Scholarlink

<p>Machine reading comprehension (MRC) requires models to integrate evidence scattered across long contexts. In many MRC settings, named entities provide strong linguistic anchors for linking distant evidence and narrowing candidate answer spans. Although pre-trained language models (PLMs) such as BERT achieve strong performance, they may not explicitly model sparse long-distance entity relations that span multiple sentences. We propose an entity-graph-enhanced MRC framework that constructs an entity co-occurrence graph from each input context (without relying on external knowledge bases), encodes the graph with a graph convolutional network (GCN), and fuses entity representations with PLM token representations for answer extraction. Two graph construction strategies—sentence-level co-occurrence and sliding-window co-occurrence weighted by normalized pointwise mutual information (NPMI)—are compared. Experiments on the DuReader dataset show consistent improvements over BERT/PERT/MacBERT baselines measured by BLEU-4 and ROUGE-L, and the sliding-window graph yields the best overall performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on Machine Reading Comprehension Integrating Long-Distance Entity Relations and Entity-Level Linguistic Cues

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月28日

查看全文
  • MA Ning,
  • Wu Di

摘要

Machine reading comprehension (MRC) requires models to integrate evidence scattered across long contexts. In many MRC settings, named entities provide strong linguistic anchors for linking distant evidence and narrowing candidate answer spans. Although pre-trained language models (PLMs) such as BERT achieve strong performance, they may not explicitly model sparse long-distance entity relations that span multiple sentences. We propose an entity-graph-enhanced MRC framework that constructs an entity co-occurrence graph from each input context (without relying on external knowledge bases), encodes the graph with a graph convolutional network (GCN), and fuses entity representations with PLM token representations for answer extraction. Two graph construction strategies—sentence-level co-occurrence and sliding-window co-occurrence weighted by normalized pointwise mutual information (NPMI)—are compared. Experiments on the DuReader dataset show consistent improvements over BERT/PERT/MacBERT baselines measured by BLEU-4 and ROUGE-L, and the sliding-window graph yields the best overall performance.