Unsupervised Extractive Summarization-Based Representations for Accurate and Explainable Collaborative Filtering

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2981–2990 August 1–6, 2021. ©2021 Association for Computational Linguistics 2981 Unsupervised Extractive Summarization-Based Representations for Accurate and Explainable Collaborative Filtering Reinald Adrian Pugoy1,2 and Hung-Yu Kao1 1Intelligent Knowledge Management Lab National Cheng Kung University, Tainan City, Taiwan 2Faculty of Information and Communication Studies University of the Philippines Open University, Los Ba˜nos, Philippines rdpugoy@up.edu.ph, hykao@mail.ncku.edu.tw Abstract We pioneer the first extractive summarization- based collaborative filtering model called ES- COFILT. Our proposed model specifically pro- duces extractive summaries for each item and user. Unlike other types of explanations, summary-level explanations closely resemble real-life explanations. The strength of ES- COFILT lies in the fact that it unifies repre- sentation and explanation. In other words, ex- tractive summaries both represent and explain the items and users. Our model uniquely inte- grates BERT, K-Means embedding clustering, and multilayer perceptron to learn sentence embeddings, representation-explanations, and user-item interactions, respectively. We argue that our approach enhances both rating pre- diction accuracy and user/item explainability. Our experiments illustrate that ESCOFILT’s prediction accuracy is better than the other state-of-the-art recommender models. Further- more, we propose a comprehensive set of cri- teria that assesses the real-life explainability of explanations. Our explainability study demon- strates the superiority of and preference for summary-level explanations over other expla- nation types. 1 Introduction Collaborative filtering (CF) approaches are the most dominant and outstanding models in recom- mender systems literature. CF mainly focuses on learning accurate representations of users and items, denoting user preferences and item charac- teristics, respectively (Chen et al., 2018; Tay et al., 2018). The earliest CF models learned such rep- resentations based on user-given numeric ratings, but employing them is an oversimplification of user preferences and item characteristics (Koren et al., 2009; Musto et al., 2017). In this regard, review texts have been utilized to alleviate this issue. Reviews Received by the ‘Journaling Bible’ Item 1. I was not expecting this Bible to be so beautiful when I pre-ordered it 5 months ago, but it arrived in the mail today and it is just gor- geous! I love the concept of Bible journaling, but was always a bit intimidated by where/how to start. This removes that concern through some beautifully done artwork and lettering. I am ecstatic at the quality of this Bible! 2. I brought this as I wanted a separate Bible to do Bible journaling. It is very beautiful and has many images that can be coloured. The pages are similar to Bible paper and cream in colour. Overall a wonderful Bible to do journaling and meditate God’s Word. Generated Explanations • Review-Level: I brought this as I wanted a separate Bible to do Bible journaling. It is very beautiful and has many images that can be coloured. The pages are similar to Bible paper and cream in colour. Overall a wonderful Bible to do journaling and meditate God’s Word. • Word-Level: I brought this as …

Content truncated to 500 words