Deep-change at AXOLOTL-24: Orchestrating WSD and WSI Models for Semantic Change Modeling

TL;DR

The paper tackles AXOLOTL-24 subtask 1, automating usage-level semantic change labeling by integrating WSD (GlossReader variants), WSI (Agglomerative) and SCM models (AggloM, Cluster2Sense, Outlier2Cluster). It introduces a Novel Sense Detection (NSD) component and demonstrates that combining NSD with per-usage decisions yields state-of-the-art results across Finnish, Russian, and German datasets, especially improving ARI for gained-sense discovery. The three SCM methods each extend WSD/WSI in complementary ways, with Outlier2Cluster achieving strong performance by per-usage NSD-based relabeling and offering gained-sense discovery. The findings highlight NSD quality as a key driver for SCM performance and point to promising directions for robust, cross-language semantic-change modeling.

Abstract

This paper describes our solution of the first subtask from the AXOLOTL-24 shared task on Semantic Change Modeling. The goal of this subtask is to distribute a given set of usages of a polysemous word from a newer time period between senses of this word from an older time period and clusters representing gained senses of this word. We propose and experiment with three new methods solving this task. Our methods achieve SOTA results according to both official metrics of the first substask. Additionally, we develop a model that can tell if a given word usage is not described by any of the provided sense definitions. This model serves as a component in one of our methods, but can potentially be useful on its own.

Figures (5)

• Figure 1: An example of data for the first subtask of AXOLOTL-24.
• Figure 2: Outlier2Cluster pipeline. Inputs are in green and outputs are in blue. Triangles denote ML models.
• Figure 3: Proportions of target words falling into different categories in the shared task datasets.
• Figure 4: ARI and F1 on the development sets depending on the threshold of novel sense detector. Higher threshold means higher proportion of WSD predictions and less WSI predictions.
• Figure 5: Precision-recall curves of novel sense detection models. Non classifier models are distances between usages and chosen glosses from GlossReader FiEnRu. Classifier w/ extra stands for classifier trained on distance-based and non distance-based features introduced in sub subsection \ref{['sec: outlier2cluster']}. Classifier w/o extra stands for classifier trained only on distance-based features.