Closed Milozms closed 5 years ago
When predict_raw_text is set to True, it will map the span prediction to the original text (using offsets) and extract the corresponding text for final evaluation (instead of concatenating all words using whitespace). This is a common practice for both SQuAD and CoQA and it usually works slightly better.
Oh, for the pipeline model, we need that predicted rationale is tokenized so it can be fed into the seq2seq model directly so we set predict_raw_text=False.
I saw in README that reading comprehension model sets predict_raw_text=True, while the pipeline model sets predict_raw_text=False. Are there any differences?