AMR Beam Search: generating faithful summaries with semantic graph representations

In recent years, notable strides have been made in abstractive summarization. However, a persistent challenge remains: many existing summarization models often produce summaries that lack semantic consistency with the source document. These models and evaluation methods tend to focus on word level, which may not necessarily bring results for being faithful. This thesis introduces a novel decoding strategy named AMR Beam Search (ABS) which leverages the potential of Abstract Meaning Representation (AMR). ABS is designed to integrate textual and structured semantic knowledge during the decoding process, promoting factual consistency between the generated summary and the source document. It accomplishes this by reranking hypotheses during decoding, where it computes the factuality score for beams completed through saliency-enhanced greedy decoding, considering both their corresponding source documents and AMRs in the process. Experiments conducted on 2 datasets, XSum and CNN/DM, demonstrate the superiority of the proposed ABS method. Across multiple faithfulness evaluation metrics, ABS outperforms the baseline method by an average gap of 0.77%, thus underscoring its effectiveness in achieving higher faithfulness in abstractive summarization.