Context-Aware Stance Detection in Manipuri Editorials via FastText Embeddings and Keyword-Guided BiLSTM

Abstract

Although stance detection has gained substantial interest in high-resource languages, it remains largely unexplored for Manipuri, a low-resource Indo-Tibeto-Burman language with limited annotated data and linguistic resources. This paper proposes a framework for stance detection in Manipuri editorial articles. The dataset created for this research is balanced with respect to the stance categories, namely, FOR, AGAINST, and NEUTRAL. Both classical machine learning algorithms and Keyword-Guided Bidirectional Long Short-Term Memory (KG-BiLSTM) are evaluated on the dataset. The KG-BiLSTM architecture proposed in the paper consists of FastText embeddings along with a label-independent keyword-guided attention network constructed from manually curated stance lexicons. The attention mechanism enables the model to focus on potentially stance-bearing words without using ground-truth class information during mask construction. With five-fold stratified cross-validation, the KG-BiLSTM obtains an average accuracy of 71.2% and a Macro-F1 score of 70.1%, outperforming logistic regression and support vector machine baselines. The model reduces confusion between FOR and NEUTRAL editorials, which is one of the most critical problems for stance identification.