Improving the Performance of Aspect-BasedSentiment Analysis with Pre-Trained LanguageModels and Perturbed Masking Method

Aspect-based sentiment analysis (ABSA) hasbecome an influential research field attracting much attentiondue to its many potential applications. It involves analyzingtexts from various perspectives and holds crucial importancein commercial and political intelligence. Improving ABSAtechniques is paramount, with various approaches, includinglexicon-based methods, machine learning, and hybrid models,being employed. Integration of deep learning and transformernetworks has shown promise in addressing this challenge. Ourstudy leverages pre-trained language models to develop aninnovative approach. Initially, we utilize the PerturbedMasking method to incorporate syntactic information into theanalysis, enabling the model to leverage both syntactic andsemantic cues. Subsequently, the induced trees, sentences, andaspects are inputted into the pre-trained BERT model.Through experiments on the SemEval2014 benchmark dataset,covering restaurant and laptop reviews, and the Twitterdataset, our model demonstrates robust performance.

Improving the Performance of Aspect-BasedSentiment Analysis with Pre-Trained LanguageModels and Perturbed Masking Method Keywords:

sentiment analysis , Deep Learning , Transformer , Perturbed Masking