. IN CQAS SUCH AS YAHOO! ANSWERS, A RESOLVEDSERVICES (CQAS) TO AD...

Question

2007). In cQAs such as Yahoo! Answers, a resolved

services (cQAs) to address the problem of “in-

question often gets more than one answers and a

complete answer”, i.e., the “best answer” of a

“best answer” will be chosen by the asker or voted

complex multi-sentence question misses valu-

by other community participants. This {question,

able information that is contained in other an-

best answer} pair is then stored and indexed for fur-

swers. In order to automatically generate anovel and non-redundant community answer

ther uses such as question retrieval. It performs very

summary, we segment the complex original

well in simple factoid QA settings, where the an-

multi-sentence question into several sub ques-

swers to factoid questions often relate to a single

tions and then propose a general Conditional

named entity like a person, time or location. How-

Random Field (CRF) based answer summary

ever, when it comes to the more sophisticated multi-

method with groupL

1

regularization. Vari-

sentence questions, it would suffer from the problem

ous textual and non-textual QA features are

of “incomplete answer”. That is, such question often

explored. Specifically, we explore four differ-ent types of contextual factors, namely, the in-

comprises several sub questions in specific contexts

formation novelty and non-redundancy mod-

and the asker wishes to get elaborated answers for as

eling for local and non-local sentence inter-

many aspects of the question as possible. In which

actions under question segmentation. To fur-

case, the single best answer that covers just one or

ther unleash the potential of the abundant cQA

few aspects may not be a good choice (Liu et al.,

features, we introduce the group L

1

regu-

2008; Takechi et al., 2007). Since “everyone knows

larization for feature learning. Experimental

something” (Adamic et al., 2008), the use of a single

results on a Yahoo! Answers dataset showthat our proposed method significantly outper-

best answer often misses valuable human generated

forms state-of-the-art methods on cQA sum-

information contained in other answers.

marization task.

In an early literature, Liu et al.(2008) reported that

no more than 48% of the 400 best answers were in-

1 Introduction

deed the unique best answers in 4 most popular Ya-

hoo! Answers categories. Table 1 shows an example

Community Question and Answering services

of the “incomplete answer” problem from Yahoo!

(cQAs) have become valuable resources for users

Answers

¹

. The asker wishes to know why his teeth

to pose questions of their interests and share their

bloods and how to prevent it. However, the best an-

knowledge by providing answers to questions. They

swer only gives information on the reason of teeth

perform much better than the traditional frequently

asked questions (FAQ) systems (Jijkoun and Rijke

1

https://traloihay.net

, 2005; Riezler et al., 2007) which are just based

20100610161858AAmAGrV

582 blooding. It is clear that some valuable information

and incorporate four different contextual factors

about the reasons of gums blooding and some solu-

based on question segmentation to model the local

and non-local semantic sentence interactions to ad-

tions are presented in other answers.

dress the problem of redundancy and information

Question

novelty. Various textual and non-textual question

Why do teeth bleed at night and how do you prevent/stop it? This

morning I woke up with blood caked between my two front teeth.

answering features are exploited in the work.

This is the third morning in a row that it has happened. I brush and

Second, we propose a group L

₁

-regularization ap-

floss regularly, and I also eat a balanced, healthy diet. Why is this

proach in the CRF model for automatic optimal fea-

happening and how do I stop it?

ture learning to unleash the potential of the features

Best Answer - Chosen by Asker

and enhance the performance of answer summariza-

Periodontal disease is a possibility, gingivitis, or some gum infec-

tion.

tion. Teeth don’t bleed; gums bleed.

We conduct experiments on a Yahoo! Answers

dataset. The experimental results show that the

Vitamin C deficiency!

proposed model improve the performance signifi-

Ever heard of a dentist? Not all the problems in life are solved on

the Internet.

cantly(in terms of precision, recall and F1 measures)

You could be brushing or flossing too hard. Try a brush with softer

as well as the ROUGE-1, ROUGE-2 and ROUGE-L

bristles or brushing/flossing lighter and slower. If this doesn’t solve

measures as compared to the state-of-the-art meth-

your problem, try seeing a dentist or doctor. Gums that bleed could

ods, such as Support Vector Machines (SVM), Lo-

be a sign of a more serious issue like leukemia, an infection, gum

disease, a blood disorder, or a vitamin deficiency.

gistic Regression (LR) and Linear CRF (LCRF)

wash your mouth with warm water and salt, it will help to strengthen

(Shen et al., 2007).

your gum and teeth, also salt avoid infection. You probably have

The rest of the paper is arranged as follows: Sec-

weak gums, so just try to follow the advice, it works in many cases

of oral problems.

tion 2 presents some definitions and a brief review

of related research. In Section 3, we propose the

Table 1: An example of question with incomplete answer

summarization framework and then in Section 4 and

problem from Yahoo! Answers. The “best answer” seems to

5 we detail the experimental setups and results re-

miss valuable information and will not be ideal for re-use when