An Exploratory Case Study for Turkish Sentiment Classification Using Graph Convolutional Neural Networks

Abstract

Graph Convolutional Neural Networks (GCNs) are highly popular in recent years. It gives very successful results for various natural language processing (NLP) tasks such as sentiment classification. It has recently been shown to be effective and successful models to solve sentiment classification problem of texts. However, there is no research demonstrating the performance of this model on Turkish texts. In this study, we observe performance of the GCN model on the sentiment classification problem of Turkish texts as first research. Since the structure of Turkish language is agglutinative, different preprocessing approaches are presented and performance results on three real-world Turkish sentiment datasets are shown. It is observed that the TripAdv dataset, which was used in this study, yielded a 0.76 F-measure value. This can be considered a reasonable success for a sentiment classification with three sentiment classes. On the other hand, this study is presented as an exploratory case study in preparation for more detailed and extensive research in the future.