Cross-domain corpus selection for cold-start context

Abstract

Sentiment analysis is a powerful tool for monitoring attitudes towards companies, products or services and identifying specific features that drive positive or negative sentiment. However, collecting labelled data for training sentiment analysis models in a specific domain can be challenging in practical applications. One promising solution to this ‘cold-start’ problem is domain adaptation, which leverages labelled data from a related source domain to train a model for the target domain. A critical yet often neglected aspect in prior research is the measurement of similarity between the source and target domains, a factor that greatly impacts the success of domain adaptation. To fill this gap, we propose a novel measure that combines semantic, syntactic and lexical features to assess corpus-level similarity between two domains. Our experimental results demonstrate that our method achieves high precision (0.91) and recall (0.75), outperforming traditional methods. Moreover, our proposed measure can assist new domain products in selecting the most suitable training data set for their sentiment analysis tasks.