Retrieval-Augmented Controllable Review Generation

Proceedings of the 28th International Conference on Computational Linguistics, pages 2284–2295 Barcelona, Spain (Online), December 8-13, 2020 2284 Retrieval-Augmented Controllable Review Generation Jihyeok Kim Yonsei University zizi1532@yonsei.ac.kr Seungtaek Choi Yonsei University hist0613@yonsei.ac.kr Reinald Kim Amplayo University of Edinburgh reinald.kim@ed.ac.uk Seung-won Hwang ∗ Yonsei University seungwonh@yonsei.ac.kr Abstract In this paper, we study review generation given a set of attribute identifiers which are user ID, product ID and rating. This is a difficult subtask of natural language generation since models are limited to the given identifiers, without any specific descriptive information regarding the inputs, when generating the text. The capacity of these models is thus confined and dependent to how well the models can capture vector representations of attributes. We thus propose to additionally leverage references, which are selected from a large pool of texts labeled with one of the attributes, as textual information that enriches inductive biases of given attributes. With these references, we can now pose the problem as an instance of text-to-text generation, which makes the task easier since texts that are syntactically, semantically similar to the output text are provided as inputs. Using this framework, we address issues such as selecting references from a large candidate set without textual context and improving the model complexity for generation. Our experiments show that our models improve over previous approaches on both automatic and human evaluation metrics. 1 Introduction The ultimate goal of opinion mining and sentiment analysis (Pang and Lee, 2008) is to automatically digest opinions of users towards a certain product to accommodate decision making. While some of these opinions are explicitly articulated in product reviews that users write, most of them are unknown since users have not bought most of the products. Alternative solutions such as aspect-based sentiment analysis (Mukherjee and Liu, 2012; Pontiki et al., 2016) and recommendation systems (Resnick and Varian, 1997; Bobadilla et al., 2013) exist, however these only offer superficial outputs that are not as expressive as textual reviews. Thus, the task of automatically generating reviews given their attributes such as user and product, or review generation (Dong et al., 2017), is necessary to achieve this goal. Most of the previous approaches (Dong et al., 2017; Sharma et al., 2018) have framed review gener- ation as A2T (Attribute-to-Text problem), where the given input is a non-linguistic data (i.e., attribute identifiers for user, product, and rating) and the output is the review text. In this problem setup, the key challenge is to learn rich representations of the attributes, which are then used to produce the text us- ing either template-based surface realization methods (Kukich, 1983; McKeown, 1992) or neural-based decoders (Mei et al., 2016; Wiseman et al., 2017), as shown in the red box in Figure 1. However, it is difficult to learn these representations merely from the given attribute identifiers since they do not convey any semantics regarding the attributes. Our key contribution is AT2T (Attribute-matched-Text-to-Text), of augmenting inductive biases of at- tributes with their matching reference reviews, as illustrated as the blue box …

Content truncated to 500 words