基于课程学习的无监督常识问答模型

Unsupervised commonsense question-answering model based on curriculum learning

无监督常识问答是利用机器自动生成问答数据来对模型进行训练的问答模型,目前方法生成的问答数据中存在噪声数据和问题的难度随机的问题。提出一种基于课程学习的无监督常识问答模型,首先根据知识生成问答数据集,再对问答数据集进行多样化评估和流畅性评估,结合两个评估结果进行数据过滤,去除噪声数据;最后根据课程学习策略,使用干扰项与正确答案的相似度作为问题难度评估标准,使得模型根据难度等级来进行训练。在测试任务上具有1.5%~3.5%的准确率提升,证明了该模型在无监督常识问答任务上的有效性。

Unsupervised commonsense question answering is a question answering model that uses the machine to automatically generate question-answering data.There are some problems in the question-answering data generated by current methods,such as noise data and random difficulty of questions.This paper proposed an unsupervised commonsense question-answering model based on curriculum learning.Firstly,it generated a question-answering dataset according to knowledge,then evaluated the diversity and fluency of the question answering dataset,and filtered the data by combining the two evaluation results to remove noise data.Finally,according to the course learning strategy,it used the similarity between the interference item and the correct answer as the difficulty evaluation standard to train the model according to the difficulty level.The accuracy of the test tasks is improved by 1.5%~3.5%,which proves that the model is effective in unsupervised commonsense question-answering tasks.