Attention Based BiGRU-2DCNN with Hunger Game Search Technique for Low-Resource Document-Level Sentiment Classification

Abstract

Extracting features with deep learning models recorded considerable performance in classifying various sentiments. However, modeling lengthy sentences from low-resource document-level datasets, and examining the semantic relationships between these sentences is challenging. Also, recent document representation modules deployed rarely distinguish between the significance of various sentiment features from general features. To solve these challenges, we exploit a neural network-based feature extraction technique. The technique exploits an attention-based two-layer bidirectional gated recurrent unit with a two-dimension convolutional neural network (AttnBiGRU-2DCNN). Specifically, the Bi-GRU layers extract compositional semantics of the low resource document. These layers generate the semantic feature vector of the sentence and present the documents in a matrix form with a time-step dimension and a feature dimension. Due to the high dimensional multiple features generated, we propose a Hunger Games Search Algorithm to select essential features and subdue unnecessary features to increase classification accuracy. Extensive experiments on two low-resourced datasets indicate that the proposed method captures low-resource sentimental relations and also outperformed known state-of-the-art approaches.