Automatic Domain-Adaptive Sentiment Analysis with SentiMap

Abstract

Location-based sentiment analysis is a promising field of study with various applications, but faces issues due to location uncertainty and lack of domain specificity. Our proposed solution automatically builds domain-adaptive lexicons for region-specific sentiment analysis. An initial lexicon for location estimation is created using topic modeling on news articles related to the target domain. For sentiment estimation, we start with a preexisting lexicon. Both initial lexicons are expanded recursively with a word embedding trained on social media messages from the target area. The final lexicons are used for location estimation and for automatically assigning sentiment labels to data, which is then used for fine-tuning a BERT transformer network. Our approach is validated in a case study of Amsterdam, demonstrating that both the automatically expanded lexicons and the fine-tuned network outperform their respective baselines. This illustrates how our system can enhance its performance by adapting to the domain, with minimal manual input. Finally, a temporal analysis is performed at different scales, showcasing the model’s ability to automatically detect sentimentally charged events.