Combining Supervised and Unsupervised Learning to Detect and Semantically Aggregate Crisis-Related Twitter Content

Abstract

Twitter is an immediate and almost ubiquitous platform and therefore can be a valuable source of information during disasters. Current methods for identifying and classifying crisis-related content are often based on single tweets, i.e., already known information from the past is neglected. In this paper, the combination of tweet-wise pre-trained neural networks and unsupervised semantic clustering is proposed and investigated. The intention is to (1) enhance the generalization capability of pre-trained models, (2) to be able to handle massive amounts of stream data, (3) to reduce information overload by identifying potentially crisis-related content, and (4) to obtain a semantically aggregated data representation that allows for further automated, manual and visual analyses. Latent representations of each tweet based on pre-trained sentence embedding models are used for both, clustering and tweet classification. For a fast, robust and time-continuous processing, subsequent time periods are clustered individually according to a Chinese restaurant process. Clusters without any tweet classified as crisis-related are pruned. Data aggregation over time is ensured by merging semantically similar clusters. A comparison of our hybrid method to a similar clustering approach, as well as first quantitative and qualitative results from experiments with two different labeled data sets demonstrate the great potential for crisis-related Twitter stream analyses.