Topic-controllable Abstractive Summarization

Abstract

Existing approaches for topic-controllable summarization either incorporate topic embeddings or modify the attention mechanism. The incorporation of such approaches in a particular summarization model requires the adaptation of its codebase, a process that can be complex and time-consuming. Instead, we propose a model-agnostic topic-controllable summarization method employing a simple tagging-based formulation that can effortlessly work with any summarization model. In addition, we propose a new topic-oriented evaluation measure to quantitatively evaluate the generated summaries based on the topic affinity between the generated summary and the desired topic. Experimental results show that the proposed tagging-based formulation can achieve similar or even better performance compared to the embedding-based approach, while being at the same time significantly faster.

Tasks

Reproductions