1. Chapter 7 Retrieval Models 1

2.  Provide a mathematical framework for defining the search process  Includes explanation of assumptions  Basis of many ranking algorithms  Progress in retrieval models has corresponded with improvements in effectiveness  Theories about relevance 2

3. Relevance  Complex concept that has been studied for some time  Many factors to consider  People often disagree when making relevance judgments  Retrieval models make various assumptions about relevance to simplify problem  e.g., topical vs. user relevance  e.g., binary vs. multi-valued relevance 3

4. Retrieval Model Overview  Older models  Boolean retrieval  Vector Space model  Probabilistic Models  BM25  Language models  Combining evidence  Inference networks  Learning to Rank 4

5. Searching Based on Retrieved Documents  Sequence of queries driven by the number of retrieved documents Example. “Lincoln” search of news articles  president AND Lincoln  president AND Lincoln AND NOT (automobile OR car)  president AND Lincoln AND biography AND life AND birthplace AND gettysburg AND NOT (automobile OR car)  president AND Lincoln AND (biography OR life OR birthplace OR gettysburg) AND NOT (automobile OR car) 5

6. Boolean Retrieval  Two possible outcomes for query processing  TRUE and FALSE  “Exact-match” retrieval  No ranking at all  Query usually specified using Boolean operators  AND, OR, NOT 6

7. Boolean Retrieval  Advantages  Results are predictable, relatively easy to explain  Many different document features, such as metadata (e.g., type/date), can be incorporated  Efficient processing, since many documents can be eliminated from search  Disadvantages  Simple queries usually don’t work well  Complex queries are difficult to construct  Effectiveness depends entirely on users 7

8. Vector Space Model  Simple and intuitively appealing  Documents and query represented by a vector of term weights  Collection represented by a matrix of term weights 8

9. Vector Space Model 9 Vector Representation of Stemmed Documents w/o Stopwords

10. Vector Space Model  3-D pictures useful, but can be misleading for high- dimensional space 10

11. Vector Space Model  Documents ranked by differences between points representing query and documents  Using similarity measure rather than a distance or dissimilarity measure  e.g. Cosine correlation 11

12. Similarity Calculation  Consider two documents D 1 and D 2, and a query Q  D 1 = (0.5, 0.8, 0.3)  D 2 = (0.9, 0.4, 0.2) 12 Q = (1.5, 1.0, 0) More similar to Q than D 1

13. Term Weights  TF-IDF Weight  Term frequency weight measures importance in document:  Inverse document frequency measures importance in collection:  Some heuristic modifications 13 t t Ensure non-zero weight

14. Vector Space Model  Advantages  Simple computational framework for ranking  Any similarity measure or term weighting scheme could be used  Disadvantages  Assumption of term independence  No assumption on whether relevance is binary/multivalued 14

15. Probabilistic Models  According to [Greiff 02]  In probabilistic approaches to IR, the occurrence of a query term in a document D contributes to the probability that D will be judged as relevant  The weight assigned to a query term should be based on the expected value of that contribution 15 [Greiff 02] Greiff, et al. The Role of Variance in Term Weighting for Probabilistic Information Retrieval. In Proc. of Intl. Conf. on Information and Knowledge Management (CIKM). 2002.

16. IR as Classification 16 (D) Bayes Decision Rule Based on Bayes Classifier Conditional Probability

17. Bayes Classifier  Bayes Decision Rule  A document D is relevant if P(R | D) > P(NR | D), where P(R | D) and P(NR | D) are conditional probabilities  Estimating probabilities  Use Bayes Rule  Classify a document as relevant if L.H.S. is the likelihood ratio of D being relevant 17 based on the probability of occurrence of words in D that are in R A prior probability of relevance Based on the Bayes Decision Rule P(R | D) > P(NR | D)

18. Estimating P ( D | R )  Assume word independence (Naïve Bayes assumption) and use individual term/word (d i ) probabilities  Binary (weights in doc) independence (of word) model  Document represented by a vector of binary features indicating term occurrence (or non-occurrence)  p i is probability that term i occurs in relevant document, s i is probability of occurrence of i in non-relevant document  Example. Given a document representation (1, 0, 0, 1, 1), the probability of the document occurring in the relevant set is p 1 × (1 - p 2 ) × (1 - p 3 ) × p 4 × p 5 18

19. Binary Independence Model 19 word i in D (Same for all documents) Probability of i in R Probability of i not in R word i not in D Probability of i in NR Probability of i not in NR  Computing the likelihood ratio of D using p i and s i 1

20. Binary Independence Model  Scoring function is  Using log to avoid multiplying lots of small numbers  If there is no other information on the relevant set, i.e., neither R nor NR exists, then  p i is set to be a constant (= 0.5)  s i is approximated as n i / N, where n i is the number of documents in the collection N that include i  s i is similar to the IDF-like weight 20

21. Contingency Table, If (non-)Relevant Sets available Gives the scoring function: 21 where r i = number of relevant documents containing term i n i = number of documents in a collection N containing i N = number of documents in the collection R = number of relevant documents in the collection n 

22. BM25 Ranking Algorithm  A ranking algorithm based on probabilistic arguments and experimental validation, but it is not a formal model  Adds document D and query Q term weights  f i is the frequency of term i in a document D  k 1, a constant, plays the role of tf along with f i  qf i is the frequency of term i in query Q, much lower than f i  k 2 plays the role as k 1 but is applied to query term weight qf i , where dl is document length & avdl is the average length of a doc regulated by b (0 ≤ b ≤ 1)  k 1, k 2 & K are set empirically as 1.2, 0..1000, & 0.75, resp. 22

23. BM25 Example  Given a query with two terms, “president lincoln”, (qf = 1)  No relevance information (r i and R are zero)  N = 500,000 documents  “president” occurs in 40,000 documents (n 1 = 40,000)  “lincoln” occurs in 300 documents (n 2 = 300)  “president” occurs 15 times in doc D (f 1 = 15)  “lincoln” occurs 25 times in doc D (f 2 = 25)  document length is 90% of the average length (dl / avdl = 0.9)  k 1 = 1.2, b = 0.75, and k 2 = 100  K = 1.2 × (0.25 + 0.75 × 0.9) = 1.11 23

24. BM25 Example 24

25. BM25 Example  Effect of term frequencies, especially on “lincoln” 25 The large occurrence of a single important query term can yield a higher score than the occurrence of both query terms

26. Language Model (LM)  A language model is associated with each doc & has been successfully applied to search applications to represent the topical content of a document & query  LM is a probability distribution over sequences of words to represent docs, i.e., puts a probability measure over strings from some vocabulary,  s  * P(s) = 1  The original and basic method for using LMs in IR is the query likelihood model in which a document d is constructed from a language model M d, i.e.,  Estimate a LM for each document d  Rank documents by the likelihood of the query according to the estimated LM, i.e., P(Q | M d ) 26

27. LMs for Retrieval  Three possibilities models of topical relevance  Probability of generating the query text from a document (language) model, i.e., the query likelihood model P mle (Q | M d ) =  t  Q P mle (t | M d ), where P mle (t | M d ) =  Probability of generating the document text from a query (language) model, i.e., P mle (D | M q ) = P mle (D | R), where R is the set of relevant documents of a query q  Comparing the language models representing the query and document topics 27 tf t,d dl d

28. LMs for Retrieval  Three ways of developing the LM modeling approach: a) The Query Likelihood Model: P(Q | M d ) or P(Q | D) b) The Document Likelihood Model: P(D | M Q ) or P(D | R) c) Comparison of the query and document LMs using KL- Divergence 28

29. LMs for Retrieval  The Query Likelihood Model  Documents are ranked based on the probability that the query Q could be generated by the document model (i.e., on the same topic) based on doc D, i.e., P(Q | D)  Given a query Q = q 1 …q n and a document D = d 1 …d m, compute the conditional probability P(D | Q) P(D | Q) = P(Q | D) P(D) where P(Q | D) P(D) is the query likelihood given D & P(D) is the prior belief that D is relevant to any query 29 A prior probability of document D (Assumed to be uniform)

30. Estimating Probabilities  Obvious estimate for unigram probabilities is  Maximum likelihood estimate (mle)  Makes the observed value of f q i, D most likely  If any query words are missing from document, score will be zero  Missing 1 out of 4 query words same as missing 3 out of 4 30

31. LMs for Retrieval  Example. The Query Likelihood Model  Given the two (unigram) LMs, M 1 and M 2, as shown below 31 Model M 1 Model M 2 the0.2the0.15 a0.1a0.12 frog0.01frog0.0002 toad0.01toad0.0001 said0.03said0.03 likes0.02likes0.04 that0.04that0.04 dog0.005dog0.01 cat0.003cat0.015 monkey0.001Monkey0.002 ………… Q: Frog said that toad likes that dog M 1 0.01 0.03 0.04 0.01 0.02 0.04 0.005 M 2 0.0002 0.03 0.04 0.0001 0.04 0.04 0.01 P(Q | M 1 ) = 0.00000000000048 P(Q | M 2 ) = 0.00000000000000034 P(Q | M 1 ) > P(Q | M 2 )

32. The Language Models  Example. Using language models, we estimate if “Barbie car” (i.e., treated as Q) is generated from content (i.e., treated as M) targeting younger or adults  Based on the results, it is more likely that “Barbie car” is generated from content written for younger audiences 32 P( | “Barbie car”) = P(“Barbie car” | ) = 0.0056  0.0039 = 0.00002184 P( | “Barbie car”) = P(“Barbie car” | ) = 0.0048  0.0002 = 0.00000096 WordsProbability on Simple Wiki Probabilit on Wikipedia Blue0.00420.0043 Car0.00390.0048 Barbie0.00560.0002 Auburn0.000000010.0439 Software0.00030.0756 Toy0.00760.0037

33. LMs for Retrieval  The Document Likelihood Model  Instead of using the probability of a document language model (M d ) generating the query, use the probability of a query language model M q generating the document  Creating a document likelihood model, even though it is less appealing, since there is much less text available to estimate a LM based on the query text  Solution: incorporate relevance feedback into the model & update the language model M q e.g., the relevance model of Lavrenko & Croft (2001) is an instance of a document likelihood model that incorporates pseudo relevance feedback into an LM approach 33

34. Language Model (LM) - Example  Consider the query Q = “Frozen Movie” & the following documents, where R is relevant & NR is non-relevant 34 Frozen Character Anna Frozen Food from Fridge Disney Movie Elsa Frozen Disney Sven R NR RR Words (R)P(w) Frozen2/9 Character1/9 Anna1/9 Disney2/9 Movie1/9 Elsa1/9 Sven1/9

35.  Using the following query language model (M q ) we estimate the probability of docs D 1 and D 2 on their relevance for query Q = “Frozen Movie”  D 1 : “Sven, Anna, Disney, Frozen”  P(D 1 | M q ) =  t  D 1 P(t | M q ) = 1/9  1/9  2/9  2/9 = 4/6561 = 0.00061  D 2 : “Frozen ice cream”  P(D 2 | M q ) =  t  D 2 P(t | M q ) = 2/9  0  0 = 0 35 Language Model (LM) - Example Words (R)P(w) Frozen2/9 Character1/9 Anna1/9 Disney2/9 Movie1/9 Elsa1/9 Sven1/9

36. LMs for Retrieval  Comparing the Query Likelihood Model & Document Likelihood Model  Rather than generating either a document language model (M d ) or a query language model (M q ), create an LM from both the document & query & find the difference  To model the risk of returning a document d as relevant to a query q, use the KL-divergence R(d; q) = KL(M d || M q ) =  t  V P(t | M q ) log KL-divergence measures how bad the probability distribution M q is at modeling M d It has been shown that the comparison model outperforms both query-likelihood & document-likelihood models and useful for ad hoc retrieval 36 P(t | M q ) P(t | M d )

37. KL-Divergence  Example. Consider the probability distributions generated using texts from Simple Wikipedia (a simplified version of Wikipedia) and Wikipedia  Probabilities generated using text targeting younger versus more mature audiences 37 WordsProbability based on Simple Wiki Probability based on Wikipedia Blue0.00420.0043 Car0.00390.0048 Barbie0.00560.0002 Auburn0.000000010.0439 Software0.00030.0756 Toy0.00760.0037

38. KL-Divergence  Example (Cont.). Using K-L Divergence on vocabulary (0.0043  log (0.0043/0.0042) + 0.0048  log (0.0048/0.0039) + 0.0002  log (0.002/0.0056) + 0.0439  log (0.0439/0.00000001) + 0.0756  log (0.0756/0.0003) + 0.0037  log (0.0037/0.0076)) = 0.47218  The vocabulary distribution for children versus more mature audiences is indeed different 38 || ) = K-L Divergence (

39. Smoothing  Document texts are a sample from the language model  Missing words (in a document) should not have zero probability of occurring  Smoothing is a technique for estimating probabilities for missing (or unseen) words for the words that are not seen in the text  Lower (or discount) the probability estimates for words that are seen in the document text  Assign that “left-over” probability to the estimates 39

40. Smoothing  As stated in [Zhai 04]  Smoothing is the problem of adjusting the maximum likelihood estimator to compensate for data sparseness and it is required to avoid assigning zero probability to unseen words  Smoothing accuracy is directly related to the retrieval performance: The retrieval performance is generally sensitive to smoothing parameters  Smoothing plays two different roles: Making the estimated document LM more accurate Explaining the non-informative words in the query 40

41. Estimating Probabilities  Estimate for unseen words is α D P(q i | C)  P(q i | C) is the probability for query word i in the collection language model for collection C (background probability)  α D, which can be a constant, is a coefficient of the probability assigned to unseen words, depending on the document D  Estimate for words that occur is (1 − α D ) P(q i | D) + α D P(q i | C)  Different forms of estimation come from different α D 41

42. Estimating Probabilities  Different forms of estimation come from different α D (or ) 42 where p s (w | d) is the smoothed probability of a word seen in d, p(w | C) is the collection language model, α d is a coefficient controlling the probability mass assigned to unseen words, all probabilities sum to one

43. Jelinek-Mercer Smoothing  α D is a constant, λ (= 0.1 for short queries or = 0.7 for long queries in TREC evaluations  Gives estimate of  Ranking score  Use logs for convenience: solve the accuracy problem of multiplying small numbers 43

44. Where is tf-idf Weight?  Proportional to the term frequency (TF), inversely proportional to the collection frequency (IDF) 44 i: f q i, D > 0  i: f q i, D = 0  i: 1..n TF IDF Same for all documents

45. Dirichlet Smoothing  α D depends on document length where  is a parameter value set empirically  Gives probability estimation of  and document score 45

46. Query Likelihood  Example. Given,  Let Q be “President Lincoln”, D (a doc) in C(ollection)  For the term “President”, let f q i,D = 15, c q i = 160,000  For the term “Lincoln”, let f q i,D = 25, c q i = 2,400  Let no. of word occurrences in D (i.e., |D|) be 1,800  Let no. of word occurrences in C be 10 9 = 500,000 (docs) × 2,000 (average no. of words in a doc)  μ = 2,000 46 Negative due to the summating logs of small numbers

47. 47 Query Likelihood

48. 48 Set Theoretic Models  The Boolean model imposes a binary criterion for deciding relevance  The question of how to extend the Boolean model to accomodate partial matching and a ranking has attracted considerable attention in the past  Two set theoretic models for this extension are  Fuzzy Set Model  Extended Boolean Model

49. 49 Fuzzy Queries  Binary queries are based on binary logic, in which a document either satisfies or does not satisfy a query.  Fuzzy queries are based on fuzzy logic, in which a query term is considered as a fuzzy set & each document is given a degree of relevance with respect to a query, usually [0 – 1].  Example.  In binary logic: The set of "tall people" are defined as those > 6 feet. Thus, a person 5' 9.995" is not considered tall.  In fuzzy logic: no clear distinction on “tall people.” Every person is given a degree of membership to the set of tall people, e.g., a person 7'0" will have a grade 1.0 a person 4'5" will have a grade 0.0 a person 6‘2" will have a grade 0.85 a person 5’8" will have a grade 0.7

50. 50 Fuzzy Queries  Binary Logic vs. Fuzzy Logic Binary (crisp) Logic Fuzzy Logic 6.5  TALL 0 0 1 1

51. 51 Fuzzy Set Model  Queries and documents are represented by sets of index terms & matching is approximate from the start  This vagueness can be modeled using a fuzzy framework, as follows:  with each (query) term is associated a fuzzy set  each document has a degree of membership (0  X  1) in this fuzzy set  This interpretation provides the foundation for many models for IR based on fuzzy theory  The model proposed by Ogawa, Morita & Kobayashi (1991) is discussed

52. 52 Fuzzy Information Retrieval  Fuzzy sets are modeled based on a thesaurus  This thesaurus is built as follows:  Let c be a term-term correlation matrix (or keyword connection matrix)  Let c(i, l) be a normalized correlation factor for (k i, k l ): n i : number of documents which contain k i n l : number of documents which contain k l n(i, l): number of documents which contain both k i and k l  The notion of proximity among indexed terms c(i, l) = n(i, I) n i + n l - n(i, l)

53. 53 Fuzzy Information Retrieval  The correlation factor c(i, l) can be used to define fuzzy set membership for a document d j in thefuzzy set for query term K i as follows:  (i, j) = 1 -  (1 - c(i, l)) k l  d j  (i, j): the degree of membership of document d j in fuzzy subset associated with k i  The above expression computes an algebraic sum over all terms in d j with respect to k i, and is implemented as the complement of a negated algebraic product  A document d j belongs to the fuzzy set for k i, if its own terms are associated with k i

54. 54 Fuzzy Information Retrieval   (i, j) = 1 -  (1 - c(i, l)) k l  d j   (i, j): membership of document d j in fuzzy subset associated with query term k i  If document d j contains any index term k l which is closely related to query term k i, we have  if  l such that c(i, l) ~ 1, then  (i, j) ~ 1, and  query index k i is a good fuzzy index for document d j

55. 55 Fuzzy IR  q = k a  (k b   k c )  q dnf = (1,1,1) + (1,1,0) + (1,0,0) (binary weighted) = cc 1 + cc 2 + cc 3 (conjunctive comp)   (q, d j ) =  (cc 1 + cc 2 + cc 3, j) = 1 - (1 -  (a, j)  (b, j)  (c, j))  (1 -  (a, j)  (b, j) (1 -  (c, j)))  (1 -  (a, j) (1 -  (b, j)) (1 -  (c, j))) Example. cc 1 cc 3 cc 2 KaKa KbKb KcKc

56. 56 Fuzzy Information Retrieval  Fuzzy set are useful for representing vagueness and imprecision  Fuzzy IR models have been discussed mainly in the literature associated with fuzzy theory  Experiments with standard test collections are not available  Difficult to compare because previously conducted experiences used only small collections