Queensland University of Technology An Ontology-based Mining Approach for User Search Intent Discovery Yan Shen, Yuefeng Li, Yue Xu, Renato Iannella, Abdulmohsen.

Queensland University of Technology

An Ontology-based Mining Approach for User Search Intent Discovery

Yan Shen, Yuefeng Li, Yue Xu, Renato Iannella, Abdulmohsen Algarni and Xiaohui Tao

ADCS 2011, 2nd Dec, Canberra

CRICOS No. 00213Ja university for the worldrealR

2

Outline

• Introduction• Related work• Proposed Approach

– An overview of the architecture – World knowledge base– Personalized ontology construction– In-levels ontology mining method


3

Outline

• Evaluation– Data collections – Measures & Baseline model– Results and findings– Discussion

• Conclusion and future work


4

Introduction

• Retrieving desired information to a user is the primary objective of an effective search engines

• Many efforts are spent to improve search capabilities, e.g….

• No doubt that they are helpful, however, they are commonly encountering an issue – information mismatch (ambiguity)


5

Introduction

• To overcome the issue, more and more researchers have taken ontologies into account

• The ontologies can classify diverse knowledge into a well-structured way, which facilitate users to assess information items

• Moreover, semantic relations can be considered to enhance information navigation


6

Introduction

• Note that user search intent is a significant aspect to return desired information

• We study search intents into two means: Specificity and Exhaustivity intent

• A hierarchical concept level-finding technique is proposed to discover and characterize user search intents


7

Introduction

• an ontology-based approach is introduced

• Library of Congress Subject Headings is applied as a world knowledge base for learning personalized ontologies

• In-levels ontology mining method is fully described

• Evaluated by 100 RCV1 topics in TREC 2002 Filtering Track

• The results indicate that the performance of top precision is improved dramatically.


8

Related work

• Ontology-based techniques– Zhong proposes a learning approach for task (or domain-specic)

ontology, which employs various mining techniques and natural-language understanding methods.

– Li and Zhong present an automatic ontology learning method, in which a class is called a compound concept, assembled by primitive classes that are the smallest concepts and cannot be divided any further.

– …


9

Related work

• Ontology-based techniques

– They don't consider the purpose of discovering and characterizing user search intents in a concept level.

– To extend the previous methods, the paper uses “Is-A“ relation to build a real hierarchical structure for the backbone of personalized ontologies


10

Related work

• User information needs– Jiang and Tan aim to represent and capture users' interests in target

domain. Subsequently, a method, they called Spreading Activation Theory (SAT), is employed for providing personalized services.

– Tao et al. propose an ontology-based knowledge retrieval framework to capture user information needs by considering user knowledge background and user's local instance repository with association roles and data mining techniques.

– …

– They are normally either expensive in extraction or inaccurate in description.


11

Proposed approach

• The paper first holds a hypothesis that a user search intent should exist somewhere in an ontology.

• The intent could be general or specific, and can be represented in a range of extent


12

Proposed approach

• An overview of the approach


13

Proposed approach

• World knowledge base (LCSH)– In the LCSH, subject headings are basic semantic units for conveying

domain knowledge and concepts, they have three main types of references: Broader Term, Narrower Term and Related Term.

– Refine to ancestor and descendant lexical relations respectively in our approach


14

Proposed approach

• World knowledge base (cont.)– Definitions


15

Proposed approach

• Personalized ontology learning– Concept hierarchy is an essential object of ontology learning

– Here, we create an abstract hieratical structure


16

Proposed approach

• Personalized ontology learning (cont.)– Definitions


17

Proposed approach

• Personalized ontology learning (cont.)– An example


18

Proposed approach

• In-Levels ontology mining method– Represent feature in levels (two objectives)

• 1) to decide subjects and weights for the pilot level;

• 2) to represent it as a query

After that, do a query expansion. Then, obtain a feature as:


19

Proposed approach

• In-Levels ontology mining method (cont.)– Determine the best level for user search intents


20

Evaluation

• Data collections– A LCSH (QUT Library data in 2008) database 719 mega bytes data

stored in Microsoft Office Access Database (.mdb), totally 491,250 subjects associated with semantic relations

– TREC-11 2002 Filtering Track, RCV1, totally 806,791 xml documents in training and testing sets.

– All of them are processed by the pre-processing approach (stopwords removal, stemming)


21

Evaluation

• Measures & Baseline model– Top 20 precision (pr@20), the precision averages at 11 standard recall

levels (11-points), the Mean Average Precision (MAP), and the F1-Measure.

– ONTO model (Tao et al., 2010)

– Two uniform level settings in upper level 7 and lower level 2 respectively.


22

Evaluation

• Results and Findings


23

Evaluation

• Results and Findings (cont.)


24

Evaluation

• Results and Findings (cont.)


25

Evaluation

• Discussion– The approach by only containing new terms has better performance than

the one keeps all the terms in levels

– Demonstrate the validity of the hierarchical backbone

– The experimental results are indistinct for all the measures, and those specific terms might be able to reduce recall

– The approach is suitable to situations when precision is be considered more important than others

– LCSH is difficult to keep up to date


26

Conclusion

• The paper introduces an ontology-based approach to discover user search intents

• The approach involves a subject-based search model, a world knowledge base, and a in-levels ontology mining method

• The empirical results indicate that our approach works remarkable on top precision

• The main intellectual contribution is the hierarchical level-finding technique


27

Future work

• Investigate the usage of the rest of semantic relations in LCSH

• Combine with pattern mining methods

• Test the approach with other world knowledge base, like WordNet or Amazon


28

• Thank you for listening, any question?

Queensland University of Technology An Ontology-based Mining Approach for User Search Intent Discovery Yan Shen, Yuefeng Li, Yue Xu, Renato Iannella, Abdulmohsen.

Documents

world real r

ontologybased approach

information navigation

canberra slide

ontologybased mining

future work slide

levels ontology mining

real hierarchical structure