Statistical Semantic Change Detection via Usage Similarities

AbstractSemantic change detection comprises two subtasks: classification, which predicts whether a target word has undergone a semantic shift, and ranking, which orders words according to the degree of their semantic change. While most prior studies concentrated on ranking subtask, the classification subtask plays an equally important role, since many practical scenarios require a yes/no decision on semantic change rather than a global ranking. In this work, we propose a novel statistical method that predicts the presence or absence of semantic change. While most existing approaches infer semantic change by comparing word embeddings across time periods or domains, our method directly models the diachronic/synchronic consistency of usage-level similarity scores. Our experiments on SemEval-2020 Task 1 and WUGS datasets demonstrate that the proposed formulation outperforms existing state-of-the-art embedding-based methods, and robustly detects semantic change across languages in both diachronic and synchronic settings.