Automatic News Headline Generation and Topic Consistency Assessment Based on SimCSE and Knowledge Graph

Abstract

To address the problems of insufficient semantic alignment and topic drift in industry news and market report headline generation, this paper proposes a method that fuses SimCSE semantic constraints with knowledge graph entity associations, paying particular attention to key entities related to names of people, organizations, locations, and fields. Considering the growing demand for reliable information processing in intelligent communication systems and electromagnetic information transmission environments, the proposed framework provides a semantic enhancement mechanism that is beneficial for trustworthy content understanding in data-driven applications. Based on SimCSE, the news text is encoded at the sentence level, and contrastive learning is used to improve semantic representation consistency. Named entity recognition is performed and aligned with the Wikidata knowledge graph to construct a local knowledge subgraph containing entity relationships. A graph attention network is used to fuse SimCSE sentence embeddings with knowledge graph entity embeddings to generate knowledge-enhanced representations. The fused representation is applied during the Transformer decoding process to dynamically focus on key entities and core semantics, generating thematically consistent and factually accurate headlines. Experimental results show that the proposed method achieves an average BERTScore of at least 0.87 and an average BLEU score of at least 0.75 across different news domains. In terms of topic consistency, the Topic-F1 score remains at 0.835 ± 0.019 under the highcomplexity condition of four topics, and the topic deviation is as low as 0.165 ± 0.021. In scenarios involving the summarization of complex manufacturing processes or market trend analysis, this study effectively addresses the issues of insufficient semantic alignment and topic drift while providing a reference for semantic information processing in intelligent electromagnetic communication systems.