Dependency Network Based Real-time Query Expansion

Jiaqi Zou, Xiaojie WangCenter for Intelligence Science and Technology,

BUPT

Outline

• Introduction– What is RTQE?– Benefits of RTQE– Related Research– Improvements in our work

• Method– Query Intention– Dependency Relation Network– RTQE Method

Outline

• Experiments– Test of operation numbers– Test of expansion success percentage– Test of retrieval performance– Comparison with Bing

• Conclusion

Introduction- What is RTQE?

• RTQE is a kind of query expansion.• RTQE methods expand queries at

the same time when users type queries into the search box.

Introduction- Benefits of RTQE

• RTQE reduces user’s keystrokes and time to perform a query, especially useful for mobile device users.

• RTQE improves the query quality.

Related Research

• Most widely used method: string matching method using query log.

• Little work on RTQE takes query intention into account.– Strohmaier et al. suggested that explicit

queries containing at least one verb word might reflect possible user intentions.

Improvements in our work

• Represent query intention better.• Construct a RTQE method which

expands components of possible user query intentions.

• This RTQE method improves the retrieval performance.

Query Intention

• Task-oriented classification of query intention: – Navigational– Informational– Transactional

• Duan et al. suggested dependency related verb-noun pairs are good representation of informational and transactional query intentions.

Query Intention

• Verb-noun pair is not sufficient to represent query intention, other parts like attributes of noun are also very important.

• New representation:Verb-Attributes-Noun

• Example: buy new car tire, cook Chinese food

Dependency Relation Network

• To do query intention related RTQE, we built a dependency relation network which is a collection of numbers of query intentions.

• Steps:– Do dependency parsing on large corpus.– Extract all the verb-attributes-noun

structures. – Combine these structures to be the Network.

Dependency Relation Network

• Example : How to change a car tire• Extracted : change car tire

RTQE method

RTQE Example

Experiments

• Corpus: www.ehow.com – 915,000 articles– 20 categories(Health, Cars,

Food&Drink, etc)

Test of operation numbers

• Keystrokes and mouse clicks needed to generate a query is recorded. Each keystroke or mouse click is recorded as an operation.

1

1

( )n

x xx

n

xx

OpFull OpExpandedSaved

OpFull

Test of operation numbers

• Average saved operations is 63.75% after RTQE

Average number of operations

Without RTQE With RTQE

15.0 5.437

Test of expansion success percentage

• For a given query intention, if the user can find a query exactly related to this intention from the expanded list, we call it a successful expansion.

SuccessNumSuccessPercentage

AllNum

Test of expansion success percentage

Times

Query expansion success

Query expansion fail

168 32

• Expansion success percentage is 84%.

Test of retrieval performance

• We compare the retrieval performance of the three: – original query user typed in– the query after verb-noun expansion– the query after verb-attributes-noun

expansion.

• We use precision and nDCG score for evaluation.

Test of retrieval performance

Query Type Precision nDCG score

Original query word 0.73% 13.11%

Query after verb-noun expansion

9.47% 37.37%

Query after verb-attributes-noun expansion

79.2% 88.95%

Comparison with Bing

• The RTQE result of Bing differs a lot if the word order of a query changes.

Comparison with Bing

• Categories the RTQE result of Bing into 3 groups:– NOT: cannot get correct

recommendations – NORMAL: get correct recommendations

only in normal word order– ALL: can get correct recommendations

both in normal order and other word orders

Comparison with Bing

Group NOT NORMAL ALL

Percentage 49% 33% 18%

Conclusion

• Presented a novel RTQE method using a dependency relation network.

• This RTQE method is proved to be effective in representing user query intention and hence improve retrieval performance.

Thank you!