Interactive Query-Assisted Summarization via Deep Reinforcement Learning

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2551 - 2568 July 10-15, 2022 ©2022 Association for Computational Linguistics Interactive Query-Assisted Summarization via Deep Reinforcement Learning Ori Shapira1,3∗, Ramakanth Pasunuru2, Mohit Bansal2, Ido Dagan1, and Yael Amsterdamer1 1Bar-Ilan University 2UNC Chapel Hill 3Amazon obspp18@gmail.com {ram,mbansal}@cs.unc.edu {dagan,amstery}@cs.biu.ac.il Abstract Interactive summarization is a task that facil- itates user-guided exploration of information within a document set. While one would like to employ state of the art neural models to im- prove the quality of interactive summarization, many such technologies cannot ingest the full document set or cannot operate at sufficient speed for interactivity. To that end, we propose two novel deep reinforcement learning models for the task that address, respectively, the sub- task of summarizing salient information that adheres to user queries, and the subtask of list- ing suggested queries to assist users throughout their exploration.1 In particular, our models allow encoding the interactive session state and history to refrain from redundancy. Together, these models compose a state of the art solu- tion that addresses all of the task requirements. We compare our solution to a recent interac- tive summarization system, and show through an experimental study involving real users that our models are able to improve informativeness while preserving positive user experience. 1 Introduction Integrating human interaction into NLP tasks has been gaining the interest of the NLP community. Human-machine cooperation can improve the gen- eral quality of results, as well as provide a higher sense of control for the targeted consumer. We focus on the task of interactive summarization (INTSUMM: Shapira et al., 2021b) which enables information exploration within a document set on a topic, by means of user-guided summarization. As illustrated in Figure 1, a user can incrementally expand on a summary by submitting requests to the system, in order to expose the information of inter- est within the topic. A proper exploration session demands access to all information within the docu- ment set, and fast reaction time for smooth human ∗ This work was conducted prior to joining Amazon. 1Code and trained models at: https://github.com/ OriShapira/InterExp_DeepRL System initial summary User query 1 System response 1 User query 2 System response 2 Suggested Queries query - Process full docset - Low latency Figure 1: An INTSUMM system, ingesting a large doc- ument set. A user interactively submits queries in or- der to expand on the information. The system is re- quired to process the full document set for compre- hensive exploration, respond quickly, and expose non- redundant salient information that also complies to the input queries. See real example in Figure 5. engagement (Anderson, 2020; Attig et al., 2017). In addition, presented information must consider the session history to refrain from repetitiveness. While it is worthwhile to apply recent NLP ad- vances that excel at extracting salient and query- biased information, those advances usually come at a cost of rather small input size limits or heavy computation time. …

Content truncated to 500 words