Statistical Models for Information Retrieval and Text Mining ChengXiang Zhai (翟成祥) Department of Computer Science Graduate School of Library & Information Science Institute for Genomic Biology, Statistics University of Illinois, Urbana-Champaign http://www-faculty.cs.uiuc.edu/~czhai, czhai@cs.uiuc.edu 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 1 Course Overview Scope of the course Information Retrieval Multimedia Data Computer Vision Text Data Natural Language Processing Machine Learning Statistics 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 2 Goal of the Course • • Overview of techniques for information retrieval (IR) Detailed explanation of a few statistical models for IR and text mining – Probabilistic retrieval models (for search) – Probabilistic topic models (for text mining) • Potential benefit for you: – Some ideas working well for text retrieval may also work for computer vision – Techniques for computer vision may be applicable to IR – IR and text mining raise new challenges as well as opportunities for machine learning 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 3 Course Plan • Lecture 1: Overview of information retrieval • Lecture 2: Statistical language models for IR: Part 1 • Lecture 3: Statistical language models for IR: Part 2 • Lecture 4: Formal retrieval frameworks • Lecture 5: Probabilistic topic models for text mining 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 4 Lecture 1: Overview of IR • Basic Concepts in Text Retrieval (TR) • Evaluation of TR • Common Components of a TR system • Overview of Retrieval Models 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 5 Basic Concepts in TR 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 6 What is Text Retrieval (TR)? • There exists a collection of text documents • User gives a query to express the information need • A retrieval system returns relevant documents to users • Known as “search technology” in industry 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 7 • History of TR on One Slide Birth of TR – 1945: V. Bush’s article “As we may think” – 1957: H. P. Luhn’s idea of word counting and matching • Indexing & Evaluation Methodology (1960’s) – Smart system (G. Salton’s group) – Cranfield test collection (C. Cleverdon’s group) – Indexing: automatic can be as good as manual (controlled vocabulary) • • TR Models (1970’s & 1980’s) … Large-scale Evaluation & Applications (1990’s-Present) – TREC (D. Harman & E. Voorhees, NIST) – Web search, PubMed, … – Boundary with related areas are disappearing 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 8 Short vs. Long Term Info Need • Short-term information need (Ad hoc retrieval) – “Temporary need”, e.g., info about used cars – Information source is relatively static – User “pulls” information – Application example: library search, Web search • Long-term information need (Filtering) – “Stable need”, e.g., new data mining algorithms – Information source is dynamic – System “pushes” information to user – Applications: news filter 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 9 Importance of Ad hoc Retrieval • Directly manages any existing large collection of information • There are many many “ad hoc” information needs • A long-term information need can be satisfied through frequent ad hoc retrieval • Basic techniques of ad hoc retrieval can be used for filtering and other “non-retrieval” tasks, such as automatic summarization. 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 10 Formal Formulation of TR • Vocabulary V={w1, w2, …, wN} of language • Query q = q1,…,qm, where qi V • Document di = di1,…,dimi, where dij V • Collection C= {d1, …, dk} • Set of relevant documents R(q) C – Generally unknown and user-dependent – Query is a “hint” on which doc is in R(q) • Task = compute R’(q), an “approximate R(q)” 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 11 Computing R(q) • Strategy 1: Document selection – R(q)={dC|f(d,q)=1}, where f(d,q) {0,1} is an indicator function or classifier – System must decide if a doc is relevant or not (“absolute relevance”) • Strategy 2: Document ranking – R(q) = {dC|f(d,q)>}, where f(d,q) is a relevance measure function; is a cutoff – System must decide if one doc is more likely to be relevant than another (“relative relevance”) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 12 Document Selection vs. Ranking True R(q) + +- - + - + + --- --- 1 Doc Selection f(d,q)=? Doc Ranking f(d,q)=? 2008 © ChengXiang Zhai 0 + +- + ++ R’(q) - -- - - + - 0.98 d1 + 0.95 d2 + 0.83 d3 0.80 d4 + 0.76 d5 0.56 d6 0.34 d7 0.21 d8 + 0.21 d9 - China-US-France Summer School, Lotus Hill Inst., 2008 R’(q) 13 Problems of Doc Selection • The classifier is unlikely accurate – “Over-constrained” query (terms are too specific): no relevant documents found – “Under-constrained” query (terms are too general): over delivery – It is extremely hard to find the right position between these two extremes • Even if it is accurate, all relevant documents are not equally relevant • Relevance is a matter of degree! 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 14 Ranking is often preferred • Relevance is a matter of degree • A user can stop browsing anywhere, so the boundary is controlled by the user – High recall users would view more items – High precision users would view only a few • Theoretical justification: Probability Ranking Principle [Robertson 77] 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 15 Probability Ranking Principle [Robertson 77] • As stated by Cooper “If a reference retrieval system’s response to each request is a ranking of the documents in the collections in order of decreasing probability of usefulness to the user who submitted the request, where the probabilities are estimated as accurately a possible on the basis of whatever data made available to the system for this purpose, then the overall effectiveness of the system to its users will be the best that is obtainable on the basis of that data.” • Robertson provides two formal justifications • Assumptions: Independent relevance and sequential browsing (not necessarily all hold in reality) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 16 According to the PRP, all we need is “A relevance measure function f” which satisfies For all q, d1, d2, f(q,d1) > f(q,d2) iff p(Rel|q,d1) >p(Rel|q,d2) Most IR research has focused on finding a good function f 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 17 Evaluation in Information Retrieval 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 18 Evaluation Criteria • Effectiveness/Accuracy – Precision, Recall • Efficiency – Space and time complexity • Usability – How useful for real user tasks? 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 19 Methodology: Cranfield Tradition • Laboratory testing of system components – Precision, Recall – Comparative testing • Test collections – Set of documents – Set of questions – Relevance judgments 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 20 The Contingency Table Action Doc Relevant Retrieved Not Retrieved Relevant Retrieved Relevant Rejected Not relevant Irrelevant Retrieved Irrelevant Rejected Re l e van tRe tri e ve d Pre ci si on Re tri e ve d Re l e van tRe tri e ve d Re call Re l e van t 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 21 How to measure a ranking? • Compute the precision at every recall point • Plot a precision-recall (PR) curve precision precision x Which is better? x x x x x x recall 2008 © ChengXiang Zhai x recall China-US-France Summer School, Lotus Hill Inst., 2008 22 Summarize a Ranking: MAP • Given that n docs are retrieved – Compute the precision (at rank) where each (new) relevant document is retrieved => p(1),…,p(k), if we have k rel. docs • • • – E.g., if the first rel. doc is at the 2nd rank, then p(1)=1/2. – If a relevant document never gets retrieved, we assume the precision corresponding to that rel. doc to be zero Compute the average over all the relevant documents – Average precision = (p(1)+…p(k))/k This gives us (non-interpolated) average precision, which captures both precision and recall and is sensitive to the rank of each relevant document Mean Average Precisions (MAP) – MAP = arithmetic mean average precision over a set of topics – gMAP = geometric mean average precision over a set of topics (more affected by difficult topics) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 23 Summarize a Ranking: NDCG • • • What if relevance judgments are in a scale of [1,r]? r>2 Cumulative Gain (CG) at rank n – Let the ratings of the n documents be r1, r2, …rn (in ranked order) – CG = r1+r2+…rn Discounted Cumulative Gain (DCG) at rank n – DCG = r1 + r2/log22 + r3/log23 + … rn/log2n – We may use any base for the logarithm, e.g., base=b • – For rank positions above b, do not discount Normalized Cumulative Gain (NDCG) at rank n – Normalize DCG at rank n by the DCG value at rank n of the ideal ranking – The ideal ranking would first return the documents with the highest relevance level, then the next highest relevance level, etc • – Compute the precision (at rank) where each (new) relevant document is retrieved => p(1),…,p(k), if we have k rel. docs NDCG is now quite popular in evaluating Web search 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 24 When There’s only 1 Relevant Document • Scenarios: – known-item search – navigational queries • Search Length = Rank of the answer: – measures a user’s effort • Mean Reciprocal Rank (MRR): – Reciprocal Rank: 1/Rank-of-the-answer – Take an average over all the queries 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 25 Precion-Recall Curve Out of 4728 rel docs, we’ve got 3212 Recall=3212/4728 Precision@10docs about 5.5 docs in the top 10 docs are relevant Breakeven Point (prec=recall) Mean Avg. Precision (MAP) D1 + D2 + D3 – D4 – D5 + D6 - Total # rel docs = 4 System returns 6 docs Average Prec = (1/1+2/2+3/5+0)/4 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 26 What Query Averaging Hides 1 0.9 0.8 Precision 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Recall Slide from Doug Oard’s presentation, originally from Ellen Voorhees’ presentation 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 27 The Pooling Strategy • • • • • When the test collection is very large, it’s impossible to completely judge all the documents TREC’s strategy: pooling – Appropriate for relative comparison of different systems – Given N systems, take top-K from the result of each, combine them to form a “pool” – Users judge all the documents in the pool; unjudged documents are assumed to be non-relevant Advantage: less human effort Potential problem: – bias due to incomplete judgments (okay for relative comparison) – Favor a system contributing to the pool, but when reused, a new system’s performance may be under-estimated Reuse the data set with caution! 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 28 User Studies • Limitations of Cranfield evaluation strategy: – How do we evaluate a technique for improving the interface of a search engine? • • – How do we evaluate the overall utility of a system? User studies are needed General user study procedure: – Experimental systems are developed – Subjects are recruited as users • – Variation can be in the system or the users – Users use the system and user behavior is logged – User information is collected (before: background, after: experience with the system) Clickthrough-based real-time user studies: – Assume clicked documents to be relevant – Mix results from multiple methods and compare their clickthroughs 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 29 Common Components in a TR System 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 30 Typical TR System Architecture docs query Feedback Tokenizer Query Rep Doc Rep (Index) Indexer Index 2008 © ChengXiang Zhai judgments User Scorer results China-US-France Summer School, Lotus Hill Inst., 2008 31 Text Representation/Indexing • Making it easier to match a query with a document • Query and document should be represented using the same units/terms • Controlled vocabulary vs. full text indexing • Full-text indexing is more practically useful and has proven to be as effective as manual indexing with controlled vocabulary 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 32 What is a good indexing term? • Specific (phrases) or general (single word)? • Luhn found that words with middle frequency are most useful – Not too specific (low utility, but still useful!) – Not too general (lack of discrimination, stop words) – Stop word removal is common, but rare words are kept • All words or a (controlled) subset? When term weighting is used, it is a matter of weighting not selecting of indexing terms 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 33 Tokenization • • Word segmentation is needed for some languages – Is it really needed? Normalize lexical units: Words with similar meanings should be mapped to the same indexing term – Stemming: Mapping all inflectional forms of words to the same root form, e.g. • computer -> compute • computation -> compute • computing -> compute (but king->k?) – Are we losing finer-granularity discrimination? • Stop word removal – What is a stop word? What about a query like “to be or not to be”? 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 34 Relevance Feedback Query Retrieval Engine Updated query Document collection Feedback 2008 © ChengXiang Zhai Results: d1 3.5 d2 2.4 … dk 0.5 ... User Judgments: d1 + d2 d3 + … dk ... China-US-France Summer School, Lotus Hill Inst., 2008 35 Pseudo/Blind/Automatic Feedback Query Retrieval Engine Updated query Document collection Feedback 2008 © ChengXiang Zhai Results: d1 3.5 d2 2.4 … dk 0.5 ... Judgments: d1 + d2 + d3 + … dk ... top 10 China-US-France Summer School, Lotus Hill Inst., 2008 36 Implicit Feedback Query Retrieval Engine Updated query Results: d1 3.5 d2 2.4 … dk 0.5 ... User Document collection Feedback 2008 © ChengXiang Zhai Judgments: d1 + d2 d3 + … dk ... infer User Activities e.g. clickthroughs China-US-France Summer School, Lotus Hill Inst., 2008 37 Important Points to Remember • • • • • PRP provides a justification for ranking, which is generally preferred to document selection How to compute the major evaluation measure (precision, recall, precision-recall curve, MAP, gMAP, breakeven precision, NDCG, MRR) What is pooling What is tokenization (word segmentation, stemming, stop word removal) What are relevance feedback, pseudo relevance feedback, and implicit feedback 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 38 Overview of Retrieval Models Relevance (Rep(q), Rep(d)) Similarity Different rep & similarity P(r=1|q,d) r {0,1} Probability of Relevance Regression Model (Fox 83) Generative Model P(d q) or P(q d) Probabilistic inference Different inference system Query Doc Learn to generation generation … Inference Prob. concept Rank network space model (Joachims 02) Vector space model Prob. distr. Classical LM (Wong & Yao, 95) (Burges et al. 05) model (Turtle & Croft, 91) model prob. Model approach (Salton et al., 75) (Wong & Yao, 89) (Robertson & (Ponte & Croft, 98) Sparck Jones, 76) (Lafferty & Zhai, 01a) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 39 Retrieval Models: Vector Space 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 40 Relevance = Similarity • Assumptions – Query and document are represented similarly – A query can be regarded as a “document” – Relevance(d,q) similarity(d,q) • R(q) = {dC|f(d,q)>}, f(q,d)=(Rep(q), Rep(d)) • Key issues – How to represent query/document? – How to define the similarity measure ? 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 41 Vector Space Model • Represent a doc/query by a term vector – Term: basic concept, e.g., word or phrase – Each term defines one dimension – N terms define a high-dimensional space – Element of vector corresponds to term weight – E.g., d=(x1,…,xN), xi is “importance” of term i • Measure relevance by the distance between the query vector and document vector in the vector space 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 42 VS Model: illustration Starbucks D9 D11 ?? ?? D2 D5 D3 D10 D4 D6 Java Query D7 D8 D1 Microsoft ?? 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 43 What the VS model doesn’t say • How to define/select the “basic concept” – Concepts are assumed to be orthogonal • How to assign weights – Weight in query indicates importance of term – Weight in doc indicates how well the term characterizes the doc • How to define the similarity/distance measure 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 44 What’s a good “basic concept”? • Orthogonal – Linearly independent basis vectors – “Non-overlapping” in meaning • No ambiguity • Weights can be assigned automatically and hopefully accurately • Many possibilities: Words, stemmed words, phrases, “latent concept”, … 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 45 How to Assign Weights? • Very very important! • Why weighting – Query side: Not all terms are equally important – Doc side: Some terms carry more information about contents • How? – Two basic heuristics • TF (Term Frequency) = Within-doc-frequency • IDF (Inverse Document Frequency) – TF normalization 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 46 TF Weighting • Idea: A term is more important if it occurs more frequently in a document • Some formulas: Let f(t,d) be the frequency count of term t in doc d – Raw TF: TF(t,d) = f(t,d) – Log TF: TF(t,d)=log f(t,d) – Maximum frequency normalization: TF(t,d) = 0.5 +0.5*f(t,d)/MaxFreq(d) – “Okapi/BM25 TF”: TF(t,d) = k f(t,d)/(f(t,d)+k(1-b+b*doclen/avgdoclen)) • Normalization of TF is very important! 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 47 TF Normalization • Why? – Document length variation – “Repeated occurrences” are less informative than the “first occurrence” • Two views of document length – A doc is long because it uses more words – A doc is long because it has more contents • Generally penalize long doc, but avoid overpenalizing (pivoted normalization) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 48 TF Normalization: How? Norm. TF Raw TF Which curve is more reasonable? Should normalized-TF be up-bounded? Normalization interacts with the similarity measure 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 49 Regularized/“Pivoted” Length Normalization Norm. TF Raw TF “Pivoted normalization”: Using avg. doc length to regularize normalization 1-b+b*doclen/avgdoclen (b varies from 0 to 1) What would happen if doclen is {>, <,=} avgdoclen? Advantage: stabalize parameter setting 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 50 IDF Weighting • Idea: A term is more discriminative if it occurs only in fewer documents • Formula: IDF(t) = 1+ log(n/k) n – total number of docs k -- # docs with term t (doc freq) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 51 TF-IDF Weighting • TF-IDF weighting : weight(t,d)=TF(t,d)*IDF(t) – Common in doc high tf high weight – Rare in collection high idf high weight • Imagine a word count profile, what kind of terms would have high weights? 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 52 How to Measure Similarity? Di ( w i 1 ,...,w iN ) Q ( wq1 ,...,wqN ) w 0 if a te rmis abse n t N Dot produ ctsim ilarity : sim (Q , Di ) wqj w ij j 1 N C osin e: sim (Q , Di ) wqj wij j 1 N ( wqj ) 2 j 1 N ( wij )2 j 1 ( n orm aliz e dot d produ ct) How about Euclidean? 2008 © ChengXiang Zhai sim(Q, Di ) N 2 ( w w ) qj ij j 1 China-US-France Summer School, Lotus Hill Inst., 2008 53 What Works the Best? Error [ ] •Use single words •Use stat. phrases •Remove stop words •Stemming •Others(?) (Singhal 2001) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 54 “Extension” of VS Model • Alternative similarity measures – Many other choices (tend not to be very effective) – P-norm (Extended Boolean): matching a Boolean query with a TF-IDF document vector • Alternative representation – Many choices (performance varies a lot) – Latent Semantic Indexing (LSI) [TREC performance tends to be average] • Generalized vector space model – Theoretically interesting, not seriously evaluated 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 55 Relevance Feedback in VS • Basic setting: Learn from examples – Positive examples: docs known to be relevant – Negative examples: docs known to be non-relevant • – How do you learn from this to improve performance? General method: Query modification – Adding new (weighted) terms – Adjusting weights of old terms • – Doing both The most well-known and effective approach is Rocchio [Rocchio 1971] 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 56 Rocchio Feedback: Illustration --- --+ + + + q q -- + - +++ + + - - - + +++ + ++ -- -- -- 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 57 Rocchio Feedback: Formula Parameters New query Origial query Rel docs 2008 © ChengXiang Zhai Non-rel docs China-US-France Summer School, Lotus Hill Inst., 2008 58 Rocchio in Practice • • • • • Negative (non-relevant) examples are not very important (why?) Often truncate the vector onto to lower dimension (i.e., consider only a small number of words that have high weights in the centroid vector) (efficiency concern) Avoid overfitting by keeping a relatively high weight on the original query weights (why?) Can be used for relevance feedback and pseudo feedback Usually robust and effective 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 59 Advantages of VS Model • Empirically effective! (Top TREC performance) • Intuitive • Easy to implement • Well-studied/Most evaluated • The Smart system – Developed at Cornell: 1960-1999 – Still available • Warning: Many variants of TF-IDF! 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 60 Disadvantages of VS Model • Assume term independence • Assume query and document to be the same • Lack of “predictive adequacy” – Arbitrary term weighting – Arbitrary similarity measure • Lots of parameter tuning! 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 61 Probabilistic Retrieval Models 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 62 Overview of Retrieval Models Relevance (Rep(q), Rep(d)) Similarity Different rep & similarity P(r=1|q,d) r {0,1} Probability of Relevance Regression Model (Fox 83) Generative Model P(d q) or P(q d) Probabilistic inference Different inference system Query Doc Learn to generation generation … Inference Prob. concept Rank network space model (Joachims 02) Vector space model Prob. distr. Classical LM (Wong & Yao, 95) (Burges et al. 05) model (Turtle & Croft, 91) model prob. Model approach (Salton et al., 75) (Wong & Yao, 89) (Robertson & (Ponte & Croft, 98) Sparck Jones, 76) (Lafferty & Zhai, 01a) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 63 Probability of Relevance • Three random variables – Query Q – Document D – Relevance R {0,1} • Goal: rank D based on P(R=1|Q,D) – Evaluate P(R=1|Q,D) – Actually, only need to compare P(R=1|Q,D1) with P(R=1|Q,D2), I.e., rank documents • Several different ways to refine P(R=1|Q,D) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 64 • Refining P(R=1|Q,D) Method 1: conditional models Basic idea: relevance depends on how well a query matches a document – Define features on Q x D, e.g., #matched terms, # the highest IDF of a matched term, #doclen,.. – P(R=1|Q,D)=g(f1(Q,D), f2(Q,D),…,fn(Q,D), ) – Using training data (known relevance judgments) to estimate parameter • • – Apply the model to rank new documents Early work (e.g., logistic regression [Cooper 92, Gey 94]) – Attempted to compete with other models Recent work (e.g. Ranking SVM [Joachims 02], RankNet (Burges et al. 05)) – Attempted to leverage other models – More features (notably PageRank, anchor text) – More sophisticated learning (Ranking SVM, RankNet, …) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 65 Logistic Regression (Cooper 92, Gey 94) 6 P ( R 1 | Q, D) log 0 i X i 1 P ( R 1 | Q, D) i 1 P ( R 1 | Q, D) logit function: log x 1 x logistic (sigmoid) function: 1 1 e xp(x ) 1 e xp( x ) 1 e xp(x ) 6 1 e xp( 0 i X i ) i 1 P(R=1|Q,D) Uses 6 features X1, …, X6 1.0 X 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 66 Features/Attributes 1 X1 M M logQAF tj 1 X 2 QL 1 X3 M Average Absolute Query Frequency Query Length M log DAF tj Average Absolute Document Frequency 1 X 4 DL Document Length 1 M X5 log IDFt j M 1 N nt j IDF nt j Average Inverse Document Frequency X 6 log M Number of Terms in common between query and document -- logged Inverse Document Frequency 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 67 Learning to Rank • Advantages – May combine multiple features (helps improve accuracy and combat web spams) – May re-use all the past relevance judgments (self-improving) • Problems – Don’t learn the semantic associations between query words and document words – No much guidance on feature generation (rely on traditional retrieval models) • All current Web search engines use some kind of learning algorithms to combine many features such as PageRank and many different representations of a page 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 68 The PageRank Algorithm (Page et al. 98) Random surfing model: At any page, With prob. , randomly jumping to a page With prob. (1-), randomly picking a link to follow. 0 1 M 0 1/ 2 d1 d3 d2 1/ 2 1/ 2 0 0 0 0 “Transition matrix” 1/ 2 0 0 0 0 1 N N= # pages N 1 p (d k ) k 1 N p(di ) (1 ) mki p(d k ) k 1 d4 N [ k 1 1 (1 )mki ] p(d k ) N p ( I (1 ) M )T p Initial value p(d)=1/N Iij = 1/N Stationary (“stable”) distribution, so we ignore time Iterate until converge 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 69 PageRank in Practice • Interpretation of the damping factor (0.15): – Probability of a random jump – Smoothing the transition matrix (avoid zero’s) N • N 1 p (d k ) k 1 N p(di ) (1 ) mki p(dk ) k 1 p ( I (1 )M )T p Normalization doesn’t affect ranking, leading to some variants N N k 1 k 1 p '(di ) (1 ) mki p '(d k ) CN N N k 1 k 1 cp(di ) (1 ) mki cp(d k ) Let p '(di ) cp(di ), c constant , 1 cp(d k ) N 1 p '(d k ) N N p '(di ) (1 ) mki p '(d k ) k 1 • The zero-outlink problem: p(di)’s don’t sum to 1 – One possible solution = page-specific damping factor (=1.0 for a page with no outlink) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 70 HITS: Capturing Authorities & Hubs • Intuitions [Kleinberg 98] – Pages that are widely cited are good authorities – Pages that cite many other pages are good hubs • The key idea of HITS – Good authorities are cited by good hubs – Good hubs point to good authorities – Iterative reinforcement… 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 71 The HITS Algorithm [Kleinberg 98] d1 d3 d2 d4 0 0 1 0 A 0 1 1 1 h( d i ) 1 1 “Adjacency matrix” 0 0 0 0 0 0 Initial values: a(di)=h(di)=1 a(d j ) d j OUT ( di ) a(di ) d j IN ( di ) h Aa ; Iterate h( d j ) Normalize: a AT h 2 h AAT h ; a AT Aa 2008 © ChengXiang Zhai a(di ) h(di ) 1 i 2 i China-US-France Summer School, Lotus Hill Inst., 2008 72 Refining P(R=1|Q,D) Method 2: generative models • Basic idea – Define P(Q,D|R) – Compute O(R=1|Q,D) using Bayes’ rule O( R 1 | Q, D) P( R 1 | Q, D) P(Q, D | R 1) P( R 1) P( R 0 | Q, D) P(Q, D | R 0) P( R 0) Ignored for ranking D • Special cases – Document “generation”: P(Q,D|R)=P(D|Q,R)P(Q|R) – Query “generation”: P(Q,D|R)=P(Q|D,R)P(D|R) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 73 Document Generation P( R 1 | Q, D) P(Q, D | R 1) P( R 0 | Q, D) P(Q, D | R 0) P( D | Q, R 1) P(Q | R 1) P( D | Q, R 0) P(Q | R 0) P( D | Q, R 1) P( D | Q, R 0) Model of relevant docs for Q Model of non-relevant docs for Q Assume independent attributes A1…Ak ….(why?) Let D=d1…dk, where dk {0,1} is the value of attribute Ak (Similarly Q=q1…qk ) k P( Ai d i | Q, R 1) P ( R 1 | Q, D ) P( R 0 | Q, D) i 1 P( Ai d i | Q, R 0) Non-query terms are equally likely to appear in relevant and non-relevant docs P( Ai 1 | Q, R 1) P( Ai 0 | Q, R 1) i 1, d i 1 P ( Ai 1 | Q, R 0) i 1, d i 0 P ( Ai 0 | Q, R 0) P( Ai 1 | Q, R 1) P( Ai 0 | Q, R 0) i 1, d i 1 P ( Ai 1 | Q, R 0) P ( Ai 0 | Q, R 1) P( Ai 1 | Q, R 1) P( Ai 0 | Q, R 0) ( Assum e P( Ai 1 | Q, R 1) P( Ai 1 | Q, R 0), if qi 0) i 1, d i qi 1 P ( Ai 1 | Q, R 0) P ( Ai 0 | Q, R 1) k k k k 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 74 Robertson-Sparck Jones Model (Robertson & Sparck Jones 76) logO( R 1 | Q, D) Rank k i 1, d i qi 1 log pi (1 qi ) qi (1 pi ) (RSJ model) Two parameters for each term Ai: pi = P(Ai=1|Q,R=1): prob. that term Ai occurs in a relevant doc qi = P(Ai=1|Q,R=0): prob. that term Ai occurs in a non-relevant doc How to estimate parameters? Suppose we have relevance judgments, pˆ i # ( rel . doc with Ai ) 0.5 # ( rel .doc) 1 qˆ i # ( nonrel. doc with Ai ) 0.5 # ( nonrel.doc) 1 “+0.5” and “+1” can be justified by Bayesian estimation 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 75 RSJ Model: No Relevance Info (Croft & Harper 79) k Rank logO( R 1 | Q, D) log i 1, d i qi 1 pi (1 qi ) qi (1 pi ) (RSJ model) How to estimate parameters? Suppose we do not have relevance judgments, - We will assume pi to be a constant - Estimate qi by assuming all documents to be non-relevant N ni 0.5 logO( R 1 | Q, D) log ni 0.5 i 1, d i qi 1 Rank k IDF ' log N ni ni N: # documents in collection ni: # documents in which term Ai occurs 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 76 Improving RSJ: Adding TF Basic doc. generation model: P( R 1 | Q, D) P( D | Q, R 1) P( R 0 | Q, D) P( D | Q, R 0) Let D=d1…dk, where dk is the frequency count of term Ak k P( Ai d i | Q, R 1) P ( R 1 | Q, D ) P( R 0 | Q, D) i 1 P( Ai d i | Q, R 0) P( Ai d i | Q, R 1) k P( Ai 0 | Q, R 1) P ( A d | Q , R 0 ) i 1, d i 1 i 1, d i 0 P ( Ai 0 | Q, R 0) i i k P( Ai d i | Q, R 1) P( Ai 0 | Q, R 0) i 1, d i 1 P ( Ai d i | Q, R 0) P ( Ai 0 | Q, R 1) k 2-Poisson mixture model p ( Ai f | Q, R ) p ( E | Q, R ) p ( Ai f | E ) P ( E | Q, R ) p ( Ai f | E ) p ( E | Q, R ) E f f! e E P ( E | Q, R ) E f f! e E Many more parameters to estimate! (how many exactly?) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 77 BM25/Okapi Approximation • (Robertson et al. 94) Idea: Approximate p(R=1|Q,D) with a simpler function that share similar properties • Observations: – log O(R=1|Q,D) is a sum of term weights Wi – Wi= 0, if TFi=0 – Wi increases monotonically with TFi – Wi has an asymptotic limit • The simple function is TFi ( k1 1) pi (1 qi ) Wi log k1 TFi qi (1 pi ) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 78 Adding Doc. Length & Query TF • Incorporating doc length – Motivation: The 2-Poisson model assumes equal document length – Implementation: “Carefully” penalize long doc • Incorporating query TF – Motivation: Appears to be not well-justified – Implementation: A similar TF transformation • The final formula is called BM25, achieving top TREC performance 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 79 The BM25 Formula “Okapi TF/BM25 TF” 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 80 Extensions of “Doc Generation” Models • Capture term dependence (Rijsbergen & Harper 78) • Alternative ways to incorporate TF (Croft 83, Kalt96) • Feature/term selection for feedback (Okapi’s TREC reports) • Other Possibilities (machine learning … ) 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 81 Query Generation P (Q , D | R 1) P (Q , D | R 0 ) P (Q | D, R 1) P ( D | R 1) P (Q | D , R 0 ) P ( D | R 0 ) P ( D | R 1) P (Q | D, R 1) ( Assume P (Q | D, R 0) P (Q | R 0)) P ( D | R 0) O( R 1 | Q , D ) Query likelihood p(q| d) Document prior Assuming uniform prior, we have O( R 1 | Q, D) P (Q | D, R 1) Now, the question is how to compute P(Q | D, R 1) ? Generally involves two steps: (1) estimate a language model based on D (2) compute the query likelihood according to the estimated model Leading to the so-called “Language Modeling Approach” … 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 82 Lecture 1: Key Points • • • Vector Space Model is a family of models, not a single model Many variants of TF-IDF weighting and some are more effective than others State of the art retrieval performance is achieved through – – – – • Bag of words representation TF-IDF weighting (BM25) + length normalization Pseudo relevance feedback (mostly for recall) For web search, add PageRank, anchor text, …, plus learning to rank Principled approaches didn’t lead to good performance directly (before the “language modeling approach” was proposed); heuristic modification has been necessary 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 83 Readings • Amit Singhal’s overview: – http://sifaka.cs.uiuc.edu/course/ds/mir.pdf • My review of IR models: – http://sifaka.cs.uiuc.edu/course/ds/irmod.pdf 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 84 Discussion • Text retrieval and image retrieval – Query language: • keywords vs. image features • By example – Content representation • Bag of words vs. “bag of image features”? • Phrase indexing vs. units from image parsing? • Sentiment analysis – Retrieval heuristics • TF-IDF weighting vs. ? • Passage retrieval vs. image region retrieval? • • • Proximity Text retrieval and video retrieval Multimedia retrieval? 2008 © ChengXiang Zhai China-US-France Summer School, Lotus Hill Inst., 2008 85