This model can measure semantic similarity between pairs of texts containing figurative language. As far as we know,
this model works slightly better than sup-simCSE-roberta-base. For example :
sentence 1: I have been in seventh heaven since Harry entered my life .
sentence 2: I have been in very happy since Harry entered my life.
the cosin score of simcse: 0.897
the cosin score of us: 0.897
sentence 1: I have been in seventh heaven since Harry entered my life .
sentence 2: I have been in pain since Harry entered my life .
the cosin score of simcse: 0.846
the cosin score of us: 0.753
It's still a big challenge for us to measure semantic similarity of figurative language from the sentence embedding perspective.
unsupvised models may useless as the key is to infer the literal meaning of the figurative expression, since the annotated is rare.
