Applying Topic Attention and Variational Autoencoder in Personalized Summarization of Reviews

Abstract

With the appearance of the Web 2 0 the electronic word of mouth (eWOM) has developed rapidly In order to have increased insight regarding product quality more and more customers watch online reviews before buying However the growing number of online reviews also presents an information overload problem There are too many reviews for customers to consume Nowadays many companies utilize a recommendation system to help customers make a decision Although the system may help customers save their time it is usually based on the price stars or browsing history and does not consider the online reviews Text summarization is a good way to solve the information overload problem Along with the progress of deep learning text summarization has become an important research area Contrasting other research our study considers the customer’s interests in our summary task We propose a new text summarization method which includes personalization called TA-VAE First we use the topic model to detect customers’ interests from previously written online reviews In the topic model we select the company's category and the POS tags to find niche interests Next we use a popular generation model variational autoencoder (VAE) to produce the summary Last we integrate the customers’ interests into the model by using the Attention mechanism and generate the personalized summary of reviews Based on the results of our research considering the company's category can benefit the topic model to finding more specific customer interests Plus the TA-VAE method not only improves the quality of the summary but also generates the review's summary included customer's features

Date of Award	2020
Original language	English
Supervisor	Hei-Chia Wang (Supervisor)