A Study Of Aspect Embedding And Conflict Sentiment Recognition On Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis is a fine-grained Sentiment Analysis task,which aims to analyze sentiment polarity for each aspect that is commented.In this paper,we discuss two problems which we think are restricting the development of Aspect-Based Sentiment Analysis.The first problem is the misalignment problem of aspect-category and aspect-term in representation space.This problem is about aspect embedding which is an important component that provides semantic information about the aspects for deep learning models on Aspect-Based Sentiment Analysis.However,there is no existing research which discusses the difference between different aspect embedding methods.Questions,such as “Which type of aspect embedding is more suitable for sentiment analysis?” remain open.We hope our research can provide some insights into these questions that are related to aspect embeddings.The second problem is about conflict sentiment,which is an interesting but frequently neglected phenomenon.We argue that recognizing conflict sentiment is meaningful and discuss its necessity in this paper.However,existing models on Aspect-Based Sentiment Analysis have difficulty in recognizing conflict sentiment.Thus,we need to design a novel model that is able to recognize conflict sentiment more accurately.In order to tackle the above two problems,we propose Aligning Aspect Embedding method and Dual-Attention-GRU(D-AT-GRU)model.Aligning Aspect Embedding first represents the relation between aspect-categories and aspect-terms in data through a graph.Then it learns aspect embeddings through an unsupervised representation learning algorithm.D-AT-GRU turns multi-class classification problem into multi-label classification problem.Then it utilizes two different attentions to recognize conflict sentiment.We test the proposed methods on Sem Eval Restaurant and Laptop datasets.The experiment results show that Aligning Aspect Embedding can represent the relation between aspect-categories and aspect-terms more accurately and can improve the performance of sentiment analysis on deep learning models.D-AT-GRU outperforms all compared method in terms of sentiment analysis accuracy,especially on conflict sentiment.