STIF: Semi-Supervised Taxonomy Induction using Term Embeddings and Clustering

In this paper, we developed a semi-supervised taxonomy induction framework using term embedding and clustering methods for a blog corpus comprising 145,000 posts from 650 Ukraine-related blog domains dated between 2010-2020. We extracted 32,429 noun phrases (NPs) and proceeded to split these NPs into a pair of categories: General/Ambiguous phrases, which might appear under any topic vs. Topical/Non-Ambiguous phrases, which pertain to a topic’s specifics. We used term representation and clustering methods to partition the topical/non-ambiguous phrases into 90 groups using the Silhouette method. Next, a team of 10 communications scientists analyzed the NP clusters and inducted a two-level taxonomy alongside its codebook. Upon achieving intercoder reliability of 94%, coders proceeded to map all topical/non-ambiguous phrases into a gold-standard taxonomy. We evaluated a range of term representation and clustering methods using extrinsic and intrinsic measures. We determined that GloVe embeddings with K-Means achieved the highest performance (i.e. 74% purity) for this real-world dataset.