| Zhang Rui (张 蕊)*,Wang Lijun* **,He Yanqing* **,Xu Hongjiao*,Lan Tian*,Xu Deshan*.[J].高技术通讯(英文),2026,32(2):202~216 |
|
| AcaAT: domain-aware academic alignment tuning for abstractive text summarization |
| |
| DOI：10. 3772 / j. issn. 1006-6748. 2026. 02. 009 |
| 中文关键词: |
| 英文关键词: academic paper summarization, domain-aware, low-rank adaptation tuning, prompt engineering |
| 基金项目: |
| Author Name | Affiliation | | Zhang Rui (张 蕊)* | (* Institute of Scientific and Technical Information of China, Beijing 100038, P. R. China)
(** Key Laboratory of Rich-media Knowledge Organization and Service of Digital Publishing Content,Beijing 100038, P. R. China) | | Wang Lijun* ** | | | He Yanqing* ** | | | Xu Hongjiao* | | | Lan Tian* | | | Xu Deshan* | |
|
| Hits: 122 |
| Download times: 125 |
| 中文摘要: |
| |
| 英文摘要: |
| Large language models (LLMs) have demonstrated exceptional capabilities in a wide range of natural language processing ( NLP) tasks. Text summarization, a fundamental NLP challenge,prompts numerous efforts to employ LLMs to improve summarization quality. However, current approaches usually focus on generating generalized summaries, and tasks within specialized areas like academic text summarization still remain unique challenges due to the specific domain knowledge. In this paper, we propose an academic summary generation that not only maintains the professional characteristics of specific academic domains but also adheres to the academic description format.Specifically, we first construct a new dataset of text summarization tasks on scientific research papers from 9 academic website databases, named SciDataSum. Then, we introduce an academic prompt pattern (AcaPP) to specify the domain knowledge within a range of domains of academic text, while ensuring the generated summary is adapted to an academic structure. We also propose an advanced framework incorporating low-rank adaptation (LoRA) fine-tuning to enable the pre-trained LLMs to learn the relevant domain knowledge and generate more precise summarizations. Experimental results show that our academic alignment taning (AcaAT) outperforms existing text summarization models on three benchmark datasets, and the generated abstract is described in the professional domain style. To bridge the gap between automatic text summarization and real-world scientific paper summaries, our AcaAT model aims to reduce the percentage of misinformation in the generated abstract. |
|
View Full Text
View/Add Comment Download reader |
| Close |
|
|
|
