LGLFF: A Lightweight Aspect-Level Sentiment Analysis Model Based on Local-Global Features

Abstract

Aspect-level sentiment analysis is highly dependent on local context. However, most models are overly concerned with global context and external semantic knowledge. This approach increases the training time of the models. We propose the LGLFF (Lightweight Global and Local Feature Fusion) model. Firstly, we introduce a Distilroberta pretrained model in the LGLFF to encode the global context. Secondly, we use the SRU++ (Simple Recurrent Unit) network to extract global features. Then we adjust the SRD (Semantic-Relative Distance) threshold size by different datasets, and use SRD to mask the global context to get the local context. Finally, we use the multi-head attention mechanism to learn the global and local context features. We do some experiments on three datasets: Twitter, Laptop, and Restaurant. The results show that our model performs better than other benchmark models.