1' Introduction

1
FDUQA--an Open Domain Question Answering
SystemJunkuo Cao, Bo Li, Zhongchao Fei, Xiaofeng
Yuan, Yaqian Zhou, Xuanjing Huang and Lide WU
Department of Computer Science and Engineering,
Fudan University220 Handan Rd., Shanghai 200433,
ChinaEmail zhouyaqian, xjhuang,
ldwu_at_fudan.edu.cn
CJNLP 2006The 6th China-Japan Natural Language
Processing Joint Research Promotion Conference

• 1. Introduction
• QA (question Answering) system return an actual
  answers rather than a ranked list documents in
  response to a question asked in natural language.
  
• FDUQA is a English question answering system
  designed and developed by Fudan IR/NLP Group of
  Media Computing and Web Intelligence Lab. The
  system take in the user-input English question,
  such as How far is it from earth to Mars? and
  output the best answer for the question. Figure 1
  shows the process of FDUQA question answering
  system

possible by question target firstly, and then
apply the knowledge to pick up the question
answers. The knowledge includes online
definitions and relative words. The system is as
Fig. 3.
Fig.1 the sketch map and user interface of FDUQA
System
1.1 Background of Question Answering Question
Answering is formally put forward in the 1999
TREC (Text Retrieval Conference)
http//trec.nist.gov Then become a formal track
which is held each year till now. Each year, many
universities and research institutes take part in
QA track, and lots of top journals and
conferences in computer science, such as AAAI,
SIGIR, ACL, etc. set up related sessions on
Question Answering. 1.2 History of FDUQA FDUQA
has took part in TREC QA track each year since
1999. It is one of the earliest question
answering system in the world and the first one
in China. In the evaluation of TREC QA track,
FDUQA has good performance, and ranked high in
all the participants. 1.3 Category of Question
Answering TREC classify the question answering
into three categories Factoid question
answering, Definition question answering and List
question answering. For the factoid question
answering, the system give one short answer of
some fact. The Definition question answering
return all the sentences that is related the
target of the question. The List question
answering will find all the possible answers for
the given question.

4. List Question Answering List Question is to
return a series of answers. For example, List the
country of north American. Our list question
answering system is composed of Answer Type
Classification module, Document Searching module,
Sentence Scoring Module, Answer Extraction
Module, Answer Ranking Module and Answer
Filtering Module. The system is as Fig. 4.

2. Factoid question answering Factoid questions
are simple ,short question that cares some
specified attribute of facts, for
Example, Question When did the submarine Kursk
sink? Answer August 12th 2000 2.1 the step
of answering factoid question Step1 Resolve
target which is an anaphora used in
questions. Step2 Question Analysis, To extract
constituents from the question sentence. The
Analysis is based on LinkParser Step3 Answer
Type Classification. The type of answer is
important for selecting answers. We consider the
POS, interrogative, Sentence constituent etc. in
the sentence to do it. WordNet are included as an
external resource. Step4 Query Generation and
Web Retrieval. The queries are generated based
on question sentence. Then Synonym extension,
Preposition extension and Unit extension will be
done on the sentence. The queries will be put in
Google to search for the candidate answer. Then
the candidate answer will be put in the Aquiant
Corpus for answer projecting. Step5 Document
Retrieval and Candidate Answer Generation. Lucene
is used to retrieve documents from Aquaint
Corpus. The final answer is extract from the
retrieved documents using Named-entities
extraction technology according the answer
type. 3. Definition Question answering
Definition questions, i.e., questions like What
is Warren Moon? or Who is Paul Revere? have
drawn much attention recently. The typical task
of definitional QA is to find out conceptual
facts or essential events about the question
target. In order to automatically identify
definition sentences from a large collection of
documents, we extract related knowledge as much as
Fig.4 The structure of the List question
Answering system
5. FDUQA at Trec Table 1 lists the rank of FDUQA
system at the recent years Trec evaluation. Table
2 is this year's Trec evaluation result of FDUQA
system.