Comparing Human and BERT-Siamese Network (BertSNN) in Product Domain Similarity Ranking for Cross-Domain Sentiment Analysis

Cross-domain sentiment analysis (CDSA) aims to predict sentiment polarity of product reviews in a target domain by applying knowledge learned from source domain(s). Our study proposes BertSNN for measuring domain similarity and selecting training samples from the most relevant source domains to improve sentiment classification performance in the target domain. BertSNN encompasses a pre-trained BERT, a Siamese network, and a distance metric to generate similarity scores between domains. We first compare the source domain similarity ranking for a target domain between BertSNN and humans. Subsequently, cross-domain sentiment classification performance is used as a proxy measure to evaluate the effectiveness of the top-most similar source domains selected to train the sentiment classifier on a target domain. Findings utilizing 25 domains from Amazon product reviews reveal a high level of discrepancy between domain similarity ranking by humans and BertSNN but greater overlap of similar domains is observed when we focus on the top 5 most similar source domains to a target domain. Our study contributes through a more thorough empirical examination of identifying relevant source domains that can offer valuable training data for sentiment classification in target domains that lack specific in-domain training resources.