Automatic Title Generation for Materials Science Literature Based on Pointer-Generator Network

Abstract

With the rapid development of materials science literature, issues related to the organization and application of information are gradually coming to the forefront. To address the issue of automatic title generation, this paper proposes a domain-adaptive pointer generation network, specifically tailored to the diverse needs of materials science texts. To effectively address the issues of rare technical terms and complex language expressions, this paper employs a dual-stream encoder to incorporate domain-specific word embeddings and a dynamic gated pointer mechanism. The experiment used a large collection of well-organized materials from various fields. The proposed model has already surpassed traditional baseline methods. The BLEU score improved by 7.2%, the ROUGE-L score increased by 6.5%, and the average human evaluation increased by 0.17%. To reduce semantic drift and expand term coverage, all modules in these models are necessary. Titles are usually concise, technically accurate, and meet public expectations. In addition to providing a practical foundation for scientific titles, this article also proposes applying the system to other fields such as intelligent knowledge management and automated indexing.