HiStruct+: Improving Extractive Text Summarization with Hierarchical Structure Information

Anonymous

Abstract: Transformer-based language models usually treat texts as linear sequences. However, most texts also have an inherent hierarchical structure, i.,e., parts of a text can be identified using their position in this hierarchy. In addition, section titles usually indicate the common topic of their respective sentences. We propose a novel approach to extract, encode and inject hierarchical structure (HiStruct) information into an extractive summarization model (HiStruct+ model) based on a pre-trained, encoder-only language model. Our HiStruct+ model achieves SOTA extractive ROUGE scores on three public summarization datasets (CNN/DailyMail, PubMed, arXiv), the improvement is especially substantial on PubMed and arXiv. Using various experimental settings, our HiStruct+ model outperforms a strong baseline, which differs from our model only in that the HiStruct information is not injected. The ablation study demonstrates that the hierarchical position information is the main contributor to our model's SOTA performance.

Paper Type: long