Automatic Survey Paper Generation

Overview

Survey paper generation faces great challenges from Cognitive Linguistics and Natural Language Processing. We investigate this problem with automatic generation only from references. In the view of Cognitive Linguistics, references taxonomy and the writing logic are non-trivial, while the difficulty in NLP lies in the document-level text generation with well-organized structure. Inspired by the writing process of human beings, we split this problem into three sub-problems: automatic references taxonomy, survey outline generation, and section content generation. We first proposed a pre-trained language model by adding pre-train tasks to extract information for sub-problems. Then for each sub-problem, we propose several possible solutions and attempt to tackle their challenging issues. Compared with existing works, the innovation parts are (1) our results preserve the hierarchical structure of survey papers, which provides clear content organization, and (2) we integrate the knowledge of writing logic in the generation process to make the result more readable.

Project Framework

Demo

Achievements

• Shuaiqi Liu, Jiannong Cao, Ruosong Yang, Zhiyuan Wen: Highlight-Transformer: Leveraging Key Phrase Aware Attention to Improve Abstractive Multi-Document Summarization. Accepted by Findings of ACL-IJCNLP 2021

Members

Zhiyuan Wen, Ruosong Yang, Shuaiqi Liu