SentiMap: Domain-Adaptive Geo-Spatial Sentiment Analysis

Location-based sentiment analysis has numerous applications, but suffers from both location uncertainty and lack of domain specificity. We propose an approach to automatically build domain-adaptive lexicons for region-specific sentiment analysis. For location estimation, we collect an initial lexicon using topic modeling on a collection of news articles about the target domain. For sentiment estimation we start with a preexisting lexicon. Both initial lexicons are then expanded recursively through employment of a word embedding trained on social media messages from the target area. The final location lexicon is used for location estimation, the final sentiment lexicon for automatically annotating data that is used to fine-tune a BERT transformer network on the task of sentiment estimation. We validate our approach by using the city of Amsterdam as a case study, and show that both the automatically expanded lexicons and the fine-tuned network outperform their respective baselines. This illustrates that with little manual input, our system improves through adapting to the domain.