I’m wondering if anyone knows a good model that could discern the difference between question + answer embeddings. For example a model that knows “Question: Do you like dogs? Answer: Yes” and “Question: Do you like dogs? Answer: No” are semantically different, and that the cosine similarity between the embddings of these two sequences would shoud that dissimilarity. Have tried lots of BERT and its variations and the cosine similarity of its emebeddings are very high.