save the arrays to a directory...
retriever.save("animal_index_bm25")

# You can save the corpus along with the model
retriever.save("animal_index_bm25", corpus=corpus)

# ...and load them when you need them
import bm25s
reloaded_retriever = bm25s.BM25.load("animal_index_bm25", load_corpus=True)
# set load_corpus=False if you don't need the corpus

For an example that shows how to quickly index a 2M-documents corpus (Natural Questions), check out examples/index_nq.py.

Flexibility

bm25s provides a flexible API that allows you to customize the BM25 model and the tokenization process. Here are some of the options you can use:

# You can provide a list of queries instead of a single query
queries = ["What is a cat?", "is the bird a dog?"]

# Provide your own stopwords list if you don't like the default one
stopwords = ["a", "the"]

# For stemming, use any function that is callable on each word list
stemmer_fn = lambda lst: [word for word in lst]

# Tokenize the queries
query_token_ids = bm25s.tokenize(queries, stopwords=stopwords, stemmer=stemmer_fn)

# If you want the tokenizer to return strings instead of token ids, you can do this
query_token_strs = bm25s.tokenize(queries, return_ids=False)

# You can use a different corpus for retrieval, e.g., titles instead of full docs
titles = ["About Cat", "About Dog", "About Bird", "About Fish"]

# You can also choose to only return the documents and omit the scores
# note: if you pass a new corpus here, it must have the same length as your indexed corpus
results = retriever.retrieve(query_token_ids, corpus=titles, k=2, return_as="documents")

# The documents are returned as a numpy array of shape (n_queries, k)
for i in range(results.shape[1]):
    print(f"Rank {i+1}: {results[0, i]}")

Memory Efficient Retrieval

bm25s is designed to be memory efficient. You can use the mmap option to load the BM25 index as a memory-mapped file, which allows you to load the index without loading the full index into memory. This is useful when you have a large index and want to save memory:

# Create a BM25 index
# ...

# let's say you have a large corpus
corpus = [
    "a very long document that is very long and has many words",
    "another long document that is long and has many words",
    # ...
]
# Save the BM25 index to a file
retriever.save("bm25s_very_big_index", corpus=corpus)

# Load the BM25 index as a memory-mapped file, which is memory efficient
# and reduce overhead of loading the full index into memory
retriever = bm25s.BM25.load("bm25s_very_big_index", mmap=True)

For an example of how to use retrieve using the mmap=True mode, check out examples/retrieve_nq.py.

Tokenization

In addition to using the simple function bm25s.tokenize, you can also use the Tokenizer class to customize the tokenization process. This is useful when you want to use a different tokenizer, or when you want to use a different tokenization process for queries and documents:

from bm25s.tokenization import Tokenizer

corpus = [
      "a cat is a feline and likes to purr",
      "a dog is the human's best friend and loves to play",
      "a bird is a beautiful animal that can fly",
      "a fish is a creature that lives in water and swims",
]

# Pick your favorite stemmer, and pass 
stemmer = None
stopwords = ["is"]
splitter = lambda x: x.split() # function or regex pattern
# Create a tokenizer
tokenizer = Tokenizer(
      stemmer=stemmer, stopwords=stopwords, splitter=splitter
)

corpus_tokens = tokenizer.tokenize(corpus)

# let's see what the tokens look like
print("tokens:", corpus_tokens)
print("vocab:", tokenizer.get_vocab_dict())

# note: the vocab dict will either be a dict of `word -> id` if you don't have a stemmer, and a dict of `stemmed word -> stem id` if you do.
# You can save the vocab. it's fine to use the same dir as your index if filename doesn't conflict
tokenizer.save_vocab(save_dir="bm25s_very_big_index")

# loading:
new_tokenizer = Tokenizer(stemmer=stemmer, stopwords=[], splitter=splitter)
new_tokenizer.load_vocab("bm25s_very_big_index")
print("vocab reloaded:", new_tokenizer.get_vocab_dict())

# the same can be done for stopwords
print("stopwords before reload:", new_tokenizer.stopwords)
tokenizer.save_stopwords(save_dir="bm25s_very_big_index")
new_tokenizer.load_stopwords("bm25s_very_big_index")
print("stopwords reloaded:", new_tokenizer.stopwords)

You can find advanced examples in examples/tokenizer_class.py, including how to:

• Pass a stemmer, stopwords, and splitter function/regex pattern
• Control whether vocabulary is updated by tokenizer.tokenize calls or not (by default, it will only be updated during the first call)
• Reset the tokenizer to its initial state with tokenizer.reset_vocab()
• Use the tokenizer in generator mode to save memory by yielding one document at a time.
• Pass different outputs of the tokenizer to the BM25.retrieve function.

Variants

You can use the following variants of BM25 in bm25s (see Kamphuis et al. 2020 for more details):

• Original implementation (method="robertson") - we set idf>=0 to avoid negatives
• ATIRE (method="atire")
• BM25L (method="bm25l")
• BM25+ (method="bm25+")
• Lucene (method="lucene")

By default, bm25s uses method="lucene", which is Lucene's BM25 implementation (exact version). You can change the method by passing the method argument to the BM25 constructor:

# The IR book recommends default values of k1 between 1.2 and 2.0, and b=0.75
retriever = bm25s.BM25(method="robertson", k1=1.5, b=0.75)

# For BM25+, BM25L, you need a delta parameter (default is 0.5)
retriever = bm25s.BM25(method="bm25+", delta=1.5)

# You can also choose a different "method" for idf, while keeping the default for the rest
# for example, this is equivalent to rank-bm25 when `epsilon=0`
retriever = bm25s.BM25(method="atire", idf_method="robertson")
# and this is equivalent to bm25-pt
retriever = bm25s.BM25(method="atire", idf_method="lucene")

Hugging Face Integration

bm25 can naturally work with Hugging Face's huggingface_hub, allowing you to load and save to the model hub. This is useful for sharing BM25 indices and using community models.

First, make sure you have a valid access token for the Hugging Face model hub. This is needed to save models to the hub, or to load private models. Once you created it, you can add it to your environment variables (e.g. in your .bashrc or .zshrc):

export HUGGING_FACE_HUB_TOKEN="hf_..."

Now, let's install the huggingface_hub library:

pip install huggingface_hub

Let's see how to use BM25SHF.save_to_hub to save a BM25 index to the Hugging Face model hub:

import os
import bm25s
from bm25s.hf import BM25HF

# Create a BM25 index
retriever = BM25HF()
# Create your corpus here
corpus = [
    "a cat is a feline and likes to purr",
    "a dog is the human's best friend and loves to play",
    "a bird is a beautiful animal that can fly",
    "a fish is a creature that lives in water and swims",
]
corpus_tokens = bm25s.tokenize(corpus)
retriever.index(corpus_tokens)

# Set your username and token
user = "your-username"
token = os.environ["HF_TOKEN"]
retriever.save_to_hub(f"{user}/bm25s-animals", token=token, corpus=corpus)
# You can also save it publicly with private=False

Then, you can use the following code to load a BM25 index from the Hugging Face model hub:

import bm25s
from bm25s.hf import BM25HF

# Load a BM25 index from the Hugging Face model hub
user = "your-username"
retriever = BM25HF.load_from_hub(f"{user}/bm25s-animals")

# you can specify revision and load_corpus=True if needed
retriever = BM25HF.load_from_hub(
    f"{user}/bm25s-animals", revision="main", load_corpus=True
)

# if you want a low-memory usage, you can load as memory map with `mmap=True`
retriever = BM25HF.load_from_hub(
    f"{user}/bm25s-animals", load_corpus=True, mmap=True
)

# Query the corpus
query = "does the fish purr like a cat?"

# Tokenize the query
query_tokens = bm25s.tokenize(query)

# Get top-k results as a tuple of (doc ids, scores). Both are arrays of shape (n_queries, k)
results, scores = retriever.retrieve(query_tokens, k=2)

For a complete example, check out:

• examples/index_to_hf.py for indexing a corpus and upload to Huggingface Hub
• examples/retrieve_from_hf.py for loading an index alongside corpus from Huggingface Hub and querying it.

Comparison

Here are some benchmarks comparing bm25s to other popular BM25 implementations. We compare the following implementations:

• bm25s: Our implementation of BM25 in pure Python, powered by Scipy sparse matrices.
• rank-bm25 (Rank): A popular Python implementation of BM25.
• bm25_pt (PT): A Pytorch implementation of BM25.
• elasticsearch (ES): Elasticsearch with BM25 configurations.

OOM means the implementation ran out of memory during the benchmark.

Throughput (Queries per second)

We compare the throughput of the BM25 implementations on various datasets. The throughput is measured in queries per second (QPS), on a single-threaded Intel Xeon CPU @ 2.70GHz (found on Kaggle). For BM25S, we take the average of 10 runs. Instances exceeding 60 queries/s are in bold.

Dataset	BM25S	Elastic	BM25-PT	Rank-BM25
arguana	573.91	13.67	110.51	2
climate-fever	13.09	4.02	OOM	0.03
cqadupstack	170.91	13.38	OOM	0.77
dbpedia-entity	13.44	10.68	OOM	0.11
fever	20.19	7.45	OOM	0.06
fiqa	507.03	16.96	20.52	4.46
hotpotqa	20.88	7.11	OOM	0.04
msmarco	12.2	11.88	OOM	0.07
nfcorpus	1196.16	45.84	256.67	224.66
nq	41.85	12.16	OOM	0.1
quora	183.53	21.8	6.49	1.18
scidocs	767.05	17.93	41.34	9.01
scifact	952.92	20.81	184.3	47.6
trec-covid	85.64	7.34	3.73	1.48
webis-touche2020	60.59	13.53	OOM	1.1

More detailed benchmarks can be found in the bm25-benchmarks repo.

Disk usage

bm25s is designed to be lightweight. This means the total disk usage of the package is minimal, as it only requires wheels for numpy (18MB), scipy (37MB), and the package itself is less than 100KB. After installation, the full virtual environment takes more space than rank-bm25 but less than pyserini and bm25_pt:

Package	Disk Usage
venv (no package)	45MB
rank-bm25	99MB
bm25s (ours)	479MB
bm25_pt	5346MB
pyserini	6976MB
elastic	1183MB

Show Details

The disk usage of the virtual environments is calculated using the following command:

$ du -s *env-* --block-size=1MB
6976    conda-env-pyserini
5346    venv-bm25-pt
479     venv-bm25s
45      venv-empty
99      venv-rank-bm25

For pyserini, we use the recommended installation with conda environment to account for Java dependencies.

Optimized RAM usage

bm25s allows considerable memory saving through the use of memory-mapping, which allows the index to be stored on disk and loaded on demand.

Using the index_nq.py to create an index, we can retrieve with:

• examples/retrieve_nq.py: setting mmap=False in the main function to load the index in memory, and mmap=True to load the index as a memory-mapped file.
• examples/retrieve_nq_with_batching.py: This takes it a step further by batching the retrieval process, which allows for reloading the index after each batch (see Mmap+Reload below). This is useful when you have a large index and want to save memory.

We show the following results on the NQ dataset (2M+ documents, 100M+ tokens):

Method	Load Index (s)	Retrieval (s)	RAM post-index (GB)	RAM post-retrieve (GB)
In-memory	8.61	21.09	4.36	4.45
Memory-mapped	0.53	20.22	0.49	2.16
Mmap+Reload	0.48	20.96	0.49	0.70

We can see that memory-mapping the index allows for a significant reduction in memory usage, with comparable retrieval times.

Similarly, for MSMARCO (8M+ documents, 300M+ tokens), we show the following results (running on the validation set), although the retrieval did not complete for the in-memory case:

Method	Load Index (s)	Retrieval (s)	RAM post-index (GB)	RAM post-retrieve (GB)
In-memory	25.71	93.66	10.21	10.34
Memory-mapped	1.24	90.41	1.14	4.88
Mmap+Reload	1.17	97.89	1.14	1.38

Acknowledgement

• The central idea behind the scoring mechanism in this library is originally from bm25_pt, which was a major inspiration to this project.
• The API of the BM25 class is also heavily inspired by the design of BM25-pt, as well as that of rank-bm25.
• The multilingual stopwords are sourced from the NLTK stopwords lists.
• The numba implementation are inspired by numba implementations originally proposed by baguetter and retriv.
• The function bm25s.utils.beir.evaluate is taken from the BEIR library. It follows the same license as the BEIR library, which is Apache 2.0.

Citation

If you use bm25s in your work, please use the following bibtex:

@misc{bm25s,
      title={BM25S: Orders of magnitude faster lexical search via eager sparse scoring}, 
      author={Xing Han Lù},
      year={2024},
      eprint={2407.03618},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2407.03618}, 
}