Joint Contrastive Learning for Factual Consistency Evaluation of Cross-Lingual Abstract Summarization

Current summarization models tend to generate erroneous or irrelevant summaries, i.e., factual inconsistency, which undoubtedly hinders the real-world application of summarization models. The difficulty in language alignment makes factual inconsistency in cross-lingual summarization (CLS) more common and factual consistency checking more challenging. Research on factual consistency has paid little attention to CLS due to the above difficulties, focusing mainly on monolingual summarization (MS). In this paper, we investigate the cross-lingual domain and propose a weakly supervised factual consistency evaluation model for CLS. In particular, we automatically synthesize large-scale datasets by a series of rule-based text transformations and manually annotate the test and validation sets. In addition, we also train the model jointly with contrastive learning to enhance the model’s ability to recognize factual errors. The experimental results on the manually annotated test set show that our model can effectively identify the consistency between the summaries and the source documents and outperform the baseline models.