Is averaging sentence embeddings, the right way to get representation for documents. Say I have a list of sentence embeddings representing symptoms. A data point looks like these: x|S1,S2,S3 --> Y|D1,D3(One hot representation). Here S1,S2 represents symptom definitions. Labels are disease(One hot). So we use Googles Universal sentence encoder to get sentence embeddings and average S1,S2,S3 and use it for training. Is there a better way other then averaging sentence embeddings?