Dynamic domain information modulation algorithm for multi-domain sentiment analysis

Abstract

Multidomain sentiment classification aims to improve model performance constrained by limited labeled data in a single domain by utilizing labeled data from multiple domains. Models that simultaneously train domain classifiers and sentiment classifiers have shown benefits. In this framework, domain classification serves as an auxiliary task, supplying crucial information for sentiment analysis. It is generally assumed that the importance of sentiment classification tasks remains consistent across all domains. By contrast, domain classification tasks exhibit variability because the impact of domain information on sentiment analysis differs among fields. This variability can be managed through adjustable weights or hyperparameters. However, as the number of domains grows, existing hyperparameter optimization algorithms face several challenges, including (1) high computational requirements, (2) convergence difficulties, and (3) increased algorithmic complexity. To efficiently generate the domain-specific information required for sentiment classification, we propose a dynamic information modulation algorithm. Specifically, the training process is divided into two phases. In the first phase, a global modulation factor that controls the proportion of domain classification tasks across all domains is established. In the second phase, we introduce an innovative cross-domain balancing modulation algorithm to refine the domain information embedded in the input text. This refinement is achieved using a gradient- and loss-based method. Experimental results show that our approach consistently enhances performance across most domains, achieving improvements of 0.3–1.0 % on 10 of 16 Amazon domains and 0.5–1.5 % on 3 of 5 Yelp domains, while maintaining performance comparable to baseline models in other domains.