MAPA BiLSTM-BERT: multi-aspects position aware attention for aspect level sentiment analysis

Mayur Wankhade,Chandra Sekhara Rao Annavarapu,Ajith Abraham
DOI: https://doi.org/10.1007/s11227-023-05112-7
IF: 3.3
2023-02-28
The Journal of Supercomputing
Abstract:Sentiment categorization at the aspect level tries to provide fine-grained sentiment polarities for distinct aspects inside a sentence. Some issues remain unresolved in the previous work. First, the specific position context is not fully addressed. Second, the distinct aspect of an opinionated sentence is evaluated independently. Also, the present, attentive approaches neglect the disruption caused by all the aspects in the same sentence while measuring the current aspect attention vector. We proposed multi-aspect-specific position attention bidirectional long short-term memory (MAPA BiLSTM)-bidirectional encoder representations from transformers (BERT) model to address these issues. The MAPA BiLSTM BERT introduces the explicit multiple aspect position-aware attention between the aspect word and the closest context words, also BERT aspect-specific attention investigates how to model multiple aspects using the aspect position attention mechanism. The parallel fused MAPA BiLSTM-BERT gains the multiple aspect contextual knowledge in a sentence for aspect classification. We conduct an empirical evaluation of the proposed method on the laptop review, restaurant review (SemEval2014 datasets), Twitter review, and multi aspect multi-sentiment (MAMS) dataset; results indicate a significant performance improvement over state-of-the-art methods.
computer science, theory & methods,engineering, electrical & electronic, hardware & architecture
What problem does this paper attempt to address?