Towards More Natural Explanations of User Preferences

Abstract

Explainable recommendations refer to algorithms or methods that enable recommender systems to provide recommendations to the users, as well as to explain the reason why the items or products are being recommended. Recently, a concept of explainability in terms of user preferences is introduced. It provides a mechanism for recommender systems to explain their understanding of the user's preferences by generating user preference statements in the form of text.

In this thesis, we explore different approaches to making the user preference statements to sounding more natural through paraphrasing, while at the same time still preserving relevancy of the sentence, with correct grammar. Two main approaches are: (1) the template-based approach which includes enhancing the template with various sentence patterns and mining more colorful expressions from movie reviews; (2) employing neural language generation techniques by experimenting on state-of-the-art neural network models explicitly built for paraphrase generation, and on transfer learning method by fine-tuning pre-trained neural models. The objective of this work is to discover which of these approaches can be devised in generating paraphrases for user preference statements, that is sounding relevant, grammatically correct, and sounding natural.

We found that some methods or architectures did not work as expected during the experiment, but we also managed to develop a better alternative solution to one of the methods. The experiment results show that both approaches have potential, with their strength and challenges.