Vous êtes sur la page 1sur 17

Adaptive User Profile Management

for cQA System


Kigon Lyu, Jungyong Lee, Hyeoncheol Kim, and Heuiseok Lim
Dept. of Computer science Education, Korea University, Korea
{gon0121, popobo, harrykim, limhseok} @ korea.ac.kr
Contents
ü
ü Research Background


ü Purpose of the Study


ü Architecture


ü Experiments


ü Conclusion

Intelligent Information Analysis Lab.


Research Background (1/6)

• Importance of information retrieval


– To satisfy the user’s information need
– To search for the relevant information from the information overload


• However, information retrieval often cannot offer the exact
information to users

• For example,
can I book the hotel which provides Korean food in Bali?

Intelligent Information Analysis Lab.


Research Background (2/6)

• Results are received from search engines



1. Hotel Bali – Great Deals on Bali Hotels, Bali Villas and Bali ...

 2. Bali hotels, Bali Accommodation Indonesia, Hotel.com.au Bali Hotel ...

3. Bali Hotel and Villa - Indonesia Travel Portal



4. St Regis Bali Hotels: The St. Regis Bali Resort - Hotel Rooms at ...

 5. Bali Bar and Restaurant - Foods, Dining, Eating out, Cafes, Bars ...
.
.
 .

• Not enough to satisfy Intelligent Information Analysis Lab.


Research Background (3/6)

• To solve the problem


– Many alternative approaches have been studied
– Enhanced web-crawler, indexer, search algorithm and so on


• People-powered approach
– Utilize user’s direct/indirect participation
– Users can be a information provider as well as a consumer


Question Direct
People-powered
Answer Search Indirect
Interne
t

Intelligent Information Analysis Lab.


Research Background (4/6)

• Community-based Question Answering


– One of the people-powered approach
– Users participate in direct answering the question
– Yahoo! Answers, Naver, ChaCha.com, WonDir.com, and so on

Pros Cons
 Get answers to complex or obscure information needs Users can be either experts or non-experts
 

See multiple, possibly differing opinions about a


  Possibility of receiving no answer at all to a question
question 

  Having to wait for an answer (possibly days)


Interact with other users who may share common 

interests, problems, and goals Receiving answers that are incorrect, misleading,

offensive, or spam

Intelligent Information Analysis Lab.


Research Background (5/6)

• Good way to get answers in our example


– From someone who have been to there
– From someone who work at travel agency or guide


• Suppose that the user ask to cQA Visitor
Guide

 Question
Answer

 Answer
cQA


Travel agency

Interne
 t

 Intelligent Information Analysis Lab.


Research Background (6/6)

• Importance of user profile


– Automatically find potential users who can answer the question


• Constructing user profile
– User profile can be constructed by oneself

Pros – To construct reliable profile require


Cons their detailed information



Filtering non-relevant users to given questions 


Very hard and annoying to users

Increase the possibility of offering high-quality


  low-quality inputs can make low-quality profile
answer – 

 Users profile does not change over time





Intelligent Information Analysis Lab.
Purpose of the study
• Systematical collection
– Collecting interest and expertise of users by mandatory input
– Collecting implicit information by user’s activity histories


• Effective representation
– Using conventional weighting techniques by keywords
– Finding potential relevant user profile by concepts


• Adaptive update
– Updating user profile by analyzing adaptively their implicit
information

Intelligent Information Analysis Lab.


Architecture
Query
Community-based
Users
Another User Answer
Question Answering

Query Answer Community Query Analysis


Log Log Activity
Explicit
Profile Search
Information
Implicit
Information

Adaptive User Profile Management User Profile

Collection

Representation

Updating

Intelligent Information Analysis Lab.


Architecture - Collection
• Explicit Information
– Direct interaction with system
• Select fields, radio buttons, text fields and so on
• Collect personal information, interest, and expertise from
registration form
• Also, interest and expertise of user are required as the mandatory
input
– Weakness
• Can not guarantee consistency
• Does not change over time
• Consist of words or short-sentences


• Implicit Information
– Systematical analysis
• Analyzing words in weblog, querylog, and answerlog
– Advantage
• Collect implicit information without user’s participation
Intelligent Information Analysis Lab.
Architecture - Representation
• Keyword-based
– A vector is composed of
User1: travel agent

one more words schedule book flight landscape
·······
– It is very hard to find 0.63 0.49 0.31 0.23


relevant user profile
• Word ambiguity
• Representing implicit information

– User1: travel agent

• Concept-based schedule book


·······
flight landscape

– Representing user profile 0.63 0.49 0.31 0.23


as more abstract concept
– Find potential relevant travel agency tour


user profile 1.66 0.54

– Does not considered the ambiguity

Intelligent Information Analysis Lab.


Architecture - Updating
• Rocchio’s algorithm can be applied to update
– Initial profile as initial query
– Updated profile as expanded query


• In addtion, we created the new equation
– Excepting a negative operation
– Calculate automatically without user feedback


t
pi = α pi + β ∑ a j
• Updated profile are calculated according to:
new

j =1


(0 ≤ α , β ≤ 1) (α + β = 1)

α andβ are parameters for control weighting between initial profile and updated profile

Intelligent Information Analysis Lab.


Experiments (1/2)

• Experimental data
– Questions and answers submitted to Wispoon for a month
– 170 QA pairs, 10 pairs from each 17 categories


Dataset
Category
Avg. # of
questions per
users
Avg. of answers Avg. query
per question length (in # of
words)
Avg. response
time (sec.) per
question
Computers & Internet 5.3 5 5.6 2.0


Business & Economy 3.0 3 6.4 3.0
Health 4.3 4 5.9 8.0


Entertainment
Shopping
3.6
3.0
3
3
3.5
6.6
2.4
3.0


Society & Culture 3.6 3 6.5 2.4





– Intelligent Information Analysis Lab.
Experiments (2/2)

• Experimental results
 Dataset Category MOS
Before After
Computers & Internet 4.0 5.0

Business & Economy 3.5 4.0

 Health 4.0 4.5

Entertainment 3.5 4.5



Shopping 3.0 4.5

Society & Culture 3.5 4.5


• Considerably higher after than before the updating


Intelligent Information Analysis Lab.
Conclusion

• Collect implicit information by analyzing user activity
histories


• Create and update user profile using implicit information


• Performance of cQA can be enhanced by a implicit
information


• User profile can be reliable and consistent by proposed
technique

Intelligent Information Analysis Lab.


Thank you

Intelligent Information Analysis Lab.

Vous aimerez peut-être aussi