Improving Aspect Sentiment Classification via Retrieving from Training Data

Aspect sentiment classification (ASC) is an essential subtask of aspect-based sentiment analysis. Recently, pre-trained language models (PLMs) have gradually become the mainstream building block for the ASC task and retrieval-based methods are shown to be effective in helping PLMs better understand various downstream tasks. However, retrieval-based methods need to introduce external knowledge, and building an index of largescale corpus leads to huge cost. To this end, we propose an effective method that utilizes training data to construct a retrieval corpus and retrieve instances most similar to current input to enhance semantic representations. In the proposed method, three kinds of retrieval metrics are applied, allowing us to quickly adapt to ASC tasks. Experimental results show that this simple retrieval augmented method can achieve significantly better performance on three benchmark ASC datasets (Twitter, Laptop, and Restaurant). Our code is available at https://github.com/rickltt/ re- bert.