A Reinforcement Learning Method with Attention Mechanism for Generating English Abstractive Summary of Products - Using Amazon E-Commerce as Examples

Abstract

Today shopping online has become a daily practice for many people A customer being interested in something on a web shop and not able to examine and try the real product might turn to the custom reviews to see the opinions of buyers on the product before making the final decision But the reviews may be too colloquial or the summary may be too short to mention important key characteristics and the detail specs of the product This makes the customer informed only that the product is good or bad but not that important key characteristics of the product are good or bad Such reviews of a web shop may not actually be helpful for its potential customers This research mainly aims at Amazon review and summary information and uses deep learning approach to analyze and learn reviews and summaries of different classes of products and to produce summaries with import key characteristics of products for customers The approach first uses part-of-speech (POS) tagging syntactic dependence and noun phrases to find the keywords in the reviews and summaries and then uses reinforcement learning to learn the keywords and their review contents in order to understand the semantic of sentences in the customer reviews Given a product review the network produces a summary that is easy to read and contains information of important key characteristics of the product and is helpful for consumers to quickly understand the quality of the product and make the final decision The major research works of this study are as follows: 1 Design grammar and syntactic dependence rules of sentence and use them to extract keywords from different types of sentences in the reviews Those rules can be expanded in the future when there are new requirements 2 Replace the Attention mechanism of a Pointer-Generator with the Intra Attention mechanism to make the decoder reconsider the temporal Attention scores of previously generated sequences when generating summary vocabulary This makes the model avoid focusing on a same word when the model is in generation mode 3 Add extra keyword semantic information in the original Attention mechanism which makes the generated attention weights focus more on the key words than the original Attention mechanism 4 Apply the self-critical-sequence training method to optimize the Pointer-Generator network This study conducts totally thirteen experiments The first experiment focuses on the accuracy of keyword extractions The second to the fourth experiments focus on the analysis of the vocabulary distribution of different Attention mechanisms in the Pointer-Generator network The fifth to the eleventh experiments are to compare the accuracy with some extractive and abstractive summary methods proposed in recent years The twelfth to the thirteenth experiments by using the best model in this study generate summary for different types of product reviews and evaluate the accuracy of the model by three methods namely Rouge BLEU and METEOR

Date of Award	2020
Original language	English
Supervisor	Tzone-I Wang (Supervisor)