Speaker
Description
The Word Sketch is a core tool for modern corpus-based lexicography, providing a concise summary of a word's grammatical and collocational behavior. A longstanding limitation of standard Word Sketches is the lack of polysemy awareness: when a lexicographer examines a highly polysemous word such as bank, the resulting collocate lists mix financial terms like account and deposit with geographical terms like river and sand. In this paper, we present and evaluate a fully automated method for enriching Word Sketches with word sense information. Using the Adaptive Skip-gram model, we perform unsupervised word sense induction and project the induced senses onto the type-level collocate lists of a Word Sketch, so that collocates with sufficiently confident predictions are associated with the sense they most strongly represent. On a manually annotated test set of 23 polysemous English words, our method achieves a Precision of 0.64 and Recall of 0.77 (F1 ≈ 0.70). The method requires no manual training data and is already deployed in Sketch Engine for eight languages. We describe the user interface, which allows lexicographers to filter Word Sketches by sense, and discuss implications for dictionary-writing workflows.