Ind-MSA: Bridging the Linguistic Gap for Inductive Multilingual Sentiment Analysis with Graph Neural Networks

Abstract

Graph neural networks (GNNs) are a powerful class of models that learn effective representations over dependencies between entities in a graph. Despite the achievement of GNNs in capturing both long-and short-distance semantics, challenges persist within text classification problems, such as multilingual sentiment analysis. Most existing GNNs methods are unable to capture word ordering and do not effectively support inductive learning with new data. In this work, to handle these challenges, we propose Ind-MSA, an Inductive Multilingual Sentiment Analysis approach. Our initial step involves constructing a single heterogeneous text graph, leveraging diverse information to effectively model the multilingual corpus. Subsequently, the learned word representations are employed to train a Bi-LSTM with attention mechanism, enhancing the proposed approach by incorporating sequential information into the analysis. The proposed approach can capture both short-and long-distance semantics while supporting the word ordering, which is extremely critical in sentiment analysis. Comprehensive experiments on various distinct datasets reveal that Ind-MSA leads to state-of-the-art results, significantly outperforming methods which focus on local consecutive word sequences and global word co-occurrence.