HSE at TempoWiC: Detecting Meaning Shift in Social Media with Diachronic Language Models

Abstract

This paper describes our methods for temporal meaning shift detection, implemented during the TempoWiC shared task. We present two systems: with and without time span data usage. Our approaches are based on the language models fine-tuned for Twitter domain. Both systems outperformed all the competition’s baselines except TimeLMs-SIM. Our best submission achieved the macro-F1 score of 70.09% and took the 7th place. This result was achieved by using diachronic language models from the TimeLMs project.