Sentence coherence evaluation based on neural network and textual features for official documents

Abstract

Sentence coherence is an essential foundation for discourse coherence in natural language processing, as it plays a vital role in enhancing language expression, text readability, and improving the quality of written documents. With the development of e-government, automatic generation of official documents can significantly reduce the writing burden of government agencies. To ensure that the automatically generated official documents are coherent, we propose a sentence coherence evaluation model integrating repetitive words features, which introduces repetitive words features with neural network-based approach for the first time. Experiments were conducted on official documents dataset and THUCNews public dataset, our method has achieved an averaged 3.8% improvement in accuracy indicator compared to past research, reaching a 96.2% accuracy rate. This result is significantly better than the previous best method, proving the superiority of our approach in solving this problem.

Keywords

• sentence coherence
• xl-net
• repetitive words
• feature fusion