Soft cosine and extended cosine adaptation for pre-trained language model semantic vector analysis

Funebi Francis Ijebu, Yuanchao Liu, Chengjie Sun, Patience Usoro Usip

Abstract: Highlights•Soft cosine and extended cosine similarities are adapted for semantic vector comparison.•Novel multi-lateral evaluation of alternative similarity measures under same conditions.•Cosine similarity measure is significantly outperformed in multiple test cases of the study.•Non-conventional similarity measure produced state-of-the-art result in STS-14 dataset.•Results of adapted similarity measures improve previously reported STS model performances.