simplify the prediction process for a finetuned bert model

text_similarity_predict.py shows the demo