Net2text: An Edge Labelling Language Model For Personalized Review Generation

Writing an item review for online shopping or sharing the dining experience of a restaurant has become major Internet activities of young people. This kind of review system could not only help users express and exchange experience but also prompt business to improve service quality. Instead of taking time to type in the review, we would like to make the review process more automated. In this work, we study an edge labelling language model for personalized review generation, e.g., the problem of generating text (e.g., a review) on the edges of the network (e.g., online shopping). It is related to both network structure and rich text semantic information. Previously, link prediction models have been applied to recommender system and event prediction. However, they could not migrate to text generation on the edges of networks since most of them are numerical prediction or tag labelling tasks. To bridge the gap between link prediction and natural language generation, in this paper, we propose a model called Net2Text, which can simultaneously learn the structural information in the network and build a language model over text on the edges. The performance of Net2Text is demonstrated in our experiments, showing that our model performs better than other baselines, and is able to produce reasonable reviews between users and items.