1 CS 430 / INFO 430 Information Retrieval Lecture 8 Query Refinement: Relevance Feedback Information Filtering
2 Course Administration
3 Query Refinement Search Reformulate query Display retrieved information new query reformulated query Query formulation EXIT
4 Reformulation of Query Manual Add or remove search terms Change Boolean operators Change wild cards Automatic Change the query vector: Remove/add search terms Change weighting of search terms
5 Manual Reformulation: Vocabulary Tools Feedback to user Information about stop lists, stemming, etc. Numbers of hits on each term or phrase Suggestions to user Thesaurus Browse lists of terms in the inverted index Controlled vocabulary
6 Manual Reformulation: Document Tools Feedback to user consists of document excerpts or surrogates Shows the user how the system has interpreted the query Effective at suggesting how to restrict a search Shows examples of false hits Less good at suggesting how to expand a search No examples of missed items
7 Relevance Feedback: Document Vectors as Points on a Surface Normalize all document vectors to be of length 1 Then the ends of the vectors all lie on a surface with unit radius For similar documents, we can represent parts of this surface as a flat region Similar document are represented as points that are close together on this surface
8 Relevance Feedback: Results of a Search x x x x x x x hits from search x documents found by search query
9 Relevance Feedback (Concept) x x x x o o o hits from original search x documents identified by user as non-relevant o documents identified by user as relevant original query reformulated query
10 Theoretically Best Query x x x x o o o optimal query x non-relevant documents o relevant documents o o o x x x x x x x x x x x x x x
11 Theoretically Best Query For a specific query, q, let: D R be the set of all relevant documents D N-R be the set of all non-relevant documents sim (q, D R ) be the mean similarity between query q and documents in D R sim (q, D N-R ) be the mean similarity between query q and documents in D N-R A theoretically best query would maximize: F = sim (q, D R ) - sim (q, D N-R )
12 Estimating the Best Query In practice, D R and D N-R are not known. (The objective is to find them.) However, the results of an initial query can be used to estimate sim (q, D R ) and sim (q, D N-R ).
13 Rocchio's Modified Query Modified query vector = Original query vector + Mean of relevant documents found by original query - Mean of non-relevant documents found by original query
14 Rocchio's Modified Query q 1 = q 0 + r i - s i i =1 n1n1 n1n1 1 n2n2 n2n2 1 q 0 = vector for the initial query q 1 = vector for the modified query r i = vector for relevant document i s i = vector for non-relevant document i n 1 = number of relevant documents n 2 = number of non-relevant documents
15 Difficulties with Relevance Feedback x x x x o o o optimal query x non-relevant documents o relevant documents original query reformulated query o o o x x x x x x x x x x x x x x Hits from the initial query are contained in the gray shaded area
16 Difficulties with Relevance Feedback x x x x o o o optimal results set x non-relevant documents o relevant documents original query reformulated query o o o x x x x x x x x x x x x x x What region provides the optimal results set?
17 Effectiveness of Relevance Feedback Best when: Relevant documents are tightly clustered (similarities are large) Similarities between relevant and non-relevant documents are small
18 When to Use Relevance Feedback Relevance feedback is most important when the user wishes to increase recall, i.e., it is important to find all relevant documents. Under these circumstances, users can be expected to put effort into searching: Formulate queries thoughtfully with many terms Review results carefully to provide feedback Iterate several times Combine automatic query enhancement with studies of thesauruses and other manual enhancements
19 Relevance Feedback: Clickthrough Data Relevance feedback methods have suffered from the unwillingness of users to provide feedback. Joachims and others have developed methods that use Clickthrough data from online searches. Concept: Suppose that a query delivers a set of hits to a user. If a user skips a link a and clicks on a link b ranked lower, then the user preference reflects rank(b) < rank(a).
20 Clickthrough Example Ranking Presented to User: 1. Kernel Machines 2. Support Vector Machine 3. SVM-Light Support Vector Machine light/ 4. An Introduction to Support Vector Machines 5. Support Vector Machine and Kernel... References Ranking: (3 < 2) and (4 < 2) User clicks on 1, 3 and 4 Joachims
21 Adjusting Parameters: Relevance Feedback q 1 = q 0 + r i - s i i =1 n1n1 n1n1 1 n2n2 n2n2 1 , and are weights that adjust the importance of the three vectors. If = 0, the weights provide positive feedback, by emphasizing the relevant documents in the initial set. If = 0, the weights provide negative feedback, by reducing the emphasis on the non-relevant documents in the initial set.
22 Adjusting Parameters by Weighting The modified query can be written: q 1 = w 1 e 1 + w 2 e w n e n where the e i are a basis for the term vector space of unit vectors corresponding to the terms in the word list and the w i are corresponding weights. If a query is used repeatedly, optimal values of the w i can be estimated using machine learning.
23 Task Application Information Information Agents: FilteringWhich news articles are interesting to a particular person? Text Routing Help-Desk Support: Who is an appropriate expert for a particular problem? Text Knowledge Management: Categorization Organizing a document database by semantic categories. Adjusting Parameters by Machine Learning: Tasks and Applications Joachims
24 Information Filtering d 1, d 2, d 3,... is a stream of incoming documents that are to be divided into two sets: R - documents judged relevant to an information need S - documents judged not relevant to the information need A query is defined as the vector in the term vector space: q = (w 1, w 2,..., w n ) where w i is the weight given to term i d j will be assigned to R if similarity(q, d j ) > What is the optimal query, i.e., the optimal values of the w i ?
25 Seeking Optimal Parameters Theoretical approach (not successful) Develop a theoretical model Derive parameters Test with users Heuristic approach (historically important) Develop a heuristic Vary parameters Test with users Machine learning (modern approach)
26 Information Filtering: Seeking Optimal Parameters using Machine Learning GENERAL:EXAMPLE: Text RetrievalInput: training examples queries with relevance judgments design space parameters of retrieval functionTraining: automatically find the solution find parameters so that many in design space that works well relevant documents are ranked on the training data highlyPrediction: predict well on new examples rank relevant documents high also for new queries
27 Learning to Rank Assume: distribution of queries P(q) distribution of target rankings for query P(r | q) Given: collection D of documents independent, identically distributed training sample (q i, r i ) Design: set of ranking functions F loss function l(r a, r b ) learning algorithm Goal: find f F that minimizes l(f (q), r) integrated across all queries
28 A Loss Function for Rankings For two orderings r a and r b, a pair is: concordant, if r a and r b agree in their ordering P = number of concordant pairs discordant, if r a and r b disagree in their ordering Q = number of discordant pairs Loss function: l(r a, r b ) = Q Example: r a = (a, c, d, b, e, f, g, h) r b = (a, b, c, d, e, f, g, h) The discordant pairs are: (c, b), (d, b) l(r a, r b ) = 2 Joachims
29 Machine Learning: Algorithms The choice of algorithms is a subject of active research, which is covered in several courses, notably CS 478 and CS/INFO 630. Some effective methods include: Naive Bayes Rocchio Algorithm C4.5 Decision Tree k-Nearest Neighbors Support Vector Machine