Attention-based Multi-Channel Gated Recurrent Neural Networks: A Novel Feature-centric Approach for Aspect-based Sentiment Classification
Date
2023Item Type
ArticleAbstract
Sentiment analysis is an active research domain of the current era, thanks to its vast applications. In the meantime, its main objective is to classify the polarity of the context as positive, negative, or neutral. Thus, researchers’ focus shifted towards the aspect or feature-based sentiment analysis because overall polarity does not determine the people’s views towards certain features. Therefore, Aspect-Based Sentiment Analysis (ABSA) helps us to identify the sentiments about various aspects of different products and services. However, their accurate identification and extraction are still challenging for the research community due to the disambiguation of natural languages. This paper presents a method named Attention-based Multi-Channel Gated Recurrent Neural Network (Att-MC-GRU), which extracts aspects and classifies their sentiments from textual reviews. It used the hybrid approach by combining word embedding, contextual position information, and part of speech (POS) tags. The main novelty lies in proposal of a Multi-Channel Gated Recurrent Neural Network (MC-GRU), in contrast to the existing studies that consider Recurrent Neural Networks (RNN) comprising only a single input channel. In addition, word embedding, POS tags, and contextual position information collectively enhance the accuracy of aspects identification and prediction of their associated sentiments. Due to the application of the filtering by the attention mechanism that figured out first the significant words, which helps to identify entities and their aspects related to the sentiment expressed. The empirical analysis proves the effectiveness of the proposed approach compared to the existing techniques in the relevant literature using standard datasets. The experimental results represent that the proposed model outperforms in the F1-measure with an overall achievement of 94% in the aspect term extraction task and 93% in the sentiment classification task.
Author
Ahmad, Waqas
Khan, Hikmat Ullah
Iqbal, Tassawar
Iqbal, Saqib