Sentiment Analysis with Graph Convolutional Networks using directed PMI

Abstract

Sentiment analysis is a fast-growing area of research in natural language processing (NLP) and more specifically text classification which makes it possible to automatically detect the sentiment of data and is used in a wide range of applications. Many studies have applied deep learning techniques to text classification trying to mend the shortcomings of previous methods and improve their performance. In recent years several studies have applied the powerful graph convolutional networks (GCNs) to text classification problems, demonstrating promising results. However, most GCNs can’t directly work on directed graphs and make full use of the informative graph structure and these models’ capabilities. As order is extremely important in sentiment analysis tasks and most datasets in this field contain short texts, considering the order of words and creating a directed graph can add extra information to model and improve the training process. In this study we introduce DPMI (Directed PMI) as a measure to calculate word associations between word nodes considering the order of word occurrence and enhance one of the powerful existing GCN models named Text GCN, to accept directed graphs. The results show that adding order and using the proper aggregation function can improve the accuracy of predictions without adding complexity to the model.