科研之友
微信
新浪微博
Facebook
分享链接
摘要
With the development of the Internet, extractive question answering became an important component in the modern search engine systems. Given the question and the related passage, extractive reading comprehension models aim to find answer texts from the related passage to answer the question. Existing works only considered the setting where the answer to the question is comprised of a single text span from the passage, and they regarded the problem as taking the passage and the question as input and predicting two position indices in the passage to indicating the answer. However, in fact there are many questions whose answers are comprised of many text spans from the related passage, and the models needs to locate multiple spans from the passage to answer the question, rather than the single answer span. Existing works for reading comprehension models mainly focused on the design of the lower layer for text understanding and lacked of exploration on the multi-span answers. In this paper, we proposed BERT-Boundary model which uses the pre-trained BERT as lower layer. Specifically, BERT employs the self-attention mechanism and the fully-connected feed forward network to encode the related passage and the question, and it used the pre-trained parameters which are trained on the large-scale unsupervised corpus in advance. Our proposed model employs an answer-boundary labeling model for locating multiple answer spans. Specifically, it conducts sequence labeling for the start positions and the end positions of the answer spans over the related passage and a classification is conducted over each token of the passage. BERT-boundary combines the power of text understanding and multi-span locating. We conduct extensive experiments on the newly constructed multi-span answer dataset. The experimental results demonstrate the effectiveness of our model and a clear improvement on F1 measurement over existing models. First, we compare our language understanding layer of our model with RNN-based lower text understanding layer, and we find that our BERT-based model achieve better result. Second, we also compare the answer component of our model with content-based method which models whether a token is inside an answer span or not, and the result demonstrates that our boundary labeling strategy can locate answer spans more accurately than the content-based method. Finally, we compare our model with original BERT model based on the SQuAD data set, and we can find that our model can locate multiple answer span from the passage, meanwhile our model can still process single-span answer in usual reading comprehension data set. We further analyze the performance with the number of the answer span and the length of the answer span. As the increase of the number of the answer span, our model performance drop slowly while performance of previous BERT model drops quickly. As the increase of the length of the answer span, we can find that our model performs well on locating longer answer. However other baseline models cannot handle the long answer texts, especially for the content-based method, its performance drops as the length of the answer increases. We conclude that our model achieves consistent performance gain compared to existing baseline methods. Our code is publicly available at https://github.com/lixinsu/multi_span.
- 单位
