A Cross-Lingual Summarization method based on cross-lingual Fact-relationship Graph Generation

Abstract

The aim of cross-lingual summarization (CLS) is to condense the content of a document in one language into a summary in another language. In essence, a CLS model requires both translation and summarization capabilities, which presents a unique challenge, as the model must effectively tackle the difficulties associated with both tasks simultaneously (e.g., semantic alignment, information compression and factual inconsistency). Graph-based semantic representation can model important text information in a structured manner, which may alleviate these challenges. Therefore, in this paper, we propose a Cross-Lingual Summarization method based on cross-lingual Fact-relationship Graph Generation (FGGCLS). Specifically, we first construct fact-relationship graphs for source language documents and target language summaries. Then, we introduce a cross-lingual fact-relationship graph generation method, which converts the CLS problem into a cross-lingual fact-relationship graph generation problem. This approach simplifies semantic alignment and information compression through the generation of graphs and leads to improved fact consistency. Finally, the generated fact-relationship graph of the target language summary serves as a draft for generating the summary, which enhances the quality of the generated summary. We conduct systematic experiments on the Zh2EnSum and En2ZhSum datasets, and the results demonstrate that our method can effectively improve the performance of CLS and alleviate factual inconsistency.