Closed guotong1988 closed 4 years ago
Depends what you want to use it for. If you're using the embeddings directly to estimate semantic textual similarity, then Universal Sentence Encoder (lite) performs slightly better overall while InferSent does a bit better at distinguishing between a sentence and its negation. InferSent is also way better at representing analogical properties of single-word inputs, so any application that compares the embeddings for sentences to the embeddings for words is likely to do better with InferSent. On the other hand, USE lite does better at detecting important differences in word order.
If you're using the embeddings as pre-trained input features for downstream tasks, I'm afraid I have no insight to offer. I generally use the embeddings directly, and don't have a good feel for their effectiveness at transfer learning.
Thank you
https://www.aclweb.org/anthology/D18-2029
Thank you very much!