SumRecom: A Personalized Summarization Approach by Learning from Users' Feedback

TL;DR

SumRecom addresses the challenge of user specific multi-document summarization by introducing a human in the loop framework that learns user interests from concept level preferences without relying on reference summaries. It integrates a user preference extractor employing active preference learning with a summarizer that uses inverse reinforcement learning to learn rewards and reinforcement learning to optimize the summary policy. The content selection is further guided by an ILP based generator, with a cross- input reward signal learned from domain expertise, and reinforced by a policy learner to produce user tailored outputs. Empirical results on DUC datasets, including a human study, demonstrate improved personalized summaries and lower cognitive burden, highlighting practical relevance for information overload scenarios and suggesting avenues for implicit preference learning and cross domain transfer.

Abstract

Existing multi-document summarization approaches produce a uniform summary for all users without considering individuals' interests, which is highly impractical. Making a user-specific summary is a challenging task as it requires: i) acquiring relevant information about a user; ii) aggregating and integrating the information into a user-model; and iii) utilizing the provided information in making the personalized summary. Therefore, in this paper, we propose a solution to a substantial and challenging problem in summarization, i.e., recommending a summary for a specific user. The proposed approach, called SumRecom, brings the human into the loop and focuses on three aspects: personalization, interaction, and learning user's interest without the need for reference summaries. SumRecom has two steps: i) The user preference extractor to capture users' inclination in choosing essential concepts, and ii) The summarizer to discover the user's best-fitted summary based on the given feedback. Various automatic and human evaluations on the benchmark dataset demonstrate the supremacy SumRecom in generating user-specific summaries. Document summarization and Interactive summarization and Personalized summarization and Reinforcement learning.

Figures (8)

• Figure 1: An overview of the SumRecom Approach: Active and preference learning are used to extract user's preferences. The learned preference ranking function is used to produce the desired summary using inverse reinforcement learning (IRL) for learning the reward and reinforcement learning for learning the optimal policy (RL).
• Figure 2: Example of comparing two summaries that put a substantial cognitive burden on users.
• Figure 3: SumRecom approach in more detail: 1) The left side is the user preference extractor using active preference learning over concepts, and 2) the right side is the summarizer including reward learning (IRL) and policy learning (RL).
• Figure 4: Features for estimating the ranker function for preferred concepts.
• Figure 5: Comparing different strategies used in active learning.