Fine-Tuning for Cross-Domain Aspect-Based Sentiment Classification

Abstract

Aspect-Based Sentiment Classification (ABSC) is a subfield of sentiment analysis concerned with classifying sentiment attributed to pre-identified aspects. A problem in ABSC nowadays is the limited availability of labeled data for certain domains. This study aims to improve sentiment classification accuracy for these domains where labeled data is scarce. Our proposed approach is to apply cross-domain fine-tuning to a state-of-the-art deep learning method designed for ABSC: LCR-Rot-hop++. For this purpose, we initially train the model on a domain that has a lot of labeled data available and consecutively fine-tune the upper layers with training data of the target domain. The performance of the fine-tuning method is evaluated relative to a model that is trained from scratch for each target domain. For the initial training, restaurant review data is used. For the fine-tuning and from-scratch training we use review data for laptops, books, hotels, and electronics. Our results show that when comparing the fine-tuning with the from-scratch method (for the same training set), the fine-tuning method on average outperforms the from-scratch method when the training set is small for all considered domains and is considerably faster.