1. 1/28 Efficient Top-k Queries for XML Information Retrieval Gerhard Weikum weikum@mpi-sb.mpg.de http://www.mpi-sb.mpg.de/~weikum/ Joint work with Ralf Schenkel and Martin Theobald Max-Planck- Gesellschaft

2. 2/28 A Few Challenging Queries (on Web / Deep Web / Intranet / Personal Info) Which drama has a scene in which a woman makes a prophecy to a Scottish nobleman that he will become king? Which professors from Saarbruecken (SB) are teaching IR and have research projects on XML? SB IR XML Who was the woman from Paris that I met at the PC meeting where Alon Halevy was PC Chair? Which gene expression data from Barrett tissue in the esophagus exhibit high levels of gene A01g? Are there any published theorems that are equivalent to or subsume my latest mathematical conjecture?

3. 3/28 Professor Name: Gerhard Weikum Address... Country: Germany Teaching:Research: Course Title: IR Description: Information retrieval... Syllabus... BookArticle... Project Title: Intelligent Search of XML Data Sponsor: German Science Foundation... XML-IR Example (1) Select P, C, R From Index Where Professor As P And P = „ Saarbruecken“ And P// Course = „ IR“ As C And P// Research = „ XML“ As R City: SB

4. 4/28 Article... Select P, C, R From Index Where Professor As P And P = „ Saarbruecken“ And P// Course = „ IR“ As C And P// Research = „ XML“ As R Select P, C, R From Index Where ~ Professor As P And P = „ ~ Saarbruecken“ And P// ~ Course = „ ~ IR“ As C And P// ~ Research = „ ~ XML“ As R Professor Name: Gerhard Weikum Address... Country: Germany Teaching:Research: Course Title: IR Description: Information retrieval... Syllabus... Book... Project Title: Intelligent Search of XML Data Sponsor: German Science Foundation... XML-IR Example (2) City: SB Name: Ralf Schenkel Teaching: Literature Book Lecturer Interests: Semistructured Data, IR Address: Max-Planck Institute for CS, Germany Seminar Title: Statistical Language Models... Contents: Ranked Search...

5. 5/28 XML-IR: History and Related Work IR on structured docs (SGML): 1995 2000 2005 IR on XML: XIRQL (U Dortmund) XXL (U Saarland / MPI) XRank (Cornell U) JuruXML (IBM Haifa ) Commercial software (MarkLogic, Verity?, Oracle?, Google?,...) XML query languages: XQuery (W3C) XPath 2.0 (W3C) INEX benchmark Compass (U Saarland / MPI) XPath 1.0 (W3C) XML-QL (AT&T Labs) Web query languages: Lorel (Stanford U) Araneus (U Roma) W3QS (Technion Haifa) TeXQuery (AT&T Labs) FleXPath (AT&T Labs) ELIXIR (U Dublin) HyperStorM (GMD Darmstadt) HySpirit (U Dortmund) WHIRL (CMU) PowerDB-IR (ETH Zurich) ApproXQL (U Berlin / U Munich) Timber (U Michigan) XSearch (Hebrew U)

6. 6/28 XML-IR Concepts Where clause: conjunction of restricted path expressions with binding of variables Select P, C, R From Index Where ~Professor As P And P = „Saarbruecken“ And P//~Course = „Information Retrieval“ As C And P//~Research = „~XML“ As R Elementary conditions on names and contents „Semantic“ similarity conditions on names and contents Relevance scoring based on tf*idf similarity of contents, ontological similarity of names, aggregation of local scores into global scores ~Research = „~XML“ Query result: query is a path/tree/graph pattern results are isomorphic paths/subtrees/subgraphs of the data graph Query result: query is a pattern with relaxable conditions results are approximate matches to query with similarity scores applicable to both XML and HTML data graphs

7. 7/28 Ontologies/Thesauri: Example WordNet woman, adult female – (an adult female person) => amazon, virago – (a large strong and aggressive woman) => donna -- (an Italian woman of rank) => geisha, geisha girl -- (...) => lady (a polite name for any woman)... => wife – (a married woman, a man‘s partner in marriage) => witch – (a being, usually female, imagined to have special powers derived from the devil)

8. 8/28 Ontology Graph An ontology graph is a directed graph with concepts (and their descriptions) as nodes and semantic relationships as edges (e.g., hypernyms). woman human body personality character lady witch nanny Mary Poppins fairy Lady Di heart... syn (1.0) hyper (0.9) part (0.3) mero (0.5) part (0.8) hypo (0.77) hypo (0.3) hypo (0.35) hypo (0.42) instance (0.2) instance (0.61) instance (0.1) Weighted edges capture strength of relationships  key for identifying closely related concepts

9. 9/28 Query Expansion Threshold-based query expansion: substitute ~w by (c 1 |... | c k ) with all c i for which sim(w, c i )   „Old hat“ in IR; highly disputed for danger of topic dilution Approach to careful expansion: determine phrases from query or best initial query results (e.g., forming 3-grams and looking up ontology/thesaurus entries) if uniquely mapped to one concept then expand with synonyms and weighted hyponyms Problem: choice of threshold   see Top-k QP

10. 10/28 Outline Motivation & Preliminaries Prob-k: Efficient Approximative Top-k TopX: Top-k for XML

11. 11/28 Top-k Query Processing with Scoring Naive join&sort QP algorithm: algorithm B+ tree on terms 17: 0.3 44: 0.4... performance... z-transform... 52: 0.1 53: 0.8 55: 0.6 12: 0.5 14: 0.4... 28: 0.1 44: 0.2 51: 0.6 52: 0.3 17: 0.1 28: 0.7... 17: 0.3 17: 0.144: 0.4 44: 0.2 11: 0.6 index lists with (DocId, s = tf*idf) sorted by DocId Given: query q = t 1 t 2... t z with z (conjunctive) keywords similarity scoring function score(q,d) for docs d  D, e.g.: Find: top k results w.r.t. score(q,d) =aggr{s i (d)}(e.g.:  i  q s i (d)) Google: > 10 mio. terms > 8 bio. docs > 4 TB index q: algorithm performance z-transform top-k (  [term=t 1 ] (index)  DocId  [term=t 2 ] (index)  DocId...  DocId  [term=t z ] (index) order by s desc)

12. 12/28 TA (Fagin’01; Güntzer/Kießling/Balke; Nepal et al.) scan all lists L i (i=1..m) in parallel: consider d j at position pos i in Li; high i := s i (d j ); if d j  top-k then { look up s (d j ) in all lists L with  i; // random access compute s(d j ) := aggr {s (dj) | =1..m}; if s(d j ) > min score among top-k then add d j to top-k and remove min-score d from top-k; }; if min score among top-k  aggr {high | =1..m} then exit; m=3 aggr: sum k=2 f: 0.5 b: 0.4 c: 0.35 a: 0.3 h: 0.1 d: 0.1 a: 0.55 b: 0.2 f: 0.2 g: 0.2 c: 0.1 h: 0.35 d: 0.35 b: 0.2 a: 0.1 c: 0.05 f: 0.05 f: 0.75 a: 0.95 top-k: b: 0.8 but random accesses are expensive !  TA-sorted  Prob-sorted applicable to XML data: course = „~ Internet“ and ~topic = „performance“

13. 13/28 TA-Sorted (aka. NRA) scan index lists in parallel: consider d j at position pos i in Li; E(d j ) := E(d j )  {i}; high i := s i (q,d j ); bestscore(d j ) := aggr{x 1,..., x m ) with x i := s i (q,d j ) for i  E(d j ), high i for i  E(d j ); worstscore(dj) := aggr{x 1,..., x m ) with x i := si(q,d j ) for i  E(d j ), 0 for i  E(d j ); top-k := k docs with largest worstscore; if min worstscore among top-k  max bestscore{d | d not in top-k} then exit; m=3 aggr: sum k=2 a: 0.55 b: 0.2 f: 0.2 g: 0.2 c: 0.1 h: 0.35 d: 0.35 b: 0.2 a: 0.1 c: 0.05 f: 0.05 top-k: candidates: f: 0.5 b: 0.4 c: 0.35 a: 0.3 h: 0.1 d: 0.1 f: 0.7 + ?  0.7 + 0.1 a: 0.95 h: 0.35 + ?  0.35 + 0.5 b: 0.8 d: 0.35 + ?  0.35 + 0.5 c: 0.35 + ?  0.35 + 0.3 g: 0.2 + ?  0.2 + 0.4 h: 0.45 + ?  0.45 + 0.2 d: 0.35 + ?  0.35 + 0.3

14. 14/28 Evolution of a Candidate’s Score scan depth drop d from the candidate queue Approximate top-k “What is the probability that d qualifies for the top-k ?” bestscore(d) worstscore(d) min-k score ? Worst- and best-scores slowly converge to final score Add d to top-k result, if worstscore(d) > min-k Drop d only if bestscore(d) < min-k, otherwise keep it in candidate queue  Overly conservative threshold & long sequential index scans TA family of algorithms based on invariant (with sum as aggr) worstscore(d)bestscore(d)

15. 15/28 Top-k Queries with Probabilistic Guarantees TA family of algorithms based on invariant (with sum as aggr) Relaxed into probabilistic invariant where the RV S i has some (postulated and/or estimated) distribution in the interval (0,high i ] f: 0.5 b: 0.4 c: 0.35 a: 0.3 h: 0.1 d: 0.1 a: 0.55 b: 0.2 f: 0.2 g: 0.2 c: 0.1 h: 0.35 d: 0.35 b: 0.2 a: 0.1 c: 0.05 f: 0.05 S1S1 S2S2 S3S3 Discard candidates with p(d) ≤  Exit index scan when candidate list empty high i worstscore(d)bestscore(d)

16. 16/28 Probabilistic Threshold Test postulating uniform or Zipf score distribution in [0, high i ] compute convolution using LSTs use Chernoff-Hoeffding tail bounds or generalized bounds for correlated dimensions (Siegel 1995) fitting Poisson distribution (or Poisson mixture) over equidistant values: easy and exact convolution distribution approximated by histograms: precomputed for each dimension dynamic convolution at query-execution time with independent Si‘s or with correlated Si‘s engineering-wise histograms work best! 0 f 2 (x) 1 high 2 Convolution (f 2 (x), f 3 (x)) 2 0 δ(d) f 3 (x) high 3 1 0  cand doc d with 2  E(d), 3  E(d)

17. 17/28 Prob-sorted Algorithm (Smart Variant) Prob-sorted (RebuildPeriod r, QueueBound b):... scan all lists Li (i=1..m) in parallel: …same code as TA-sorted… // queue management for all priority queues q for which d is relevant do insert d into q with priority bestscore(d); // periodic clean-up if step-number mod r = 0 then // rebuild; single bounded queue if strategy = Smart then for all queue elements e in q do update bestscore(e) with current high_i values; rebuild bounded queue with best b elements; if prob[top(q) can qualify for top-k] <  then exit; if all queues are empty then exit;

18. 18/28 TA-sortedProb-sorted (smart) #sorted accesses2,263,652527,980 elapsed time [s]148.715.9 max queue size10849400 relative recall10.69 rank distance039.5 score error00.031 Performance Results for.Gov Queries on.GOV corpus from TREC-12 Web track: 1.25 Mio. docs (html, pdf, etc.) 50 keyword queries, e.g.: „Lewis Clark expedition“, „juvenile delinquency“, „legalization Marihuana“, „air bag safety reducing injuries death facts“ speedup by factor 10 at high precision/recall (relative to TA-sorted); aggressive queue mgt. even yields factor 100 at 30-50 % prec./recall

19. 19/28.Gov Expanded Queries on.GOV corpus with query expansion based on WordNet synonyms: 50 keyword queries, e.g.: „juvenile delinquency youth minor crime law jurisdiction offense prevention“, „legalization marijuana cannabis drug soft leaves plant smoked chewed euphoric abuse substance possession control pot grass dope weed smoke“ TA-sortedProb-sorted (smart) #sorted accesses22,403,49018,287,636 elapsed time [s]79081066 max queue size70896400 relative recall10.88 rank distance014.5 score error00.035

20. 20/28 Performance Results for IMDB Queries on IMDB corpus (Web site: Internet Movie Database): 375 000 movies, 1.2 Mio. persons (html/xml) 20 structured/text queries with Dice-coefficient-based similarities of categorical attributes Genre and Actor, e.g.: Genre  {Western}  Actor  {John Wayne, Katherine Hepburn}  Description  {sheriff, marshall}, Genre  {Thriller}  Actor  {Arnold Schwarzenegger}  Description  {robot} TA-sortedProb-sorted (smart) #sorted accesses1,003,650403,981 elapsed time [s]201.912.7 max queue size12628400 relative recall10.75 rank distance0126.7 score error00.25

21. 21/28 response time: 0.7 37: 0.9 44: 0.8... 22: 0.7 23: 0.6 51: 0.6 52: 0.6 throughput: 0.6 92: 0.9 67: 0.9... 52: 0.9 44: 0.8 55: 0.8 Handling Ontology-Based Query Expansions algorithm B+ tree index on terms 57: 0.6 44: 0.4... performance 52: 0.4 33: 0.3 75: 0.3 12: 0.9 14: 0.8... 28: 0.6 17: 0.55 61: 0.5 44: 0.5 44: 0.4 ontology / meta-index  i  q {max j  onto(i) { sim(i,j)*sj(d)) }} performance response time: 0.7 throughput: 0.6 queueing: 0.3 delay: 0.25... consider expandable query „algorithm and ~performance“ with score dynamic query expansion with incremental on-demand merging of additional index lists + much more efficient than threshold-based expansion + no threshold tuning + no topic drift

22. 22/28 Outline Motivation & Preliminaries Prob-k: Efficient Approximative Top-k TopX: Top-k for XML

23. 23/28 Top-k Search on XML TA-style algorithm should handle also tag-value conditions such as title = algorithm and ~topic = ~performance using standard indexes and exact path conditions of the form / book // literature using path indexes (pre/post index, HOPI, etc.) Example query (NEXI, XPath & IR): //book[about(.//„Information Retrieval“ „XML“) and.//affiliation[about(.// „Stanford“)] and.//reference[about(.// „Page Rank“)] ]//publisher//country Handling arbitrary combinations of tag-term conditions and path conditions TopX Algorithm

24. 24/28 Problems Addressed by TopX Problems: 1) content conditions (CC) on both tags and terms 2) scores for elements or subtrees, docs as results 3) score aggregation not necessarily monotonic 4) test path conditions (PC), but avoid random accesses Solutions: 0) disk space is cheap, disk I/O is not ! 1) build index lists for each tag-term pair 2) block-fetch all elements for the same doc in desc. order of MaxScore(e) = max{Score(e‘) | e‘  doc(e)} 3) precompute and store scores for entire subtrees 4a) test PCs on candidates in memory 4b) postpone evaluation of remaining PCs until after threshold test Example query (NEXI, XPath & IR): //book[about(.//„Information Retrieval“ „XML“) and.//affiliation[about(.// „Stanford“)] and.//reference[about(.// „Page Rank“)] ]//publisher//country

25. 25/28 Simplified Scoring Model for TopX 2:A 1:R 6:B 3:X7:X 4:B5:C aaccabab 8:B9:C bbbcccxy 2:X 1:A 6:B 3:B7:C 4:B5:C cccabb abcabc 2:B 1:Z 3:X 4:C5:A aaaabb 6:B8:X 7:C acc 9:B10:A bb11:C12:C aabbcxyz d1d2d3 restricted to tree data (disregarding links) using only tag-term conditions for scoring score (doc d or subtree rooted at d.n) for query q with tag-term conditions A 1 [a 1 ],..., A m [a m ] ) matched by nodes n 1,..., n m Example:

26. 26/28 TopX Pseudocode based on index table L (Tag, Term, MaxScore, DocId, Score, ElemId, Pre, Post) decompose query: content conditions (CC) & path conditions (PC); for each index list L i (extracted from L by tag and/or term) do: block-scan next elements from same doc d; E(d) := E(d)  {i}; for each CC l  E(d) – {i} do: for each element pair (e, e‘)  (elems(d,L i )  elems(d,L l )) do: test PC(i,l) connecting e and e‘ using pre & post of e, e‘; delete e  elems(d,L i ) if not  e‘ such that (e,e‘) satisifes PC(i,l); delete e‘  elems(d,L l ) if not  e such that (e,e‘) satisifes PC(i,l); bestscore(d) :=  j  E(d) max{Score(e) | e  elems(L j )} +  j  E(d) high j ; worstscore(d) := 0; if E(d) is complete then worstscore(d) := bestscore(d);... // proceed with standard top-k algorithm test remaining PCs on d and drop d if not satisifed;

27. 27/28 TopX Example Data Example query: //A [.//“a“ &.//B[.//“b“] &.//C[.//“c“] ] Tag Term MaxScore DocId Score ElemId Pre Post pre- computed index table with appropriate B+ tree index on (Tag, Term, MaxScore, DocId, Score, ElemId) 2:A 1:R 6:B 3:X7:X 4:B5:C aaccabab 8:B9:C bbbcccxy 2:X 1:A 6:B 3:B7:C 4:B5:C cccabb abcabc 2:B 1:Z 3:X 4:C5:A aaaabb 6:B8:X 7:C acc 9:B10:A bb11:C12:C aabbcxyz d1d2d3 block-scans: (A, a, d3,...) (B, b, d1,...) (C, c, d2,...) (A, a, d1,...) (B, b, d3,...) (C, c, d3,...) (A, a, d2,...) (B, b, d2,...) (C, c, d1,...) A a 1 d3 1 e5 5 2 A a 1 d3 1/4 e10 10 9 A a 1/2 d1 1/2 e2 2 4 A a 2/9 d2 2/9 e1 1 7 B b 1 d1 1 e8 8 5 B b 1 d1 1/2 e4 4 1 B b 1 d1 3/7 e6 6 8 B b 1 d3 1 e9 9 7 B b 1 d3 1/3 e2 2 4 B b 2/3 d2 2/3 e4 4 1 B b 2/3 d2 1/3 e3 3 3 B b 2/3 d2 1/3 e6 6 6 C c 1 d2 1 e5 5 2 C c 1 d2 1/3 e7 7 5 C c 2/3 d3 2/3 e7 7 5 C c 2/3 d3 1/5 e11 11 8 C c 3/5 d1 3/5 e9 9 6 C c 3/5 d1 1/2 e5 5 2

28. 28/28 Experimental Results: INEX Benchmark  (join&sort)TopX (  =0.0)TopX (  =0.1) #sorted accesses472,22770,6745,534 #random accesses226206 elapsed time [s]35.63.2 relative recall110.85 on IEEE-CS journal and conference articles: 12,000 XML docs with 12 Mio. elements,7.9 GB for all indexes 20 CO queries, e.g.: „XML editors or parsers“ 20 CAS queries, e.g.: //article[.//bibl[about(.//„QBIC“)] and.//p[about(.//„image retrieval“)] ] CO  (join&sort)TopX (  =0.0)TopX (  =0.1) #sorted accesses1,077,302280,249163,084 #random accesses2,0221,931 elapsed time [s]118.121.1 relative recall110.94 CAS

29. 29/28 INEX with Query Expansion static expansionincremental merge (  =0.8,  =0.1)(  =0.1) #sorted accesses471,128533,385 #random accesses4,228212 elapsed time [s]319.633.3 max #terms165 - 16 20 CO queries, e.g.: „XML editors or parsers“ 20 CAS queries, e.g.: //article[.//bibl[about(.//„QBIC“)] and.//p[about(.//„image retrieval“)] ] CO CAS static expansionincremental merge (  =0.8,  =0.1)(  =0.1) #sorted accesses814,1231,271,379 #random accesses12,755648 elapsed time [s]429.8103.0 max #terms117 - 11

30. 30/28 Conclusion: Ongoing and Future Work Observation: Approximations with statistical guarantees are key to obtaining Web-scale efficiency (e.g., TREC’04 Terabyte benchmark: ca. 25 Mio. docs, ca. 700 000 terms, 5-50 terms per query) Challenges: Generalize TopX to arbitrary graphs Efficient consideration of correlated dimensions Integrated support for all kinds of XML similarity search: content & ontological sim, structural sim Scheduling of index-scan steps and few random accesses Integration of top-k operator into physical algebra and query optimizer of XML engine