Multilingual Keyphrase Generation

Code & Dataset for our NAACL 2022 Findings paper: Retrieval-Augmented Multilingual Keyphrase Generation with Retriever-Generator Iterative Training https://aclanthology.org/2022.findings-naacl.92/

The dataset is released under CDLA-Permissive-2.0.

If you use our data, please cite our paper as follows:

@inproceedings{gao-etal-2022-retrieval,
    title = "Retrieval-Augmented Multilingual Keyphrase Generation with Retriever-Generator Iterative Training",
    author = "Gao, Yifan  and
      Yin, Qingyu  and
      Li, Zheng  and
      Meng, Rui  and
      Zhao, Tong  and
      Yin, Bing  and
      King, Irwin  and
      Lyu, Michael",
    booktitle = "Findings of the Association for Computational Linguistics: NAACL 2022",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-naacl.92",
    doi = "10.18653/v1/2022.findings-naacl.92",
    pages = "1233--1246",
    abstract = "Keyphrase generation is the task of automatically predicting keyphrases given a piece of long text. Despite its recent flourishing, keyphrase generation on non-English languages haven{'}t been vastly investigated. In this paper, we call attention to a new setting named multilingual keyphrase generation and we contribute two new datasets, EcommerceMKP and AcademicMKP, covering six languages. Technically, we propose a retrieval-augmented method for multilingual keyphrase generation to mitigate the data shortage problem in non-English languages. The retrieval-augmented model leverages keyphrase annotations in English datasets to facilitate generating keyphrases in low-resource languages. Given a non-English passage, a cross-lingual dense passage retrieval module finds relevant English passages. Then the associated English keyphrases serve as external knowledge for keyphrase generation in the current language. Moreover, we develop a retriever-generator iterative training algorithm to mine pseudo parallel passage pairs to strengthen the cross-lingual passage retriever. Comprehensive experiments and ablations show that the proposed approach outperforms all baselines.",
}

Code

https://github.com/amzn/multilingual-keyphrase-generation

Dataset in E-Commerce Domain (EcommerceMKP)

A multilingual keyphrase generation dataset on the e-commerce domain, built on Amazon product data. It includes (passage, keyphrases) pairs in Spanish, French, German and Italian.

We have splitted the dataset into train, dev and test set (./e-commerce/<split>.json). Each instance in our dataset is in the jsonl format including the following entries:

• id: the unique id for the instance, it is the md5 hashing result of the product ASIN id.

• title: the title of the product.

• context: the passage description of the product.

• keywords: the keywords of the product.

• lang: the language of current instance.

In addition, we provide the Amazon English keyphrase generation dataset (./e-commerce/en.json) for experimental replication purpose. The instances in the en.json are of the same format with our multilingual dataset. The product appears in different languages share the same id.

Dataset in Academic Domain (AcademicMKP)

A multilingual keyphrase generation dataset on the academic domain, built on Microsoft Academic Graph. It includes (passage, keyphrases) pairs in Korean and Chinese.

We have splitted the dataset into train, dev and test set (./academic/<split>.json). Each instance in our dataset is in the jsonl format including the following entries:

• asin: the unique id for the instance.

• context: the passage description of the product.

• keywords: the keywords of the abstract.

• lang: the language of current instance.

In addition, we provide the our processed KP20K (./academic/en.json) for experimental replication purpose.