.png){kind=icon}
返回
A Token-based transition-aware joint framework for multi-span question answering
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ipm.2024.103678.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Multi-span question answering has gained prominence as it aligns more closely with real -world user requirements compared to single-span question answering. The utilization of pretrained language models has shown promise in improving multi-span question answering, particularly for factoid questions that necessitate entity-based answers. However, existing methods tend to overlook critical information regarding answer span boundaries, resulting in limited accuracy when generating descriptive answers. To address this limitation, we propose TOAST, a novel joint learning framework specialized in token-based neighboring transitions that capture answer span boundaries through adjacent word relations. Our approach extracts high-quality multispan answers and is general-purpose, applicable to both alphabet languages like English and logographic languages like Chinese. Furthermore, we introduce CLEAN, a comprehensive opendomain Chinese multi-span question answering dataset, which includes a substantial number of descriptive questions. Extensive experiments demonstrate the superior performance of TOAST over previous top-performing QA models in terms of both EM F1 and overlapped F1 scores. Specifically, the TOAST models, leveraging BERT,se and RoBERTa,,se, achieve substantial improvements in EM F1 scores, with increments of 3.03/2.13, 4.82/3.73, and 16.26/11.53, across three publicly available datasets, respectively.
Keyword:
Reading comprehension
Multi-span question answering
Pretrained language models
Multitask learning
Chinese datasets
期刊
I
IF:
6.9
论文数:
5.2K
被引数:
1.4W
机构
引用论文
.png){kind=icon}
TYDI QA: A Benchmark for Information-Seeking Question Answering in Typologically Diverse Languages
OAAI
OAAI分享收藏
查看更多内容