advisor: hsin-hsi chen reporter: chi-hsin yu date: 2010.08.05
TRANSCRIPT
Advisor: Hsin-Hsi Chen
Reporter: Chi-Hsin Yu
Date: 2010.08.05
Query Log Related Topics
IntroductionTopic A: Relevance Assignment
Using Query Log to Assign Actual Relevance of Documents for User Communities
Topic B: Knowledge Transfer (cloud wisdom in QL)Transfer Learning in Query Log
Topic C: Query UnderstandingSemantic Relation of Query Terms in Query
LogDiscussions
Outlines
Query Log
Introduction (1/3)
From: Intent Based Clustering of Search Engine Query Log, 2009
IssuesPerceived relevance v.s. Actual relevance Clicks
Click bias in positions Relevance
Query – documentUser intent/goal – queries – documents User community – queries – documents
Cost Editorial judgments v.s. model predicted judgments
Introduction (2/3)
Introduction (3/3)
From: SIGIR 2010 Tutorials
Task (original)Assign relevance judgment for a q-d pair
Relevance Assignment (1/4)
21121
Actual Relevance
From: Intent Based Clustering of Search Engine Query Log, 2009
Applications of the predicted relevance judgments (pr)As meta-features
As actual relevanceLow cost
Relevance Assignment (2/4)
Samples(Matrix)
pr
RankingAlgorithms
Performance in dataset (editorial judgment)
ButA Dynamic Bayesian
Network Click Model for Web Search Ranking (WWW2009, Track: Data
Mining/Session: Click Models)
ExperimentsPredicting click-through ratePredicted relevance as a
ranking featureLearning a ranking function
with predicted relevance
Relevance Assignment (3
From: (WWW2009, Track: Data Mining/Session: Click Models)
Task (Revised) Assign relevance judgment for a ((user community, q), d) pair
Not q-d pair Pseudo-query: (user community, q)
Models GA DBN: same as proposed click models in WWW09 papers
Difficulties Pseudo-query generation (include user information) User clustering/classification
Evaluation Joint training (as in the WWW09 papers)
Application For detail analysis of personalized search
because we can use predicted relevance to substitute the editorial judgment
Relevance Assignment (4/4)
Query log = cloud wisdom
Task: Mining/leverage cloud wisdom in QLUse transfer learningUse QL to learn meta-features
Knowledge Transfer (1/3)
Task Leverage useful structure/knowledge in QL to boost
performance of existing datasets (human judgment)Algorithm
SCL: structure correspondence learning Difficulties
Selection of extended feature s in QLEvaluation
As common IR evaluation metricsExpected results (planned experiments)
Can improve performance when use whole training datasetCan improve performance when using small training
dataset
Knowledge Transfer (2/3)
SCL ACL 2005, EMNLP 2006Domain Adaptation with Structural
Correspondence Learning (EMNLP 2006)
Knowledge Transfer (3/3)
From: EMNLP 2006
From Google search suggestions
Interpretation “machine learning wiki/amazon” Concept + in site “machine learning stanford” concept + in organization “machine learning tutorial/tool/ppt/journal” concept + in
topic/resources “machine learning kernel” concept + topic
Semantic Relations of Query Terms (1/6)
Compare to compound noun semanticsDiarmuid ´O S´eaghdha, 2008
Semantic Relations of Query Terms (2/6)
From: Diarmuid ´O S´eaghdha, 2008
Beyond static semantic relations
Dynamic semantic relations recognition What is the patterns in the process of query
reformulation?Is this useful to identify user goal in a session? Can we build new click model based on semantic
relation?
Semantic Relations of Query Terms (3/6)
Pseudo-session A 1. Apple2. Apple ipod3. Apple ipod discount
Pseudo-session A 1. <concept>2. <organization> <product>3. <organization> <product>
<topic/concept>
Current works (incomplete)
Introduction – Semantic Relations of Query Terms (4/6)
From: SIGIR 2010 Tutorials
Research planDefinition of semantic relations in QL
Use Google query suggestions to study the types of semantic relations
Segmentation of query termsMapping segmented query terms to ontologyClassification of semantic relation in QLMining important statistics from QLApplications
Ranking strategies based on SRsClick models based on SRs
Semantic Relations of Query Terms (5/6)
Task Study of semantic relations of query terms
AlgorithmQuery Segmentation, classification, statistics
miningDifficulties
Depends ...Evaluation
Depends ...Expected results
New problem in NLP and in IR
Semantic Relations of Query Terms (6/6)
Thanks for Your Attention. Discussions