Towards the development of typhoon-related tweet classifiers despite the sparseness of strongly-annotated data

We employ a method for weakly annotating typhoon-related tweets in situations where there is only a small strongly-annotated dataset. The idea is that classifiers can be trained on a larger weakly-annotated dataset generated from this smaller dataset. In this research specifically, four binary classifiers were trained to detect the presence of the following themes in tweets: (1) Agonism or Engagement in Debate, (2) Celebrification, (3) Solidaristic, and (4) Tweeting about a Charity Event. The classifiers were then evaluated qualitatively on tweets about the original typhoon (Haiyan in 2013) from which the annotations came from, and on tweets about a typhoon years later (Haima in 2016). This was done by running LDA topic modelling on the positive classifications and qualitatively analyzing the resulting topics for relevance to the themes. Results suggest that the trained classifiers could be useful for the typhoon with which it was trained on, but not necessarily for other typhoons.