Learning to Generate Product Reviews from Attributes

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 623–632, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Learning to Generate Product Reviews from Attributes Li Dong†, Shaohan Huang‡, Furu Wei‡, Mirella Lapata†, Ming Zhou‡ and Ke Xu⊺ †University of Edinburgh, Edinburgh, United Kingdom ‡Microsoft Research, Beijing, China ⊺Beihang University, Beijing, China li.dong@ed.ac.uk, {fuwei, mingzhou}@microsoft.com, buaahsh@gmail.com, mlap@inf.ed.ac.uk, kexu@nlsde.buaa.edu.cn Abstract Automatically generating product reviews is a meaningful, yet not well-studied task in sentiment analysis. Traditional natu- ral language generation methods rely ex- tensively on hand-crafted rules and prede- fined templates. attention-enhancedThis paper presents an attribute-to-sequence model to generate product reviews for given attribute information, such as user, product, and rating. The attribute en- coder learns to represent input attributes as vectors. Then, the sequence decoder generates reviews by conditioning its out- put on these vectors. We also introduce an attention mechanism to jointly gener- ate reviews and align words with input at- tributes. The proposed model is trained end-to-end to maximize the likelihood of target product reviews given the attributes. We build a publicly available dataset for the review generation task by leveraging the Amazon book reviews and their meta- data. Experiments on the dataset show that our approach outperforms baseline meth- ods and the attention mechanism signif- icantly improves the performance of our model. 1 Introduction Nowadays, there are many popular online review sites (such as Amazon, and Yelp) that allow users to read and post reviews about books, electronics, restaurants, etc. The reviews are used to express opinions for different aspects of products, and have a wide variety of writing styles and differ- ent polarity strengths. As a result, much previous work has focused on how opinions are expressed in review data. For example, previous studies on User Product Rating Attribute Encoder Sequence Decoder LSTM I loved this family story , it was touching . Attention Layer Figure 1: Our model learns to encode attributes into vectors, and then uses recurrent neural net- works based on long short-term memory (LSTM) units to generate reviews by conditioning on the encoding vectors. An attention layer is used to learn soft alignments between attributes and gen- erated words. sentiment analysis identify and extract subjective content in review data (Liu, 2015). However, few studies have explored building data-driven mod- els that can generate product reviews for the given products and ratings, which is helpful to under- stand how a specific user comments for products. As shown in Figure 1, the input to our model is a set of attributes (such as user, product, and rat- ing information), and our goal is to generate user- and product-specific reviews that agree with the input rating. These automatically generated re- views are useful for companies. For example, we could promote a product to users who have not bought it, by generating novel and personalized recommendations. We could also build a review writing assistant for E-commerce websites. Af- ter the website generates some candidate …

Content truncated to 500 words