Controllable Summarization with Constrained Markov Decision Process

Abstract We study controllable text summarization, which allows users to gain control on a particular attribute (e.g., length limit) of the generated summaries. In this work, we propose novel training framework based Constrained Markov Decision Process (CMDP), conveniently includes reward function along with set constraints, facilitate better summarization control. The encourages generation resemble human-written reference, while constraints are used explicitly prevent summaries from violating user-imposed requirements. Our can be applied important attributes including length, covered entities, and abstractiveness, as devise specific for each these aspects. Extensive experiments popular benchmarks show that our CMDP helps generate informative complying given attribute’s requirement.1