A method of improving the performance of extractive reading comprehension model by using adversarial samples
Extractive machine reading comprehension is a crucial task in natural language processing,where machines are required to extract the answer(fragments in inputting text)to a given question on the base of reading and understanding natural language text,and refuse to provide any answer when the given question is unanswerable.The task becomes more challenging when faced with unanswerable questions,where machines should refrain from providing any answer,especially when an inputting text contains plausible text fragments.Existing models are easily confused such fragments as an answer to the given question,and then wrongly judge answerability of question.To further improve the effect of extractive machine reading comprehension model,this paper takes plausible answers in SQuAD 2.0 dataset as the adversarial samples,which are used as positive examples extracted from text fragments of the answer and negative examples for judging answer-ability of question.Thus,this method increases the ranking loss based on the cross-entropy loss of the an-swers in the existing models.Experiments on SQuAD 2.0 shows that the proposed method can improve ro-bustness of existing extractive machine reading comprehension models,and significantly improve the effect of answerability judgment and extraction of text fragments about answer.
国家科技期刊平台
NSTL
万方数据
