1. A Whirlwind Tour of Search Engine Design Issues CMSC 498W April 4, 2006 Douglas W. Oard

2. Information Access Problems Collection Information Need Stable Different Each Time Retrieval Filtering Different Each Time

3. Design Strategies Foster human-machine synergy –Exploit complementary strengths –Accommodate shared weaknesses Divide-and-conquer –Divide task into stages with well-defined interfaces –Continue dividing until problems are easily solved Co-design related components –Iterative process of joint optimization

4. Human-Machine Synergy Machines are good at: –Doing simple things accurately and quickly –Scaling to larger collections in sublinear time People are better at: –Accurately recognizing what they are looking for –Evaluating intangibles such as “quality” Both are pretty bad at: –Mapping consistently between words and concepts

5. Supporting the Search Process Source Selection Search Query Selection Ranked List Examination Document Delivery Document Query Formulation IR System Query Reformulation and Relevance Feedback Source Reselection NominateChoose Predict

6. Supporting the Search Process Source Selection Search Query Selection Ranked List Examination Document Delivery Document Query Formulation IR System Indexing Index Acquisition Collection

7. Taylor’s Model of Question Formation Q1 Visceral Need Q2 Conscious Need Q3 Formalized Need Q4 Compromised Need (Query) End-user Search Intermediated Search

8. Search Goal Choose the same documents a human would –Without human intervention (less work) –Faster than a human could (less time) –As accurately as possible (less accuracy) Humans start with an information need –Machines start with a query Humans match documents to information needs –Machines match document & query representations

9. Search Component Model Comparison Function Representation Function Query Formulation Human Judgment Representation Function Retrieval Status Value Utility Query Information NeedDocument Query RepresentationDocument Representation Query Processing Document Processing

10. Relevance Relevance relates a topic and a document –Duplicates are equally relevant, by definition –Constant over time and across users Pertinence relates a task and a document –Accounts for quality, complexity, language, … Utility relates a user and a document –Accounts for prior knowledge We seek utility, but relevance is what we get!

11. Problems With Word Matching Word matching suffers from two problems –Synonymy: paper vs. article –Homonymy: bank (river) vs. bank (financial) Disambiguation in IR: seek to resolve homonymy –Index word senses rather than words Synonymy usually addressed by –Thesaurus-based query expansion –Latent semantic indexing

12. Stemming Stem: in IR, a word equivalence class that preserves the main concept. –Often obtained by affix-stripping (Porter, 1980) –{destroy, destroyed, destruction}: destr Inexpensive to compute Usually helps IR performance Can make mistakes! (over-/understemming) –{centennial,century,center}: cent –{acquire,acquiring,acquired}: acquir {acquisition}: acquis

13. N-gram Indexing Consider a Chinese document c 1 c 2 c 3 … c n Don’t segment (you could be wrong!) Instead, treat every character bigram as a term – _ c 1 c 2, c 2 c 3, c 3 c 4, …, c n-1 c n Break up queries the same way

14. “Bag of Terms” Representation Bag = a “set” that can contain duplicates  “The quick brown fox jumped over the lazy dog’s back”  {back, brown, dog, fox, jump, lazy, over, quick, the, the} Vector = values recorded in any consistent order  {back, brown, dog, fox, jump, lazy, over, quick, the, the}  [1 1 1 1 1 1 1 1 2]

15. Bag of Terms Example The quick brown fox jumped over the lazy dog’s back. Document 1 Document 2 Now is the time for all good men to come to the aid of their party. the quick brown fox over lazy dog back now is time for all good men to come jump aid of their party 0 0 1 1 0 1 1 0 1 1 0 0 1 0 1 0 0 1 1 0 0 1 0 0 1 0 0 1 1 0 1 0 1 1 Term Document 1Document 2 Stopword List

16. Boolean IR Strong points –Accurate, if you know the right strategies –Efficient for the computer Weaknesses –Often results in too many documents, or none –Users must learn Boolean logic –Sometimes finds relationships that don’t exist –Words can have many meanings –Choosing the right words is sometimes hard

17. Proximity Operators More precise versions of AND –“NEAR n” allows at most n-1 intervening terms –“WITH” requires terms to be adjacent and in order Easy to implement, but less efficient –Store a list of positions for each word in each doc Stopwords become very important! –Perform normal Boolean computations Treat WITH and NEAR like AND with an extra constraint

18. Proximity Operator Example time AND come –Doc 2 time (NEAR 2) come –Empty quick (NEAR 2) fox –Doc 1 quick WITH fox –Empty quick brown fox over lazy dog back now time all good men come jump aid their party 01 (9) Term 1 (13) 1 (6) 1 (7) 1 (8) 1 (16) 1 (1) 1 (2) 1 (15) 1 (4) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 (5) 1 (9) 1 (3) 1 (4) 1 (8) 1 (6) 1 (10) Doc 1Doc 2

19. Advantages of Ranked Retrieval Closer to the way people think –Some documents are better than others Enriches browsing behavior –Decide how far down the list to go as you read it Allows more flexible queries –Long and short queries can produce useful results

20. Ranked Retrieval Challenges “Best first” is easy to say but hard to do! –The best we can hope for is to approximate it Will the user understand the process? –It is hard to use a tool that you don’t understand Efficiency becomes a concern –Only a problem for long queries, though

21. Similarity-Based Queries Treat the query as if it were a document –Create a query bag-of-words Find the similarity of each document –Using the coordination measure, for example Rank order the documents by similarity –Most similar to the query first Surprisingly, this works pretty well! –Especially for very short queries

22. Counting Terms Terms tell us about documents –If “rabbit” appears a lot, it may be about rabbits Documents tell us about terms –“the” is in every document -- not discriminating Documents are most likely described well by rare terms that occur in them frequently –Higher “term frequency” is stronger evidence –Low “collection frequency” makes it stronger still

23. The Document Length Effect Humans look for documents with useful parts –But probabilities are computed for the whole Document lengths vary in many collections –So probability calculations could be inconsistent Two strategies –Adjust probability estimates for document length –Divide the documents into equal “passages”

24. Incorporating Term Frequency High term frequency is evidence of meaning –And high IDF is evidence of term importance Recompute the bag-of-words –Compute TF * IDF for every element

25. TF*IDF Example 4 5 6 3 1 3 1 6 5 3 4 3 7 1 nuclear fallout siberia contaminated interesting complicated information retrieval 2 123 2 3 2 4 4 0.50 0.63 0.90 0.13 0.60 0.75 1.51 0.38 0.50 2.11 0.13 1.20 123 0.60 0.38 0.50 4 0.301 0.125 0.602 0.301 0.000 0.602 Unweighted query: contaminated retrieval Result: 2, 3, 1, 4 Weighted query: contaminated(3) retrieval(1) Result: 1, 3, 2, 4 IDF-weighted query: contaminated retrieval Result: 2, 3, 1, 4

26. Document Length Normalization Long documents have an unfair advantage –They use a lot of terms So they get more matches than short documents –And they use the same words repeatedly So they have much higher term frequencies Normalization seeks to remove these effects –Related somehow to maximum term frequency –But also sensitive to the of number of terms

27. “Okapi” Term Weights TF componentIDF component

28. Passage Retrieval Another approach to long-document problem –Break it up into coherent units Recognizing topic boundaries is hard –But overlapping 300 word passages work fine Document rank is best passage rank –And passage information can help guide browsing

29. Summary Goal: find documents most similar to the query Compute normalized document term weights –Some combination of TF, DF, and Length Optionally, get query term weights from the user –Estimate of term importance Compute inner product of query and doc vectors –Multiply corresponding elements and then add

30. Another Perspective We ask “is this document relevant?” –Vector space: we answer “somewhat” –Probabilistic: we answer “probably”

31. Probability Ranking Principle Assume binary relevance/document independence –Each document is either relevant or it is not –Relevance of one doc reveals nothing about another Assume the searcher works down a ranked list –Seeking some number of relevant documents Theorem (provable from assumptions): –Documents should be ranked in order of decreasing probability of relevance to the query, P(d relevant-to q)

32. Some Questions for Indexing How long will it take to find a document? –Is there any work we can do in advance? If so, how long will that take? How big a computer will I need? –How much disk space? How much RAM? What if more documents arrive? –How much of the advance work must be repeated? –Will searching become slower? –How much more disk space will be needed?

33. The Indexing Process quick brown fox over lazy dog back now time all good men come jump aid their party 0 0 1 1 0 0 0 0 0 1 0 0 1 0 1 1 0 0 1 0 0 1 0 0 1 0 0 1 1 0 0 0 0 1 Term Doc 1Doc 2 0 0 1 1 0 1 1 0 1 1 0 0 1 0 1 0 0 1 1 0 0 1 0 0 1 0 0 1 0 0 0 0 0 1 Doc 3 Doc 4 0 0 0 1 0 1 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 0 1 Doc 5Doc 6 0 0 1 1 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 1 0 0 1 0 0 1 1 1 1 0 0 0 Doc 7Doc 8 A B C F D G J L M N O P Q T AI AL BA BR TH TI 4, 8 2, 4, 6 1, 3, 7 1, 3, 5, 7 2, 4, 6, 8 3, 5 3, 5, 7 2, 4, 6, 8 3 1, 3, 5, 7 2, 4, 8 2, 6, 8 1, 3, 5, 7, 8 6, 8 1, 3 1, 5, 7 2, 4, 6 Postings Inverted File

34. The Finished Product quick brown fox over lazy dog back now time all good men come jump aid their party Term A B C F D G J L M N O P Q T AI AL BA BR TH TI 4, 8 2, 4, 6 1, 3, 7 1, 3, 5, 7 2, 4, 6, 8 3, 5 3, 5, 7 2, 4, 6, 8 3 1, 3, 5, 7 2, 4, 8 2, 6, 8 1, 3, 5, 7, 8 6, 8 1, 3 1, 5, 7 2, 4, 6 PostingsInverted File

35. How Big Is the Postings File? Very compact for Boolean retrieval –About 10% of the size of the documents If an aggressive stopword list is used! Not much larger for ranked retrieval –Perhaps 20% Enormous for proximity operators –Sometimes larger than the documents!

36. Building an Inverted Index Simplest solution is a single sorted array –Fast lookup using binary search –But sorting large files on disk is very slow –And adding one document means starting over Tree structures allow easy insertion –But the worst case lookup time is linear Balanced trees provide the best of both –Fast lookup and easy insertion –But they require 45% more disk space

37. How Big is the Inverted Index? Typically smaller than the postings file –Depends on number of terms, not documents Eventually, most terms will already be indexed –But the postings file will continue to grow Postings dominate asymptotic space complexity –Linear in the number of documents

38. Index Compression CPU’s are much faster than disks –A disk can transfer 1,000 bytes in ~20 ms –The CPU can do ~10 million instructions in that time Compressing the postings file is a big win –Trade decompression time for fewer disk reads Key idea: reduce redundancy –Trick 1: store relative offsets (some will be the same) –Trick 2: use an optimal coding scheme

39. Summary Slow indexing yields fast query processing –Key fact: most terms don’t appear in most documents We use extra disk space to save query time –Index space is in addition to document space –Time and space complexity must be balanced Disk block reads are the critical resource –This makes index compression a big win

40. Problems with “Free Text” Search Homonymy –Terms may have many unrelated meanings –Polysemy (related meanings) is less of a problem Synonymy –Many ways of saying (nearly) the same thing Anaphora –Alternate ways of referring to the same thing

41. Two Ways of Searching Write the document using terms to convey meaning Author Content-Based Query-Document Matching Document Terms Query Terms Construct query from terms that may appear in documents Free-Text Searcher Retrieval Status Value Construct query from available concept descriptors Controlled Vocabulary Searcher Choose appropriate concept descriptors Indexer Metadata-Based Query-Document Matching Query Descriptors Document Descriptors

42. Applications When implied concepts must be captured –Political action, volunteerism, … When terminology selection is impractical –Searching foreign language materials When no words are present –Photos w/o captions, videos w/o transcripts, … When user needs are easily anticipated –Weather reports, yellow pages, …

43. Supervised Learning f 1 f 2 f 3 f 4 … f N v 1 v 2 v 3 v 4 … v N CvCv w 1 w 2 w 3 w 4 … w N CwCw LearnerClassifier New example x 1 x 2 x 3 x 4 … x N CxCx Labelled training examples CwCw

44. Example: kNN Classifier

45. Challenges Changing concept inventories –Literary warrant and user needs are hard to predict Accurate concept indexing is expensive –Machines are inaccurate, humans are inconsistent Users and indexers may think differently –Diverse user populations add to the complexity Using thesauri effectively requires training –Meta-knowledge and thesaurus-specific expertise

46. Relevance Feedback Make Profile Vector Compute Similarity Select and Examine (user) Assign Ratings (user) Update User Model New Documents Vector Documents, Vectors, Rank Order Document, Vector Rating, Vector Vector(s) Make Document Vectors Initial Profile Terms Vectors

47. Rocchio Formula 040800 124001 201104 6370-3 040800 248002 8044016 Original profile Positive Feedback Negative feedback (+) (-) New profile

48. Implicit Feedback Observe user behavior to infer a set of ratings –Examine (reading time, scrolling behavior, …) –Retain (bookmark, save, save & annotate, print, …) –Refer to (reply, forward, include link, cut & paste, …) Some measurements are directly useful –e.g., use reading time to predict reading time Others require some inference –Should you treat cut & paste as an endorsement?

49. Some Observable Behaviors

50. Behavior Category

51. Minimum Scope

52. Some Examples Read/Ignored, Saved/Deleted, Replied to (Stevens, 1993) Reading time (Morita & Shinoda, 1994; Konstan et al., 1997) Hypertext Link (Brin & Page, 1998)

53. Estimating Authority from Links Authority Hub

54. Collecting Click Streams Browsing histories are easily captured –Make all links initially point to a central site Encode the desired URL as a parameter –Build a time-annotated transition graph for each user Cookies identify users (when they use the same machine) –Redirect the browser to the desired page Reading time is correlated with interest –Can be used to build individual profiles –Used to target advertising by doubleclick.com

55. 0 20 40 60 80 100 120 140 160 180 No Interest Low Interest Moderate Interest High Interest Rating Reading Time (seconds) Full Text Articles (Telecommunications) 50 32 58 43

56. Critical Issues Protecting privacy –What absolute assurances can we provide? –How can we make remaining risks understood? Scalable rating servers –Is a fully distributed architecture practical? Non-cooperative users –How can the effect of spamming be limited?

57. Gaining Access to Observations Observe public behavior –Hypertext linking, publication, citing, … Policy protection –EU: Privacy laws –US: Privacy policies + FTC enforcement Architectural assurance of privacy –Distributed architecture –Model and mitigate privacy risks

58. Adversarial IR Search is user-controlled suppression –Everything is known to the search system –Goal: avoid showing things the user doesn’t want Other stakeholders have different goals –Authors risk little by wasting your time –Marketers hope for serendipitous interest Metadata from trusted sources is more reliable

59. Index Spam Goal: Manipulate rankings of an IR system Multiple strategies: –Create bogus user-assigned metadata –Add invisible text (font in background color, …) –Alter your text to include desired query terms –“Link exchanges” create links to your page

60. Putting It All Together Free TextBehaviorMetadata Topicality Quality Reliability Cost Flexibility

61. Evaluation Criteria Effectiveness –System-only, human+system Efficiency –Retrieval time, indexing time, index size Usability –Learnability, novice use, expert use

62. IR Effectiveness Evaluation User-centered strategy –Given several users, and at least 2 retrieval systems –Have each user try the same task on both systems –Measure which system works the “best” System-centered strategy –Given documents, queries, and relevance judgments –Try several variations on the retrieval system –Measure which ranks more good docs near the top

63. Good Measures of Effectiveness Capture some aspect of what the user wants Have predictive value for other situations –Different queries, different document collection Easily replicated by other researchers Easily compared –Optimally, expressed as a single number

64. Comparing Alternative Approaches Achieve a meaningful improvement –An application-specific judgment call Achieve reliable improvement in unseen cases –Can be verified using statistical tests

65. Evolution of Evaluation Evaluation by inspection of examples Evaluation by demonstration Evaluation by improvised demonstration Evaluation on data using a figure of merit Evaluation on test data Evaluation on common test data Evaluation on common, unseen test data

66. Which is the Best Rank Order? R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R R abcd ef gh

67. IR Test Collection Design Representative document collection –Size, sources, genre, topics, … “Random” sample of representative queries –Built somehow from “formalized” topic statements Known binary relevance –For each topic-document pair (topic, not query!) –Assessed by humans, used only for evaluation Measure of effectiveness –Used to compare alternate systems

68. Set-Based Effectiveness Measures Precision –How much of what was found is relevant? Often of interest, particularly for interactive searching Recall –How much of what is relevant was found? Particularly important for law, patents, and medicine Fallout –How much of what was irrelevant was rejected? Useful when different size collections are compared

69. Effectiveness Measures Relevant Retrieved False AlarmIrrelevant Rejected Miss Relevant Not relevant RetrievedNot Retrieved Doc Action User- Oriented System- Oriented

70. Single-Figure Set-Based Measures Balanced F-measure –Harmonic mean of recall and precision –Weakness: What if no relevant documents exist? Cost function –Reward relevant retrieved, Penalize non-relevant For example, 3R + - 2N + –Weakness: Hard to normalize, so hard to average

71. Ranked Retrieval Query Ranked List Documents Abstract Evaluation Model Evaluation Measure of Effectiveness Relevance Judgments

72. Uninterpolated MAP R R 1/3=0.33 2/7=0.29 AP=0.31 R R R R 1/1=1.00 2/2=1.00 3/5=0.60 4/9=0.44 AP=0.76 R R R 1/2=0.50 3/5=0.60 4/8=0.50 AP=0.53 MAP=0.53

73. Why Mean Average Precision? It is easy to trade between recall and precision –Adding related query terms improves recall But naive query expansion techniques kill precision –Limiting matches by part-of-speech helps precision But it almost always hurts recall Comparisons should give some weight to both –Average precision is a principled way to do this –More “central” than other available measures

74. What MAP Hides Adapted from a presentation by Ellen Voorhees at the University of Maryland, March 29, 1999

75. Hypothesis Testing Paradigm Choose some experimental variable X Establish a null hypothesis H 0 Identify distribution of X, assuming H 0 is true Do experiment to get an empirical value x Ask: “What’s the probability I would have gotten this value of x if H 0 were true?” If this probability p is low, reject H 0

76. Statistical Significance Testing System A 0.02 0.39 0.16 0.58 0.04 0.09 0.12 System B 0.76 0.07 0.37 0.21 0.02 0.91 0.46 Query 12345671234567 Average0.200.40 Sign Test +-+--+-+-+--+- p=1.0 Wilcoxon +0.74 - 0.32 +0.21 - 0.37 - 0.02 +0.82 - 0.38 p=0.9375 0 95% of outcomes Try some out at: http://www.fon.hum.uva.nl/Service/Statistics.html

77. How Much is Enough? Measuring improvement –Achieve a meaningful improvement Guideline: 0.05 is noticeable, 0.1 makes a difference –Achieve reliable improvement on “typical” queries Wilcoxon signed rank test for paired samples Know when to stop! –Inter-assessor agreement limits max precision Using one judge to assess the other yields about 0.8

78. Pooled Assessment Methodology Systems submit top 1000 documents per topic Top 100 documents for each are judged –Single pool, without duplicates, arbitrary order –Judged by the person that wrote the query Treat unevaluated documents as not relevant Compute MAP down to 1000 documents –Treat precision for complete misses as 0.0

79. Alternative Ways to Get Judgments Exhaustive assessment is usually impractical –Topics * documents = a large number! Pooled assessment leverages cooperative evaluation –Requires a diverse set of IR systems Search-guided assessment is sometimes viable –Iterate between topic research/search/assessment –Augment with review, adjudication, reassessment Known-item judgments have the lowest cost –Tailor queries to retrieve a single known document –Useful as a first cut to see if a new technique is viable

80. Ten Commandments of Evaluation Thou shalt define insightful evaluation metrics Thou shalt define replicable evaluation metrics Thou shalt report all relevant system parameters Thou shalt establish upper bounds on performance Thou shalt establish lower bounds on performance Thou shalt test differences for statistical significance Thou shalt say whether differences are meaningful Thou shalt not mingle training data with test data

81. Automatic Evaluation Search Query Ranked List

82. User Studies Search Query Selection Ranked List Examination Document Query Formulation Query Reformulation and Relevance Feedback

83. User Studies Goal is to account for interface issues –By studying the interface component –By studying the complete system Formative evaluation –Provide a basis for system development Summative evaluation –Designed to assess performance

84. Quantitative User Studies Select independent variable(s) –e.g., what info to display in selection interface Select dependent variable(s) –e.g., time to find a known relevant document Run subjects in different orders –Average out learning and fatigue effects Compute statistical significance –Null hypothesis: independent variable has no effect –Rejected if p<0.05

85. Qualitative User Studies Observe user behavior –Instrumented software, eye trackers, etc. –Face and keyboard cameras –Think-aloud protocols –Interviews and focus groups Organize the data –For example, group it into overlapping categories Look for patterns and themes Develop a “grounded theory”

86. Questionnaires Demographic data –For example, computer experience –Basis for interpreting results Subjective self-assessment –Which did they think was more effective? –Often at variance with objective results! Preference –Which interface did they prefer? Why?

87. Summary Qualitative user studies suggest what to build Design decomposes task into components Automated evaluation helps to refine components Quantitative user studies show how well it works

88. Source: Global Reach English 2000 2005 Global Internet User Population Chinese