To make these new techniques accessible to researchers outside of Computational Linguistics, we have developed the DURel tool (Schlechtweg et al., 2024). The basic annotation data gathered in the system are judgments of semantic proximity between word uses (Blank, 1997; Erk et al., 2013), created using the DURel relatedness scale (Schlechtweg et al., 2018; Schlechtweg 2023, p. 33). DURel’s computational annotators enable us to generate word sense clusters for large sets of words and word uses, and to systematically search unlabelled data for new senses. The most important annotator is XL-Lexeme, a bi-encoder that vectorizes the input sequences using Fine-Tuned LLMs to Discover Non-Recorded Senses in Multiple Languages 33 XLMR-based Siamese Network (Cassotti et al., 2023), which has been trained to minimize the contrastive loss with cosine distance on several WiC datasets and predicts either cosine similarities or relatedness scores derived from these. The DURel tool then creates Word Usage Graphs (WUGs) out of these proximity judgments, where nodes represent word uses, and weights on edges represent the semantic relatedness of two nodes. Various graph clustering techniques can subsequently be used to identify word senses.