DKCS: A Dual Knowledge-Enhanced Abstractive Cross-Lingual Summarization Method Based on Graph Attention Networks

Shuyu Jiang, Dengbiao Tu, Xingshu Chen, Rui Tang, Wenxian Wang, Haizhou Wang

Abstract: Cross-Lingual Summarization (CLS) is the task of generating summaries in a target language for source articles in a different language. Previous studies on CLS mainly take pipeline methods or train an attention-based end-to-end model on translated parallel datasets. However, challenges arising from lengthy sources and non-parallel mappings hamper the accurate summarization and translation of pivotal information. To address this, this paper proposes a novel Dual Kknowledge-enhanced abstractive CLS model (DKCS) via a graph-encoder-decoder architecture. DKCS implements a clue-focused graph encoder that utilizes a graph attention network to simultaneously capture inter-sentence structures and significant information guided by extracted salient internal knowledge. Additionally, a bilingual lexicon is introduced in the decoder with an attention layer for enhanced translation. We construct the first hand-written CLS dataset for evaluation as well. Experimental results demonstrate the model’s robustness and significant performance gains over the existing SOTA on both automatic and human evaluations.