HAS-QA: Hierarchical Answer Spans Model
for Open-domain Question Answering
Liang Pang
†
, Yanyan Lan
†∗
, Jiafeng Guo
†
, Jun Xu
†
, Lixin Su
†
and Xueqi Cheng
†
†CAS Key Laboratory of Network Data Science and Technology,
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
†University of Chinese Academy of Sciences, Beijing, China
∗Department of Statistics, University of California, Berkeley
{pangliang,lanyanyan,guojiafeng,sulixin,cxq}@ict.ac.cn, junxu@ruc.edu.cn
Abstract
This paper is concerned with open-domain question answer-
ing (i.e., OpenQA). Recently, some works have viewed this
problem as a reading comprehension (RC) task, and directly
applied successful RC models to it. However, the perform-
ances of such models are not so good as that in the RC task. In
our opinion, the perspective of RC ignores three characterist-
ics in OpenQA task: 1) many paragraphs without the answer
span are included in the data collection; 2) multiple answer
spans may exist within one given paragraph; 3) the end pos-
ition of an answer span is dependent with the start position.
In this paper, we first propose a new probabilistic formula-
tion of OpenQA, based on a three-level hierarchical structure,
i.e., the question level, the paragraph level and the answer
span level. Then a Hierarchical Answer Spans Model (HAS-
QA) is designed to capture each probability. HAS-QA has the
ability to tackle the above three problems, and experiments
on public OpenQA datasets show that it significantly outper-
forms traditional RC baselines and recent OpenQA baselines.
1 Introduction
Open-domain question answering (OpenQA) aims to seek
answers for a broad range of questions from a large know-
ledge sources, e.g., structured knowledge bases (Berant et
al. 2013; Mou et al. 2017) and unstructured documents from
search engine (Ferrucci et al. 2010). In this paper we fo-
cus on the OpenQA task with the unstructured knowledge
sources retrieved by search engine.
Inspired by the reading comprehension (RC) task flour-
ishing in the area of natural language processing (Wang
and Jiang 2016; Seo et al. 2016; Xiong, Zhong, and Socher
2016), some recent works have viewed OpenQA as an
RC task, and directly applied the existing RC models to
it (Chen et al. 2017; Joshi et al. 2017; Wang and Jiang 2016;
Clark and Gardner 2018). However, these RC models do not
well fit for the OpenQA task.
Firstly, they directly omit the paragraphs without answer
string
1
. RC task assumes that the given paragraph contains
the answer string (Figure 1 top), however, it is not valid
Copyright
c
2019, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
1
The answer string is a piece of text that can answer the ques-
tion. If the answer string is obtained in a paragraph as a consecutive
text, we call it the answer span.
Reading Comprehension
Open-domain Question Answering
Question: What does a camel store in its hump?
Paragraph1(multiple-answer-spans): The humps are reservoirs of
fatty tissue: concentrating body fat in their humps minimizes th e
insulating effect fat would have if distributed over the rest of their
bodies, helping camels survive in hot climates.
Paragraph2(no-answer-span): Camels with one hump are called
Arabian camels, or Dromedaries, and come from North Africa. Camels
with two humps are from Asia, and are called Bactrian camels.
Paragraph: At standard temperature and pressure, two atoms of the
element bind to form dioxygen, a colorless and odorless diatomic gas
with the formula O2.
Question: How many atoms combine to form dioxygen?
Answer: two
Answer: fat
Satisfied
Relevant
Figure 1: Examples of RC task and OpenQA task.
for the OpenQA task (Figure 1 bottom). That’s because the
paragraphs to provide answer for an OpenQA question is
collected from a search engine, where each retrieved para-
graph is merely relevant to the question. Therefore, it con-
tains many paragraphs without answer string, for instance, in
Figure 1 Paragraph2. When applying RC models to OpenQA
task, we have to omit these paragraphs in the training phase.
However, during the inference phase, when model meets one
paragraph without answer string, it will pick out a text span
as an answer span with high confidence, since RC model
has no evidence to justify whether a paragraph contains the
answer string.
Secondly, they only consider the first answer span in
the paragraph, but omit the remaining rich multiple answer
spans. In RC task, the answer and its positions in the para-
graph are provided by the annotator in the training data.
Therefore RC models only need to consider the unique an-
swer span, e.g., in SQuAD (Rajpurkar et al. 2016). How-
ever, the OpenQA task only provides the answer string as
the ground-truth. Therefore, multiple answer spans are de-
tected in the given paragraph, which cannot be considered
by the traditional RC models. Take Figure 1 as an example,
all text spans contain ‘fat’ are treated as answer span, so we
detect two answer spans in Paragraph1.
Thirdly, they assume that the start position and end posi-
下载后可阅读完整内容，剩余7页未读，立即下载
weixin_38630324
- 粉丝: 3
- 资源: 890
我的内容管理
展开
最新资源
- 十种常见电感线圈电感量计算公式详解
- 军用车辆：CAN总线的集成与优势
- CAN总线在汽车智能换档系统中的作用与实现
- CAN总线数据超载问题及解决策略
- 汽车车身系统CAN总线设计与应用
- SAP企业需求深度剖析：财务会计与供应链的关键流程与改进策略
- CAN总线在发动机电控系统中的通信设计实践
- Spring与iBATIS整合：快速开发与比较分析
- CAN总线驱动的整车管理系统硬件设计详解
- CAN总线通讯智能节点设计与实现
- DSP实现电动汽车CAN总线通讯技术
- CAN协议网关设计：自动位速率检测与互连
- Xcode免证书调试iPad程序开发指南
- 分布式数据库查询优化算法探讨
- Win7安装VC++6.0完全指南：解决兼容性与Office冲突
- MFC实现学生信息管理系统：登录与数据库操作
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
