Interactively learning to summarise timelines by reinforcement learning

Abstract

Timeline Summarisation is a special type of automatic summarisation task, in which temporal information is as important as summaries. It aims at generating a time-ordered event list containing concise event summaries and their precise happened times described in source documents like new articles. However, current approaches are neither capable of adjusting their content according to a certain user’s interest nor being adapted to new domains without a huge amount of reference timelines. Therefore, I propose a Reinforcement Learning-based interactive timeline summarisation system which can interactively learn from the user’s feedback to generate timelines that meet the user’s demands. I define a compound reward function that can update automatically according to the received feedback to ensure topical coherence, factual consistency and linguistic fluency of the generated summaries. The system use this reward function to adjust its output content via Reinforcement Learning to fine-tune an abstractive Multi-document Summarisation model. The advantage of my system is to avoid the need of using reference summaries in training. The experiments show that my system outperforms state of the art on the benchmark Timeline Summarisation dataset, Timeline17, and could generate accurate and precise timelines tailored for each user.

Date of Award	22 Mar 2022
Original language	English
Awarding Institution	University of Bristol
Supervisor	Edwin D. Simpson (Supervisor) & Dima Damen (Supervisor)