Ensemble BiLSTM: A Novel Approach for Aspect Extraction From Online Text

Abstract

Aspect extraction poses a significant challenge in Natural Language Processing (NLP). Extracting explicit and implicit aspects from online text data remains an ongoing challenge despite significant research efforts. Enhancing the accuracy and effectiveness of aspect extraction is an important area for improvement. This research introduces Ensemble BiLSTM, a novel approach to aspect extraction that addresses these challenges. Ensemble BiLSTM leverages the syntactic, semantic, and contextual properties of unstructured texts present in BERT word embeddings, along with their sequential properties captured using an ensemble of Bidirectional Long Short-Term Memory (BiLSTM) models. The proposed Ensemble BiLSTM model was evaluated extensively using the SemEval-2014 Restaurant, SemEval-2015 Restaurant, SemEval-2016 Laptop, and Financial Opinion Mining and Question Answering (FiQA) datasets. The experimental results demonstrate its efficacy in extracting aspects from text, achieving 91.28%, 87.39%, 95.85%, and 94.59% accuracy on the respective datasets. These promising results highlight the effectiveness of the ensemble approach and the incorporation of sequential models combined with BERT embeddings. The contributions of this research lie in the aspect category features extracted by the proposed Ensemble BiLSTM model, which can be expanded upon to generate accurate aspect-level sentiment features in future work.

Keywords

• Aspect extraction
• aspect-oriented features
• ensemble BiLSTM
• BERT embeddings