Optimizing Representation for Abstractive Multidocument Summarization Based on Adversarial Learning Strategy
Bin Cao, Xinxin Guan, Songlin Bao, Jiawei Wu, Jing Fan · IEEE Transactions on Cognitive and Developmental Systems · 2025
Abstractive Multi-Document Summarization (MDS) is a crucial technique in cognitive computing, enabling the efficient synthesis of a documents cluster into a concise and complete summary. Despite recent advances, existing approaches still face challenges in representation learning when processing large-scale documents clusters: (1) incomplete semantic learning caused by documents truncation or exclusion; (2) the incorporation of noise, such as irrelevant or redundant information from documents; and (3) the potential omission of critical content due to partial coverage of documents. These limitations collectively undermine the semantic integrity and conciseness of the generated summaries. To address these issues, we proposeTALER, a two-stage representation architecture enhanced by adversarial learning for abstractive MDS, which reformulates the MDS task as a single-document optimization problem. In Stage I,TALERfocuses on enhancing single-document representations by maximizing semantic learning from each document in the cluster and employing the adversarial learning to suppress the introduction of documents noise. In Stage II,TALERconducts multi-document semantic fusion and summary generation by aggregating the learned document embeddings based on Stage I into a cluster-level representation through a pooling mechanism, followed by a self-attention module to capture salient content and produce the final summary. Experimental results on the Multi-News, DUC04, and Multi-XScience datasets demonstrate thatTALERconsistently outperforms existing baseline models across multiple evaluation metrics.