190 likes | 212 Views
Explore the impact of using query logs in improving search engine results, incorporating history instances to refine language models for better information retrieval. Experiment with different weighting techniques.
E N D
Relevance feedback using query-logs Gaurav Pandey Supervisors: Prof. Gerhard Weikum Julia Luxenburger
Motivation Query Search Engine Results “One size fits all”
Motivation User info Query Search Engine Results
Motivation Python
Motivation CGI code Debugging programming Python
Usage of Query Logs Clickthrough data • Past queries • Documents clicked
Query Clicked Documents History Instance Usage of Query Logs
Result Query: “python information CGI code examples program code debugging bug removal programming” But, p(python/query)=? p(CGI)/query)=? p(code)/query)=? ………………………….. Query Reformulation
Considers only the current query But, not history instances Language Model Normally(without using history), w:term d: document q:query Importance of term w in current query
Now, using history: ? Importance of term w in current query + history instances Language Model Normally(without using history), w:term d: document q:query Importance of term w in current query
History query: “CGI code” Documents: “CGI examples”, “program code” History query: “CGI code” Language Model+History Importance of term w in history instances Importance of the term w at one instance in the history Documents: “CGI examples”, “program code”
Works,but can be improved Equal Weighting
Discriminative Weighting Choose different for every history instance.. How?
Overlap if a history query has common terms with the current query then λi= 1, Else if there is no common term λi=0 Example: Current query “python information” History query:”python code” λi= 1 History query:”world cup” λi= 0
Soft overlap if a history query has common terms with the current query then λi= a, Else if there is no common term λi=b (a>b) Example: Current query “python information” History query:”python code” λi= 8 History query:”world cup” λi= 2
Decrease with time Use uniformly decreasing values If there are n history instances, 1 =n 2 =n-1 3 =n-2 …… n-1 =2 n =1
Decrease with time Use geometrically decreasing values If there are n history instances, 1 =n 2 =n/2 3 =n/3 …… n-1 =n/(n-1) n =1
Experiment • Comparison of the • 4 techniques • Equal weighting • Basic model (without history) • Use similar techniques for: • Probabilistic model • Vector space model