Dataset construction method of cross-lingual summarization based on filtering and text augmentation

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Science

Background:

• Existing cross-lingual summarization (CLS) datasets suffer from inconsistent sample quality and limited scale.
• This hinders the development and evaluation of effective CLS models.

Purpose of the Study:

• To propose a novel method for jointly supervising quality and scale in CLS dataset construction.
• To address the limitations of existing CLS datasets by improving both data quality and size.

Main Methods:

• Implemented a multi-strategy filtering algorithm to remove low-quality monolingual summarization (MS) samples based on character and semantic analysis.
• Utilized a text augmentation algorithm leveraging pretrained models to expand the CLS dataset size while ensuring quality.

Main Results:

• Successfully constructed an English-Chinese CLS dataset using the proposed method.
• Evaluated the dataset using a robust quality evaluation framework, confirming good quality and large scale.
• Demonstrated that the method improves quality and scale comprehensively at a lower cost.

Conclusions:

• The proposed method offers an effective approach to building high-quality, large-scale CLS datasets.
• This contributes to advancing research and development in cross-lingual summarization.
• The method provides a cost-effective solution for dataset creation.