Solving aspect-based sentiment analysis task with GNN models and tree reconstruction methods

Abstract

Sentiment Analysis (SA) is an important topic in NLP, which investigates the opinion polarity expressed in a sentence. To achieve more fine analysis, Aspect-Based Sentiment Analysis (ABSA) is put forward. Unlike traditional SA focusing on sentence-level analysis, ABSA is aimed at the analysis of the individual aspects in the sentences. In this way, ABSA can provide more detailed information about the opinion on the aspects of an entity. The application of ABSA contributes to the social investigation of certain products, policies, and people, and has been adopted by companies and the government to assist their decision process. To achieve the ABSA task, I choose to implement Graph Neural Networks (GNN), an emerging type of model in the machine learning area. Unlike other Artificial Neural Networks, GNN’s structure is built on the graph, a data structure that maps data to a non-Euclidean space, where the data structure is represented as vertices and edges. In this dissertation, I decided to implement three GNN models and compare their performance. The models are 1. Graph Convolution Network (GCN), a type of GNN model whose state transition function is based on spectral analysis. 2. Heterogeneous GCN, the GCN model applied on the heterogeneous graph, is a graph whose vertices and edges are classified into different types. 3. Graph Attention Network (GAT), an attention mechanism-based GNN model. The implementation procedures include 1. Building the graph structure with dependency parsing algorithm and Tree-Lifting and Tree-Truncating algorithm. 2. Applying three different models: GCN, GAT, and Heterogeneous GCN on the data. 3. Compare their results and explain. This dissertation will give the full picture of the development of solutions to ABSA tasks first. Then, the GNN models that inspire the algorithm developed in this dissertation will be introduced as well. After that, I will explain the algorithm and methodology that I utilized in this dissertation. Finally, the results of the implementation experiments will be analyzed. Among all the models, GAT had the best performance of 71.31% for SemEval 2014 Laptop task (3-way), while the result of GCN is less satisfying, reaching 68.48%. Hetero-GCN has the worst performance, resulting in 34.77% of accuracy.