基于多任务学习的生成式阅读理解(Generative Reading Comprehension via Multi-task Learning)

Jin Qian (钱锦), Rongtao Huang (黄荣涛), Bowei Zou (邹博伟), Yu Hong (洪宇)


Abstract
生成式阅读理解是机器阅读理解领域一项新颖且极具挑战性的研究。与主流的抽取式阅读理解相比,生成式阅读理解模型不再局限于从段落中抽取答案,而是能结合问题和段落生成自然和完整的表述作为答案。然而,现有的生成式阅读理解模型缺乏对答案在段落中的边界信息以及对问题类型信息的理解。为解决上述问题,本文提出一种基于多任务学习的生成式阅读理解模型。该模型在训练阶段将答案生成任务作为主任务,答案抽取和问题分类任务作为辅助任务进行多任务学习,同时学习和优化模型编码层参数;在测试阶段加载模型编码层进行解码生成答案。实验结果表明,答案抽取模型和问题分类模型能够有效提升生成式阅读理解模型的性能。
Anthology ID:
2020.ccl-1.29
Volume:
Proceedings of the 19th Chinese National Conference on Computational Linguistics
Month:
October
Year:
2020
Address:
Haikou, China
Editors:
Maosong Sun (孙茂松), Sujian Li (李素建), Yue Zhang (张岳), Yang Liu (刘洋)
Venue:
CCL
SIG:
Publisher:
Chinese Information Processing Society of China
Note:
Pages:
301–312
Language:
Chinese
URL:
https://aclanthology.org/2020.ccl-1.29
DOI:
Bibkey:
Cite (ACL):
Jin Qian, Rongtao Huang, Bowei Zou, and Yu Hong. 2020. 基于多任务学习的生成式阅读理解(Generative Reading Comprehension via Multi-task Learning). In Proceedings of the 19th Chinese National Conference on Computational Linguistics, pages 301–312, Haikou, China. Chinese Information Processing Society of China.
Cite (Informal):
基于多任务学习的生成式阅读理解(Generative Reading Comprehension via Multi-task Learning) (Qian et al., CCL 2020)
Copy Citation:
PDF:
https://aclanthology.org/2020.ccl-1.29.pdf
Data
CoQAHotpotQAMS MARCONarrativeQASearchQATriviaQA