Nai, Peiran and Li, Lin and Tao, Xiaohui 
ORCID: https://orcid.org/0000-0002-0020-077X
(2020)
A Densely Connected Encoder Stack Approach for Multi-type Legal Machine Reading Comprehension.
In: 21st International Conference on Web Information Systems Engineering (WISE 2020), 20-24 Oct 2020, Amsterdam, The Netherlands.
Abstract
Legal machine reading comprehension (MRC) is becoming increasingly important as the number of legal documents rapidly grows. Currently, the main approach of MRC is the deep neural network based model which learns multi-level semantic information with different granularities layer by layer, and it converts the original data from shallow features into abstract features. Owing to excessive abstract semantic features learned by the model at the top of layers and the large loss of shallow features, the current approach still can be strengthened when applying to the legal field. In order to solve the problem, this paper proposes a Densely Connected Encoder Stack Approach for Multi-type Legal MRC. It can easily get multi-scale semantic features. A novel loss function named multi-type loss is designed to enhance the legal MRC performance. In addition, our approach includes a bidirectional recurrent convolutional layer to learn local features and assist in answering general questions. And several fully connected layers are used to keep position features and make predictions. Both extensive experiments and ablation studies in the biggest Chinese legal dataset demonstrate the effectiveness of our approach. Finally, our approach achieves 0.817 in terms of F1 in CJRC dataset and 83.4 in the SQuAD2.0 dev.
|
Statistics for this ePrint Item |
Actions (login required)
 |
Archive Repository Staff Only |