Product review summarization

Abstract

Online product reviews are an invaluable source of information for customers to make to informed decisions when they are making a purchase. However, users will have to go through large volumes of online reviews which makes the task overwhelming. Filtering similar or duplicated reviews further adds on the user’s effort. Thus, there is a need to identify the salient information among numerous reviews and present them in a summary to reduce user’s time and effort to make their decision. Multi-Document summarization techniques can be broadly classified into extractive and abstractive approaches. Extractive approaches are often selection-based techniques where sentences of the resulting summary generated are directly selected from the dataset reviews. On the other hand, abstractive approaches generate new sentences for the summary based on the training dataset. This project will be focusing on exploring an unsupervised, abstractive approach that uses an encoder-decoder framework in order to summarize the product reviews. The purpose is to explore the outcome of using such methodology and compare it with other unsupervised extractive methods for the product review summarization task.