JMTEB-fixed

このリポジトリは元のJMTEBデータセットのUTF-8エンコーディングエラーを修正したバージョンです。

修正内容

問題

NLP Journal LaTeXコーパスの一部ファイル（例: V28N02-25.tex）が異なるエンコーディング（Shift-JIS、EUC-JPなど）で保存されており、UTF-8デコーディングエラーが発生していました。

解決策

retrieval.pyのNLPJournalHelper.load_txtメソッドを修正し、複数のエンコーディングを順次試行するようにしました。

使用方法

from datasets import load_dataset

dataset = load_dataset(
    "ks-pf/JMTEB-fixed",
    name="nlp_journal_title_abs-corpus",
    trust_remote_code=True
)


# JMTEB: Japanese Massive Text Embedding Benchmark

JMTEB is a benchmark for evaluating Japanese text embedding models. It consists of 6 tasks, currently involving 24 datasets in total.

## TL;DR

```python
from datasets import load_dataset

dataset = load_dataset("sbintuitions/JMTEB", name="<dataset_name>", split="<split>")

JMTEB_DATASET_NAMES = (
    'livedoor_news',
    'mewsc16_ja',
    'amazon_review_classification',
    'amazon_counterfactual_classification',
    'massive_intent_classification',
    'massive_scenario_classification',
    'jsts',
    'jsick',
    'paws_x_ja',
    'jaqket-query',
    'jaqket-corpus',
    'mrtydi-query',
    'mrtydi-corpus',
    'jagovfaqs_22k-query',
    'jagovfaqs_22k-corpus',
    'nlp_journal_title_abs-query',
    'nlp_journal_title_abs-corpus',
    'nlp_journal_title_intro-query',
    'nlp_journal_title_intro-corpus',
    'nlp_journal_abs_intro-query',
    'nlp_journal_abs_intro-corpus',
    'nlp_journal_abs_article-query',
    'nlp_journal_abs_article-corpus',
    'jacwir-retrieval-query',
    'jacwir-retrieval-corpus',
    'miracl-retrieval-query',
    'miracl-retrieval-corpus',
    'mldr-retrieval-query',
    'mldr-retrieval-corpus',
    'esci-query',
    'esci-corpus',
    'jqara-query',
    'jqara-corpus',
    'jacwir-reranking-query',
    'jacwir-reranking-corpus',
    'miracl-reranking-query',
    'miracl-reranking-corpus',
    'mldr-reranking-query',
    'mldr-reranking-corpus',
)

Introduction

We introduce JMTEB (Japanese Massive Text Embedding Benchmark), an evaluation benchmark including 6 tasks (Clustering, Classification, STS, PairClassification, Retrieval and Reranking). 24 datasets in total are collected to conduct these tasks. Similar with MTEB, we aim to provide a diverse and extensible evaluation benchmark for Japanese embedding models, enabling more insightful analysis on model performance, thus benefitting the emerging of more powerful models.

We also provide an easy-to-use evaluation script to perform the evaluation just with a one-line command. Refer to https://https://github.com/sbintuitions/JMTEB-eval-scripts.

We encourage anyone interested to contribute to this benchmark!

Tasks and Datasets

Here is an overview of the tasks and datasets currently included in JMTEB.

Task	Dataset	Train	Dev	Test	Document (Retrieval)
Clustering	Livedoor-News	5,163	1,106	1,107	-
	MewsC-16-ja	-	992	992	-
Classification	AmazonCounterfactualClassification	5,600	466	934	-
	AmazonReviewClassification	200,000	5,000	5,000	-
	MassiveIntentClassification	11,514	2,033	2,974	-
	MassiveScenarioClassification	11,514	2,033	2,974	-
STS	JSTS	12,451	-	1,457	-
	JSICK	5,956	1,985	1,986	-
PairClassification	PAWS-X-ja	49,401	2,000	2,000	-
Retrieval	JAQKET	13,061	995	997	114,229
	Mr.TyDi-ja	3,697	928	720	7,000,027
	NLP Journal title-abs	-	100	404	504
	NLP Journal title-intro	-	100	404	504
	NLP Journal abs-intro	-	100	404	504
	NLP Journal abs-abstract	-	100	404	504
	JaGovFaqs-22k	15,955	3,419	3,420	22,794
	JaCWIR-Retrieval	-	1,000	4,000	513,107
	MIRACL-Retrieval	2,433	1,044	860	6,953,614
	MLDR-Retrieval	2,262	200	200	10,000
Reranking	Esci	10,141	1,790	4,206	149,999
	JaCWIR-Reranking	-	1,000	4,000	513,107
	JQaRA	498	1,737	1,667	250,609
	MIRACL-Reranking	2,433	1,044	860	37,124
	MLDR-Reranking	2,262	200	200	5,339

Clustering

The goal of the Clustering task is to correctly distribute texts with similar semantics/topic to the same cluster. It is an unsupervised process in evaluating embedding models. We have 2 datasets for Clustering.

Livedoor News

Livedoor News is a dataset collected from the news reports of a Japanese news site by RONDHUIT Co, Ltd. in 2012. It contains over 7,000 news report texts across 9 categories (topics).

The dataset is licensed under CC BY-ND 2.1.

MewsC-16-ja

MewsC-16-ja is the Japanese split of MewsC-16 dataset, which consists of topic sentences from Wikinews. It has 12 types of topics.

Classification

Classification aims to predict the correct category of the text only with its dense representation. Typically, the process is conducted with supervised learning that employs statistical models like linear regression and k-NN.

AmazonCounterfactualClassification

We use the Japanese split of Amazon Multiligual Counterfactual Dataset in MTEB, which contains sentences from Amazon customer review texts. It is a binary classification of the text is/isn't a statement that describes an event that did not or cannot take place. For more details, refer to https://huggingface.co/datasets/mteb/amazon_counterfactual.

This dataset is made available under the Creative Commons Attribution-ShareAlike 4.0 International License. Refer to following page for the license information of this dataset: https://github.com/amazon-science/amazon-multilingual-counterfactual-dataset?tab=License-1-ov-file.

AmazonReviewClassification

We use the Japanese split of the Multiligual Amazon Review Corpus in MTEB. The dataset is a 5-classification of customer rating on a product, according with a review paragraph. For more details, refer to https://huggingface.co/datasets/amazon_reviews_multi.

Refer to following page for the license information of this dataset: https://docs.opendata.aws/amazon-reviews-ml/readme.html.

MassiveIntentClassification

We use the Japanese split of MASSIVE dataset. This dataset is built with Alexa user utterance and the corresponding intent. It is a 60-classification. For more detail as well as the license information, refer to https://github.com/alexa/massive.

MassiveScenarioClassification

We use the Japanese split of MASSIVE dataset. This dataset is built with Alexa user utterance and the corresponding scenario. It is an 18-classification. The texts are the same as MassiveIntentClassification. For more detail as well as the license information, refer to https://github.com/alexa/massive.

STS

STS (Semantic Textual Similarity) unsupervisedly predicts the semantic similarity between two sentences, and correlations are computed between the prediction and the annotated similarity.

JSTS

JSTS, a part of JGLUE, is a Japanese version of STS dataset. The sentences are extracted from the Japanese version of the MS COCO Caption Dataset, the YJ Captions Dataset (Miyazaki and Shimizu, 2016). Refer to https://github.com/yahoojapan/JGLUE/blob/main/README.md#jsts for more details.

This dataset is licensed under Creative Commons Attribution Share Alike 4.0 International.

JSICK

JSICK is the Japanese NLI and STS dataset by manually translating the English dataset SICK (Marelli et al., 2014) into Japanese. We use merely the STS part.

This work is licensed under a Creative Commons Attribution 4.0 International License.

PairClassification

PairClassification is a task to predict a label (typically binary) that indicates whether two sentences constitute a parapharse pair, utilizing the best binary threshold accuracy or F1.

PAWS-X-ja

PAWS-X-ja is the Japanese split of PAWS-X, which is a multiligual paraphrase identification dataset.

Regarding the license, the dataset may be freely used for any purpose, although acknowledgement of Google LLC ("Google") as the data source would be appreciated.

Retrieval

The retrieval task aims to find the most relevant document with the query from the corpus, through the computing of embedding similarities.

JAQKET

JAQKET (AIO Ver. 1.0) dataset has a quiz set and a corpus that consists of Wikipedia passages, each is a description is an entity (the title of the Wikipedia page). A quiz question is answered by looking for the most relevant Wikipedia passage with the quiz question text. For more details, refer to https://www.nlp.ecei.tohoku.ac.jp/projects/jaqket/.

The copyright for the quiz questions in the train subset belongs to the abc/EQIDEN Committee and redistributed from Tohoku University for non-commercial research purposes. This validation/test subset is licensed under CC BY-SA 4.0 DEED.

Mr.TyDi-ja

Mr.TyDi-ja is the Japanese split of Mr.TyDi, a multilingual benchmark dataset built on TyDi. The goal is to find the relevant documents with the query text. For more details, refer to https://huggingface.co/datasets/castorini/mr-tydi.

This dataset is licensed under Apache-2.0.

NLP Journal title-abs

NLP Journal title-intro

NLP Journal abs-intro

NLP Journal abs-article

These datasets are created with the Japanese NLP Journal LaTeX Corpus. We shuffled the titles, abstracts and introductions of the academic papers, and the goal is to find the corresponding abstract with the given title / introduction with the given title / introduction / full article with the given abstract, through the similarities computed with text embeddings.

These datasets are licensed under CC-BY-4.0, according to the Manuscript Guide to Journal Publication.

JaGovFaqs-22k

JaGovFaqs-22k is a dataset consisting of FAQs manully extracted from the website of Japanese bureaus. We shuffled the queries (questions) and corpus (answers), and the goal is to match the answer with the question.

This dataset is licensed under CC-BY-4.0.

JaCWIR-Retrieval

JaCWIR (Japanese Casual Web IR Dataset) is a dataset consisting of questions and webpage meta description texts collected from Hatena Bookmark. Passages that contain various genres are collected with RSS, and corresponding questions are generated with ChatGPT-3.5. JaCWIR-Retrieval reformats JaCWIR data for retrieval task.

Refer to this link for the detail of the license of JaCWIR.

MIRACL-Retrieval

MIRACL 🌍🙌🌏 (Multilingual Information Retrieval Across a Continuum of Languages) is a multilingual retrieval dataset that focuses on search across 18 different languages, which collectively encompass over three billion native speakers around the world. In JMTEB, we use the Japanese split of MIRACL. MIRACL inherits from Mr.TyDi-ja. MIRACL-Retrieval is the reformatted version of MIRACL for retrieval task.

This dataset is licensed under Apache-2.0.

MLDR-Retrieval

MLDR is a Multilingual Long-Document Retrieval dataset built on Wikipeida, Wudao and mC4, covering 13 typologically diverse languages. Specifically, we sample lengthy articles from Wikipedia, Wudao and mC4 datasets and randomly choose paragraphs from them. Then we use GPT-3.5 to generate questions based on these paragraphs. The generated question and the sampled article constitute a new text pair to the dataset. MLDR-Retrieval is the reformatted version of MLDR (Japanese split) for retrieval task.

This dataset is licensed under MIT.

Reranking

The reranking task aims to rerank the retrieved documents through computing embedding similarities.

Esci

Amazon esci is a dataset consisting of retrieval queries and products information on Amazon. For each data, the relevance between query and product is annotated with E(Exact), S(Substitute), C(Complement), and I(Irrelevant).　Each relevance label is given a different score, allowing for more detailed scoring. We employed product titles and descriptions as product information and excluded data without descriptions.

This dataset is Apache-2.0.

JQaRA

JQaRA (Japanese Question Answering with Retrieval Augmentation) is a reranking dataset consisting of questions processed from JAQKET and corpus from Japanese Wikipedia. There are 100 passages for each question, where multiple relevant passages in the 100 are relevant with the question.

This dataset is licensed with CC-BY-SA-4.0.

JaCWIR-Reranking

JaCWIR (Japanese Casual Web IR Dataset) is a dataset consisting of questions and webpage meta description texts collected from Hatena Bookmark. Passages that contain various genres are collected with RSS, and corresponding questions are generated with ChatGPT-3.5. JaCWIR-Reranking reformats JaCWIR data for reranking task. 1 out of 100 passages is relevant with the question.

Refer to this link for the detail of the license of JaCWIR.

MIRACL-Reranking

MIRACL 🌍🙌🌏 (Multilingual Information Retrieval Across a Continuum of Languages) is a multilingual retrieval dataset that focuses on search across 18 different languages, which collectively encompass over three billion native speakers around the world. In JMTEB, we use the Japanese split of MIRACL. MIRACL inherits from Mr.TyDi-ja. MIRACL-Reranking is the reformatted version of MIRACL for reranking task. One or multiple passages are relevant with the question.

This dataset is licensed under Apache-2.0.

MLDR-Reranking

MLDR is a Multilingual Long-Document Retrieval dataset built on Wikipeida, Wudao and mC4, covering 13 typologically diverse languages. Specifically, we sample lengthy articles from Wikipedia, Wudao and mC4 datasets and randomly choose paragraphs from them. Then we use GPT-3.5 to generate questions based on these paragraphs. The generated question and the sampled article constitute a new text pair to the dataset. MLDR-Reranking is the reformatted version of MLDR (Japanese split) for reranking task.

This dataset is licensed under MIT.

Reference

@misc{jmteb,
    author = {Li, Shengzhe and Ohagi, Masaya and Ri, Ryokan},
    title = {{J}{M}{T}{E}{B}: {J}apanese {M}assive {T}ext {E}mbedding {B}enchmark},
    howpublished = {\url{https://huggingface.co/datasets/sbintuitions/JMTEB}},
    year = {2024},
}

License

Our code is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License.

Regarding the license information of datasets, please refer to the individual datasets.