Question? Leave a message!




Evaluation & Result Summaries

Evaluation & Result Summaries
Introduction to Information Retrieval Introduction to Information Retrieval Evaluation Result Summaries 1Introduction to Information Retrieval Overview ❶ Recap ❷ Unranked evaluation ❸ Ranked evaluation ❹ Evaluation benchmarks ❺ Result summaries 2Introduction to Information Retrieval Outline ❶ Recap ❷ Unranked evaluation ❸ Ranked evaluation ❹ Evaluation benchmarks ❺ Result summaries 3Introduction to Information Retrieval 4 4Introduction to Information Retrieval Pivot normalization source: Lilian Lee 5 5Introduction to Information Retrieval Heuristics for finding the top k even faster  Documentatatime processing  We complete computation of the querydocument similarity score of document d before starting to compute the query i document similarity score of d . i+1  Requires a consistent ordering of documents in the postings lists  Termatatime processing  We complete processing the postings list of query term t i before starting to process the postings list of t . i+1  Requires an accumulator for each document “still in the running”  The most effective heuristics switch back and forth between termatatime and documentatatime processing. 6 6Introduction to Information Retrieval Use min heap for selecting top k ouf of N  Use a binary min heap  A binary min heap is a binary tree in which each node’s value is less than the values of its children.  It takes O(N log k) operations to construct the kheap containing the k largest values (where N is the number of documents).  Essentially linear in N for small k and large N. 7 7Introduction to Information Retrieval Binary min heap 8Introduction to Information Retrieval Selecting k top scoring documents in O(N log k)  Goal: Keep the k top documents seen so far  Use a binary min heap  To process a new document d′ with score s′:  Get current minimum h of heap (in O(1)) m  If s′ ≤ h skip to next document m  If s′ h heapdeleteroot (in O(log k)) m  Heapadd d′/s′ (in O(1))  Reheapify (in O(log k)) 9 9Introduction to Information Retrieval Tiered index 10 10Introduction to Information Retrieval Outline ❶ Recap ❷ Unranked evaluation ❸ Ranked evaluation ❹ Evaluation benchmarks ❺ Result summaries 11Introduction to Information Retrieval Measures for a search engine  How fast does it index  e.g., number of bytes per hour  How fast does it search  e.g., latency as a function of queries per second  What is the cost per query  in dollars 12 12Introduction to Information Retrieval Measures for a search engine  All of the preceding criteria are measurable: we can quantify speed / size / money  However, the key measure for a search engine is user happiness.  What is user happiness  Factors include:  Speed of response  Size of index  Uncluttered UI  Most important: relevance  (actually, maybe even more important: it’s free)  Note that none of these is sufficient: blindingly fast, but useless answers won’t make a user happy.  How can we quantify user happiness 13 13Introduction to Information Retrieval Who is the user  Who is the user we are trying to make happy  Web search engine: searcher. Success: Searcher finds what she was looking for. Measure: rate of return to this search engine  Web search engine: advertiser. Success: Searcher clicks on ad. Measure: clickthrough rate  Ecommerce: buyer. Success: Buyer buys something. Measures: time to purchase, fraction of “conversions” of searchers to buyers  Ecommerce: seller. Success: Seller sells something. Measure: profit per item sold  Enterprise: CEO. Success: Employees are more productive (because of effective search). Measure: profit of the company 14 14Introduction to Information Retrieval Most common definition of user happiness: Relevance  User happiness is equated with the relevance of search results to the query.  But how do you measure relevance  Standard methodology in information retrieval consists of three elements.  A benchmark document collection  A benchmark suite of queries  An assessment of the relevance of each querydocument pair 15 15Introduction to Information Retrieval Relevance: query vs. information need  Relevance to what  First take: relevance to the query  “Relevance to the query” is very problematic.  Information need i : “I am looking for information on whether drinking red wine is more effective at reducing your risk of heart attacks than white wine.”  This is an information need, not a query.  Query q: red wine white wine heart attack  Consider document d′: At heart of his speech was an attack on the wine industry lobby for downplaying the role of red and white wine in drunk driving.  d′ is an excellent match for query q . . .  d′ is not relevant to the information need i . 16 16Introduction to Information Retrieval Relevance: query vs. information need  User happiness can only be measured by relevance to an information need, not by relevance to queries.  Our terminology is sloppy in these slides and in IIR: we talk about querydocument relevance judgments even though we mean informationneeddocument relevance judgments. 17 17Introduction to Information Retrieval Precision and recall  Precision (P) is the fraction of retrieved documents that are relevant  Recall (R) is the fraction of relevant documents that are retrieved 18 18Introduction to Information Retrieval Precision and recall P = TP / ( TP + FP ) R = TP / ( TP + FN ) 19 19Introduction to Information Retrieval Precision/recall tradeoff  You can increase recall by returning more docs.  Recall is a nondecreasing function of the number of docs retrieved.  A system that returns all docs has 100 recall  The converse is also true (usually): It’s easy to get high precision for very low recall.  Suppose the document with the largest score is relevant. How can we maximize precision 20 20Introduction to Information Retrieval A combined measure: F  F allows us to trade off precision against recall. where 2  α ϵ 0, 1 and thus b ϵ 0,∞  Most frequently used: balanced F with b = 1 or α = 0.5  This is the harmonic mean of P and R:  What value range of β weights recall higher than precision 21 21Introduction to Information Retrieval F: Example relevant not relevant retrieved 20 40 60 not retrieved 60 1,000,000 1,000,060 80 1,000,040 1,000,120  P = 20/(20 + 40) = 1/3  R = 20/(20 + 60) = 1/4  22 22Introduction to Information Retrieval Accuracy  Why do we use complex measures like precision, recall, and F  Why not something simple like accuracy  Accuracy is the fraction of decisions (relevant/nonrelevant) that are correct.  In terms of the contingency table above, accuracy = (TP + TN)/(TP + FP + FN + TN).  Why is accuracy not a useful measure for web information retrieval 23 23Introduction to Information Retrieval Exercise  Compute precision, recall and F for this result set: 1 relevant not relevant retrieved 18 2 not retrieved 82 1,000,000,000  The snoogle search engine below always returns 0 results (“0 matching results found”), regardless of the query. Why does snoogle demonstrate that accuracy is not a useful measure in IR 24 24Introduction to Information Retrieval Why accuracy is a useless measure in IR  Simple trick to maximize accuracy in IR: always say no and return nothing  You then get 99.99 accuracy on most queries.  Searchers on the web (and in IR in general) want to find something and have a certain tolerance for junk.  It’s better to return some bad hits as long as you return something.  →We use precision, recall, and F for evaluation, not accuracy. 25 25Introduction to Information Retrieval F: Why harmonic mean  Why don’t we use a different mean of P and R as a measure  e.g., the arithmetic mean  The simple (arithmetic) mean is 50 for “returneverything” search engine, which is too high.  Desideratum: Punish really bad performance on either precision or recall.  Taking the minimum achieves this.  But minimum is not smooth and hard to weight.  F (harmonic mean) is a kind of smooth minimum. 26 26Introduction to Information Retrieval F and other averages 1  We can view the harmonic mean as a kind of soft minimum 27 27Introduction to Information Retrieval Difficulties in using precision, recall and F  We need relevance judgments for informationneed document pairs – but they are expensive to produce.  For alternatives to using precision/recall and having to produce relevance judgments – see end of this lecture. 28 28Introduction to Information Retrieval Outline ❶ Recap ❷ Unranked evaluation ❸ Ranked evaluation ❹ Evaluation benchmarks ❺ Result summaries 29Introduction to Information Retrieval Precisionrecall curve  Precision/recall/F are measures for unranked sets.  We can easily turn set measures into measures of ranked lists.  Just compute the set measure for each “prefix”: the top 1, top 2, top 3, top 4 etc results  Doing this for precision and recall gives you a precisionrecall curve. 30 30Introduction to Information Retrieval A precisionrecall curve  Each point corresponds to a result for the top k ranked hits (k = 1, 2, 3, 4, . . .).  Interpolation (in red): Take maximum of all future points  Rationale for interpolation: The user is willing to look at more stuff if both precision and recall get better.  Questions 31 31Introduction to Information Retrieval 11point interpolated average precision Recall Interpolated Precision 0.0 1.00 0.1 0.67 11point average: ≈ 0.2 0.63 0.425 0.3 0.55 0.4 0.45 How can precision 0.5 0.41 at 0.0 be 0 0.6 0.36 0.7 0.29 0.8 0.13 0.9 0.10 1.0 0.08 32 32Introduction to Information Retrieval Averaged 11point precision/recall graph  Compute interpolated precision at recall levels 0.0, 0.1, 0.2, . . .  Do this for each of the queries in the evaluation benchmark  Average over queries  This measure measures performance at all recall levels.  The curve is typical of performance levels at TREC.  Note that performance is not very good 33 33Introduction to Information Retrieval ROC curve  Similar to precisionrecall graph  But we are only interested in the small area in the lower left corner.  Precisionrecall graph “blows up” this area. 34 34Introduction to Information Retrieval Variance of measures like precision/recall  For a test collection, it is usual that a system does badly on some information needs (e.g., P = 0.2 at R = 0.1) and really well on others (e.g., P = 0.95 at R = 0.1).  Indeed, it is usually the case that the variance of the same system across queries is much greater than the variance of different systems on the same query.  That is, there are easy information needs and hard ones. 35 35Introduction to Information Retrieval Outline ❶ Recap ❷ Unranked evaluation ❸ Ranked evaluation ❹ Evaluation benchmarks ❺ Result summaries 36Introduction to Information Retrieval What we need for a benchmark  A collection of documents  Documents must be representative of the documents we expect to see in reality.  A collection of information needs  . . .which we will often incorrectly refer to as queries  Information needs must be representative of the information needs we expect to see in reality.  Human relevance assessments  We need to hire/pay “judges” or assessors to do this.  Expensive, timeconsuming  Judges must be representative of the users we expect to see in reality. 37 37Introduction to Information Retrieval Standard relevance benchmark: Cranfield  Pioneering: first testbed allowing precise quantitative measures of information retrieval effectiveness  Late 1950s, UK  1398 abstracts of aerodynamics journal articles, a set of 225 queries, exhaustive relevance judgments of all query documentpairs  Too small, too untypical for serious IR evaluation today 38 38Introduction to Information Retrieval Standard relevance benchmark: TREC  TREC = Text Retrieval Conference (TREC)  Organized by the U.S. National Institute of Standards and Technology (NIST)  TREC is actually a set of several different relevance benchmarks.  Best known: TREC Ad Hoc, used for first 8 TREC evaluations between 1992 and 1999  1.89 million documents, mainly newswire articles, 450 information needs  No exhaustive relevance judgments – too expensive  Rather, NIST assessors’ relevance judgments are available only for the documents that were among the top k returned for some system which was entered in the TREC evaluation for which the information need was developed. 39 39Introduction to Information Retrieval Standard relevance benchmarks: Others  GOV2  Another TREC/NIST collection  25 million web pages  Used to be largest collection that is easily available  But still 3 orders of magnitude smaller than what Google/Yahoo/MSN index  NTCIR  East Asian language and crosslanguage information retrieval  Cross Language Evaluation Forum (CLEF)  This evaluation series has concentrated on European languages and crosslanguage information retrieval.  Many others 40 40Introduction to Information Retrieval Validity of relevance assessments  Relevance assessments are only usable if they are consistent.  If they are not consistent, then there is no “truth” and experiments are not repeatable.  How can we measure this consistency or agreement among judges  → Kappa measure 41 41Introduction to Information Retrieval Kappa measure  Kappa is measure of how much judges agree or disagree.  Designed for categorical judgments  Corrects for chance agreement  P(A) = proportion of time judges agree  P(E) = what agreement would we get by chance  k = for (i) chance agreement (ii) total agreement 42 42Introduction to Information Retrieval Kappa measure (2)  Values of k in the interval 2/3, 1.0 are seen as acceptable.  With smaller values: need to redesign relevance assessment methodology used etc. 43 43Introduction to Information Retrieval Calculating the kappa statistic Judge 2 Relevance Yes No Total Yes 300 20 320 Judge 1 Observed proportion of No 10 70 80 Relevance the times the judges agreed Total 310 90 400 P(A) = (300 + 70)/400 = 370/400 = 0.925 Pooled marginals P(nonrelevant) = (80 + 90)/(400 + 400) = 170/800 = 0.2125 P(relevant) = (320 + 310)/(400 + 400) = 630/800 = 0.7878 Probability that the two judges agreed by chance P(E) = 2 2 2 2 P(nonrelevant) + P(relevant) = 0.2125 + 0.7878 = 0.665 Kappa statistic к = (P(A) − P(E))/(1 − P(E)) = (0.925 − 0.665)/(1 − 0.665) = 0.776 (still in acceptable range) 44 44Introduction to Information Retrieval Interjudge agreement at TREC Information number of disagreements need docs judged 51 211 6 62 400 157 67 400 68 95 400 110 127 400 106 45 45Introduction to Information Retrieval Impact of interjudge disagreement  Judges disagree a lot. Does that mean that the results of information retrieval experiments are meaningless  No.  Large impact on absolute performance numbers  Virtually no impact on ranking of systems  Suppose we want to know if algorithm A is better than algorithm B  An information retrieval experiment will give us a reliable answer to this question . . .  . . . even if there is a lot of disagreement between judges. 46 46Introduction to Information Retrieval Evaluation at large search engines  Recall is difficult to measure on the web  Search engines often use precision at top k, e.g., k = 10 . . .  . . . or use measures that reward you more for getting rank 1 right than for getting rank 10 right.  Search engines also use nonrelevancebased measures.  Example 1: clickthrough on first result  Not very reliable if you look at a single clickthrough (you may realize after clicking that the summary was misleading and the document is nonrelevant) . . .  . . . but pretty reliable in the aggregate.  Example 2: Ongoing studies of user behavior in the lab – recall last lecture  Example 3: A/B testing 47 47Introduction to Information Retrieval A/B testing  Purpose: Test a single innovation  Prerequisite: You have a large search engine up and running.  Have most users use old system  Divert a small proportion of traffic (e.g., 1) to the new system that includes the innovation  Evaluate with an “automatic” measure like clickthrough on first result  Now we can directly see if the innovation does improve user happiness.  Probably the evaluation methodology that large search engines trust most 48 48Introduction to Information Retrieval Critique of pure relevance  We’ve defined relevance for an isolated querydocument pair.  Alternative definition: marginal relevance  The marginal relevance of a document at position k in the result list is the additional information it contributes over and above the information that was contained in documents d . . . d . 1 k−1  Exercise  Why is marginal relevance a more realistic measure of user happiness  Give an example where a nonmarginal measure like precision or recall is a misleading measure of user happiness, but marginal relevance is a good measure.  In a practical application, what is the difficulty of using marginal measures instead of nonmarginal measures 49 49Introduction to Information Retrieval Outline ❶ Recap ❷ Unranked evaluation ❸ Ranked evaluation ❹ Evaluation benchmarks ❺ Result summaries 50Introduction to Information Retrieval How do we present results to the user  Most often: as a list – aka “10 blue links”  How should each document in the list be described  This description is crucial.  The user often can identify good hits (= relevant hits) based on the description.  No need to “click” on all documents sequentially 51 51Introduction to Information Retrieval Doc description in result list  Most commonly: doc title, url, some metadata . . .  . . . and a summary  How do we “compute” the summary 52 52Introduction to Information Retrieval Summaries  Two basic kinds: (i) static (ii) dynamic  A static summary of a document is always the same, regardless of the query that was issued by the user.  Dynamic summaries are querydependent. They attempt to explain why the document was retrieved for the query at hand. 53 53Introduction to Information Retrieval Static summaries  In typical systems, the static summary is a subset of the document.  Simplest heuristic: the first 50 or so words of the document  More sophisticated: extract from each document a set of “key” sentences  Simple NLP heuristics to score each sentence  Summary is made up of topscoring sentences.  Machine learning approach: see IIR 13  Most sophisticated: complex NLP to synthesize/generate a summary  For most IR applications: not quite ready for prime time yet 54 54Introduction to Information Retrieval Dynamic summaries  Present one or more “windows” or snippets within the document that contain several of the query terms.  Prefer snippets in which query terms occurred as a phrase  Prefer snippets in which query terms occurred jointly in a small window  The summary that is computed this way gives the entire content of the window – all terms, not just the query terms. 55 55Introduction to Information Retrieval A dynamic summary Query: “new guinea economic development” Snippets (in bold) that were extracted from a document: . . . In recent years, Papua New Guinea has faced severe economic difficulties and economic growth has slowed, partly as a result of weak governance and civil war, and partly as a result of external factors such as the Bougainville civil war which led to the closure in 1989 of the Panguna mine (at that time the most important foreign exchange earner and contributor to Government finances), the Asian financial crisis, a decline in the prices of gold and copper, and a fall in the production of oil. PNG’s economic development record over the past few years is evidence that governance issues underly many of the country’s problems. Good governance, which may be defined as the transparent and accountable management of human, natural, economic and financial resources for the purposes of equitable and sustainable development, flows from proper public sector management, efficient fiscal and accounting mechanisms, and a willingness to make service delivery a priority in practice. . . . 56 56Introduction to Information Retrieval Google example for dynamic summaries 57 57Introduction to Information Retrieval Generating dynamic summaries  Where do we get these other terms in the snippet from  We cannot construct a dynamic summary from the positional inverted index – at least not efficiently.  We need to cache documents.  The positional index tells us: query term occurs at position 4378 in the document.  Byte offset or word offset  Note that the cached copy can be outdated  Don’t cache very long documents – just cache a short prefix 58 58Introduction to Information Retrieval Dynamic summaries  Real estate on the search result page is limited Snippets must be short . . .  . . . but snippets must be long enough to be meaningful.  Snippets should communicate whether and how the document answers the query.  Ideally: linguistically wellformed snippets  Ideally: the snippet should answer the query, so we don’t have to look at the document.  Dynamic summaries are a big part of user happiness because . . .  . . .we can quickly scan them to find the relevant document we then click on.  . . . in many cases, we don’t have to click at all and save time. 59 59Introduction to Information Retrieval Resources  Chapter 8 of IIR  Resources at http://ifnlp.org/ir  The TREC home page – TREC had a huge impact on information retrieval evaluation.  Originator of Fmeasure: Keith van Rijsbergen  More on A/B testing  Too much A/B testing at Google  Tombros Sanderson 1998: one of the first papers on dynamic summaries  Google VP of Engineering on search quality evaluation at Google 60 60
Website URL
Comment