1. IR Theory: Web Information Retrieval

2. Web IRFusion IR Search Engine 2

3. Evolution of IR: Phase I  Brute-force Search  User  Raw Data  Library  Collection Development  Quality Control  Classification  Controlled Vocabulary  Bibliographical Records  Browsing  User  Organized/Filtered Data  Searching  User  Intermediary  Metadata  Organized/Filtered Data Search Engine 3

4. Evolution of IR: Phase II  IR System  Automatic Indexing  Pattern Matching  User  Computer  Inverted Index  Raw Data  Move from metadata to content-based search  IR Research  Goal  Rank the documents by their relevance to a given query  Approach  Query-Document Similarity  Term-Weights based on term occurrence statistics  Query  Document Term Index  Ranked list of matches  Controlled and restricted experiments with small, homogeneous, and high quality data Search Engine 4

5. Evolution of IR: Phase III  World Wide Web  Massive, uncontrolled, heterogeneous, and dynamic environment  Content-based Web Search Engines  Web Crawler + Basic IR technology  Matching of query terms to document terms  Web Directories  Browse/Search of Organized Web  Manual cataloging of Web subset  Content- & Link-based Web Search Engines  Pattern Matching + Link Analysis  Renewed interest in metadata and classification approach  Digital Libraries?  Integrated (Content, Link, Metadata) Information Discovery Search Engine 5

6. Fusion IR: Overview  Goal  To achieve the whole that is greater than sum of its parts  Approaches  Tag team  Use a method best suited for a given situation  Single method, single set of results  Integration  Use a combined method that integrates multiple methods  Combined method, single set of results  Merging  Merge the results of multiple methods  Multiple methods, multiple sets of results  Meta-fusion  All of the above Search Engine 6

7. Fusion IR: Research Areas  Data Fusion  Combining multiple sources of evidence  Single collection, multiple representations, single IR method  Collection Fusion  Merging the results of multiple collection search  Multiple collections, single representation, single IR method  Method Fusion  Combining multiple IR methods  Single collection, single representation, multiple IR methods  Paradigm Fusion  Combining content analysis, link analysis, and classification  Integrating user, system, and data Search Engine 7

8. Fusion IR: Research Findings  Findings from content-based IR experiments with small, homogeneous document collections  Different IR systems retrieve  different sets of documents  Documents retrieved by multiple systems  are more likely to be relevant  Combining different systems  is likely to be more beneficial than combining similar systems  Fusion is good for IR  Is fusion a viable approach for Web IR? Search Engine 8

9. Web Fusion IR: Motivation  Motivation  Web search has become a daily information access mechanism  2.4 Billion Internet Users (532% growth from 2000) (InternetWorldStats.com, 2012)InternetWorldStats.com  96% of Web users access the Internet daily. (www.internetsociety.org, 2012)www.internetsociety.org  91% of Web users use search engines to find information. (Pew Internet, 2012)Pew Internet  5.1 billion Google searches per day (www.statisticbrain.com, 2012)www.statisticbrain.com  New Challenges  Data: massive, dynamic, heterogeneous, noisy  Users: diverse, “transitory”  New Opportunities  Multiple sources of evidence – content, hyperlinks, document structure, user data, taxonomies  Data abundance/redundancy  Review  Yang (2005). Information Retrieval on the Web, ARIST Vol. 39  http://kiyang.kmu.ac.kr/pubs/webir_arist.pdf http://kiyang.kmu.ac.kr/pubs/webir_arist.pdf Search Engine 9

10. Link Analysis: PageRank  PageRank score: R(p i )  Propagation of R(p i ) through inlinks of the entire Web  T = total # of pages in the Web d = damping factor p i = inlink of p C(p i ) = outdegree of p i  Start w/ all R(p i )=1, repeat computation until convergence  Global Measure of a page based on link analysis only  Interpretation  Models the behavior of random Web surfer – A probability distribution/weighting function that estimates the likelihood of arriving at page p by link traversal and random jump (d).  Importance/Quality/Popularity of a Web page – A link signifies recommendation/citation – aggregate all recommendations recursively over entire Web, where each recommendation is weighted by its importance and normalized by its outdegree Search Engine 10

11. PageRank Simplified 11

12. Link Analysis: HITS  Hyperlink Induced Topic Search  Consider both inlinks & outlinks  estimates the value of a page based on aggregate value of in/outlinks  Identify “authority” & “hub” pages  authority = a page pointed to by many good hubs  hub = a page pointing to many good authority  Query-dependent measure  hub & authority scores assigned for each query  computed from a small subset of the Web – i.e. top N retrieval results  Premise  Web contains mutually reinforcing communities of hubs & authorities on broad topics Search Engine 12

13. Link Analysis: Modified HITS  HITS-based Ranking 1.Expand a set of Text-based search results  Root set S = top N documents (e.g. N=200)  Inlinks & Outlinks of S (1 or 2 hops) – Max. k inlinks per document (e.g. k=50) – Delete intrahost links, stoplist URLs 2.Compute Hub and Authority scores  Iterative algorithm  Fractional weights to links by same authors 3.Rank documents by Authority/Hub scores Search Engine 13

14. Modified HITS : Scoring Algorithm 1.Initialize all h(p) and a(p) to 1 2.Recompute h(p) and a(p) with fractional weights - normalize contribution of authorship (assumption: host=author) a(p)=  (h(q)*auth_wt(q,p)) q is a page linking to p auth_wt (q,p) = 1/m for page q, whose host has m documents linking to p h(p)=  (a(q) *hub_wt(p,q)) q is a page linked from p hub_wt(p,q) = 1/n for page q, whose host has n documents linked from p 3.Normalize scores  divide score by square root of sum of squared scores (  a(p)=  h(p)=1) 4.Repeat steps 2 & 3 until scores stabilize  Typical convergence in 10 to 50 iterations for 5000 webpages Search Engine 14

15. Modified HITS : Link Weighting Search Engine 15 p q1q1 q2q2 q3q3 q4q4 h(p)= a(q 1 ) + a(q 2 ) + a(q 3 ) + a(q 4 )/6 q1q1 q2q2 q3q3 q4q4 a(p)= h(q 1 ) + h(q 2 ) + h(q 3 ) + h(q 4 )/5 p

16. WIDIT: Web IR System Overview 1.Mine Multiple Sources of Evidence (MSE)  Document Content  Document Structure  Link Information  URL information 2.Execute Parallel Search  Multiple Document Representations  body text, anchor text, header text  Multiple Query formulations  query expansion 3.Combine the Parallel Search Results  Static Tuning of fusion formula (QT-independent) 4.Identify Query Types (QT)  Combination Classifier 5.Rerank the fusion result with MSE  Compute Reranking Feature Scores  Dynamic Tuning of reranking formulas (QT-specific) Search Engine 16

17. WIDIT: Web IR System Architecture Search Engine 17 Indexing Module Sub-indexes Body Index Anchor Index Header Index Documents Topics Queries Simple Queries Queries Expanded Queries Retrieval Module Fusion Module Sub-indexes Search Results Re-ranking Module Fusion Result Final Result Static Tuning Dynamic Tuning Query Classification Module Query Types

18. WIDIT: Dynamic Tuning Interface Search Engine 18

19. SMART  Length-Normalized Term Weights  SMART lnu weight for document terms  SMART ltc weight for query terms where:f ik = number of times term k appears in document i idf k = inverse document frequency of term k t = number of terms in document/query  Document Score  inner product of document and query vectors where:q k = weight of term k in the query d ik = weight of term k in document i t = number of terms common to query & document Search Engine 19

20.  Document term weight (simplified formula)  Query term weight Okapi  Document Ranking where:Q = query containing terms T K = k 1 ((1-b) + b*(doc_length/avg.doc_length)) tf = term frequency in a document qtf = term frequency in a query k 1, b, k 3 = parameters (1.2, 0.75, 7..1000) w RS = Robertson-Sparck Jones weight N = total number of documents in the collection n = total number of documents in which the term occur R = total number of relevant documents in the collection n = total number of relevant documents retrieved Search Engine 20