Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 188–197, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 188 Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects Jianmo Ni, Jiacheng Li, Julian McAuley University of California, San Diego {jin018,j9li,jmcauley}@ucsd.edu Abstract Several recent works have considered the problem of generating reviews (or ‘tips’) as a form of explanation as to why a recommen- dation might match a user’s interests. While promising, we demonstrate that existing ap- proaches struggle (in terms of both quality and content) to generate justifications that are relevant to users’ decision-making process. We seek to introduce new datasets and meth- ods to address this recommendation justifica- tion task. In terms of data, we first pro- pose an ‘extractive’ approach to identify re- view segments which justify users’ intentions; this approach is then used to distantly label massive review corpora and construct large- scale personalized recommendation justifica- tion datasets. In terms of generation, we design two personalized generation models with this data: (1) a reference-based Seq2Seq model with aspect-planning which can gen- erate justifications covering different aspects, and (2) an aspect-conditional masked language model which can generate diverse justifica- tions based on templates extracted from justi- fication histories. We conduct experiments on two real-world datasets which show that our model is capable of generating convincing and diverse justifications. 1 Introduction Explaining, or justifying, recommendations to users has the potential to increase their trans- parency and reliability. However providing mean- ingful interpretations remains a difficult task, partly due to the black-box nature of many rec- ommendation models, but also because we simply lack ground-truth datasets specifying what ‘good’ justifications ought to look like. Previous work has sought to learn user prefer- ences and writing styles from crowd-sourced re- views (Dong et al., 2017; Ni and McAuley, 2018) Review examples: I love this little stand! The coconut mocha chiller and caramel macchiato are delicious. Wow what a special find. One of the most unique and special date nights my husband and I have had. Tip examples: Great food. Nice ambiance. Gnocchi were very good. I can’t get enough of this place. Justification examples: The food portions were huge. Plain cheese quesadilla is very good and very cheap. Table 1: In contrast to reviews and tips, we seek to automatically generate recommendation justifications that are more concise, concrete, and helpful for deci- sion making. Examples of justifications from reviews, tips, and our annotated dataset are marked in bold. to generate explanations in the form of natural lan- guage, e.g. generating synthesized reviews simi- lar to those that users would write about a prod- uct. However, a large portion of review text (or text from ‘tips’) is often of little relevance to most users’ decision making (e.g. they describe verbose experiences or general endorsements) and may not be appropriate to use as explanations in terms of content and language style. As a result, existing models that …

Content truncated to 500 words