Abstract:

Previous work on social media text-based geolocation focused on mapping language semantic space to geospatial space, which ignores the semantic correlation between social media texts and the distance correlation between geographical locations.To take advantage of these correlations, mCLF, a new unsupervised multiple-level contrastive learning framework was proposed, three contrastive learning modules were designed: a semantic learning module, a location learning module, and a cross-learning module.Transformer encoder was used to obtain semantic representation of posts, utilizing unsupervised contrastive learning method to decrease the distance of semantic representations and location representations of posts with near locations, and then fine-tuned the model with supervised method for geographic location regression or classification outputs.Compared with five baseline methods, extensive experiments based on four datasets demonstrate the effectiveness of the proposed framework.

Key words: social media, geolocation, contrastive learning, information mining, Transformer