Abstract
Existing phrase embeddings provide low dimensional representations for phrases or sentences. However, these phrase embedding techniques treat phrases as independent units and are unable to capture contextual and sentiment dependencies among them. As a result, phrases with similar contexts and sentiments could have unrelated embedding representations. In order to overcome this research gap, we have proposed a method to infer context-aware causal phrase embeddings using a multi-relational graph neural network. Initially, a sequence-labelling deep model is used to extract rationale (causal phrases) for user opinions and a multi-relational causal graph is constructed, which is input to a graph neural network to infer context-aware causal phrase embeddings. The context enriched multi-relational embeddings are then fed to a neural network classifier to predict sentiment polarities. We have evaluated our models on three publicly available datasets and obtained mean accuracies of 96%, 95%, 80% and 95%, 95%, 80% respectively for causality extraction and sentiment inference tasks. Empirical evaluation reveals that the models proposed could outperform strong baselines in the literature with respect to causal extraction and sentiment inference tasks.
