Chapter 6 Queries and Interfaces

Keyword Queries n Simple, natural language queries were designed to enable everyone to search n Current search engines do not perform well (in general) with natural language queries n People trained (in effect) to use keywords  Compare average of about 2.3 words/Web query to average of 30 words/Community-based Question Answering (CQA) query n Keyword selection is not always easy  Query refinement techniques can help 2

Query-Based Stemming n Make decision about stemming at query time rather than during indexing  Improved flexibility and effectiveness n Query is expanded using word variants  Documents are not stemmed  e.g., “rock climbing” expanded with “climb”, not stemmed to “climb” 3

Stem Classes n Stemming generates stem classes n A stem class is the group of words that will be transformed into the same stem by the stemming algorithm  Generated by running stemmer on large corpus  e.g., Porter stemmer on TREC News on 3 stem classes with the first entry being the stem 4

Stem Classes n Stem classes are often too big and inaccurate n Modify using analysis of word co-occurrence n Assumption:  Word variants that could substitute for each other should co-occur often in documents 5

Modifying Stem Classes 6 n For all pairs of words in the stem classes, count how often they co-occur in text windows of W words, where W is in the range n Compute a co-occurrence or association metric for each pair, which measures the degree of association between the words. n Construct a graph where the vertices represent words and the edges are between words whose co-occurrence metric is above a threshold value, which is set empirically. n Find the connected components of the graph. These are the new stem classes.

Modifying Stem Classes n Dices’ Coefficient is an example of a term association measure , where n x is the number of windows (documents) containing x  a measure of the proportion of term co-occurrence n Two vertices are in the same connected component of a graph if there is a path between them  Forms word clusters n Sample output of modification 7

Context Sensitive Stemming According to [Peng 07]:  Although stemming increases recall, when considering Web searches, stemming lowers precision and requires significant additional computation.  Traditional stemming transforms all query terms, i.e., it always performs the same transformation for the same query word, regardless of the context in which the words appears.  Stemming by query expansion, on the other hand, handles each query differently, i.e., considers the context of each word in a query, which is more effective.  The Porter Stemmer algorithm does not considers the semantic meaning of words and thus can make mistakes when converting a word to its base form, e.g., “news” to “new”. 8 [Peng 07] F. Peng, et al. Context Sensitive Stemming for Web Search. In Proc. of Conf. on Research and Development in Information Retrieval

Spell Checking n Important part of query processing  10-15% of all Web queries have spelling errors n Errors include typical word processing errors but also many other types, (e.g., errors extracted from query logs) 9

Spell Checking n Basic approach: suggest corrections for words not found in spelling dictionary  Many spelling errors are related to websites, products, companies, people, etc. that are unlikely to be found n Suggestions found by comparing word to words in dictionary using similarity measure n Most common similarity measure is edit distance  Number of operations required to transform one word into the other 10

Edit Distance n Damerau-Levenshtein Distance  Counts the minimum number of insertions, deletions, substitutions, or transpositions of single characters required  e.g., Damerau-Levenshtein distance 1 (single-character errors)  Distance 2 11

Edit Distance n Different techniques used to speed up calculation of edit distances -- restrict to words that  start with same character (spelling errors rarely occur in the first letter)  come with similar length (spelling errors rarely occur on words with the same length)  sound the same (homophone, rules map words to codes) n Last option uses a (same) phonetic code to group words  e.g., Soundex 12

Soundex Code 13 (correct word may not always have the same Soundex code)

Spelling Correction Issues n Ranking corrections (> 1 possible corrections for an error)  “Did you mean...” feature requires accurate ranking of possible corrections (more likely: the best suggestion) n Context  Choosing right suggestion depends on context (other words)  e.g., lawers → lowers, lawyers, layers, lasers, lagers but trial lawers → trial lawyers n Run-on errors (word boundaries are skipped/mistyped)  e.g., “mainscourcebank”  Missing spaces can be considered another single character error in right framework 14

Noisy Channel Model n Address the issues of ranking, context, and run-on errors n User chooses word w based on probability distribution P(w)  Called the language model  Can capture context information about the frequency of occurrence of a word in text, e.g., P(w 1 | w 2 )  The probability of observing a word, given that another one has just been observed n User writes word w, but noisy channel causes word e to be written instead with probability P(e | w)  Called error model  Represents information about the frequency of spelling errors 15

Noisy Channel Model n Need to estimate probability of correction – to represent info. about the frequency of different types of errors  P(w | e) = P(e | w)P(w), i.e., the probability that given a written word e, the correct word is w n Estimate language model using context P(w) = λP(w) + (1 − λ)P(w | wp) where wp is a previous word of w, and λ is a parameter which specifies the relative importance of P(w) & P(w | wp) n Examples.  “fish tink”: “tank” and “think” both likely corrections, but 16 Error Model Probability of Word Occurrence P(tank | fish) > P(think | fish) Probability of Word Co-occurrence

Noisy Channel Model n Language model probabilities estimated using corpus and query log n Both simple and complex methods have been used for estimating error model  Simple approach: assume that all words with same edit distance have same probability, only edit distance 1 and 2 considered  More complex approach: incorporate estimates based on common typing errors Estimates are derived from large collections of text by finding many pairs of (in)correctly spelled words 17

18 Cucerzan and Brill (2004)’s Spell Checker using (i) query log (QL) and (ii) dictionary (D) 1. Tokenize the query. 2. For each token, a set of alternative words and pairs of words is found using an edit distance modified by weighting certain types of errors. The data structure that is searched for the alternatives contains words and pairs from QL and D. 3. The noisy channel model is used to select the best correction. 4. The process of looking for alternatives and finding the best correction is repeated until no better correction is found.

Query Expansion n A variety of (semi)-automatic query expansion techniques have been developed  Goal: improve effectiveness by matching related terms  Semi-automatic techniques require user interaction to select best expansion terms n Query suggestion is a related technique  Alternative queries, not necessarily more terms 19

Query Expansion n Approaches usually based on an analysis of term co- occurrence e ither in (i) the entire document collection, (ii) a large collection of queries, or (iii) the top-ranked documents in a result list  Query-based stemming also an expansion technique n Automatic expansion based on general thesaurus, but not effective  Does not take context into account 20

Term Association Measures n Key to effective expansion: choose appropriate words for the context of the query  e.g., “aquarium” for “tank” in “tropical fish tanks”, but not for “armor for tanks” n Dice’s Coefficient: computing the term-occurrence n Mutual Information: measure co-occurrence words 21 log P(a, b) P(a)P(b)P(a)P(b) rank  n a. n b n ab

Term Association Measures n Mutual Information measure favors low frequency terms n Expected Mutual Information Measure (EMIM) corrects the problem of MIM by weighting MIM values using P(a, b) where N is the number of documents in a collection  favoring high-frequency terms, which can be a problem 22

Term Association Measures n Pearson’s Chi-squared (χ 2 ) measure  Compares the number of co-occurrences of two words with the expected number of co-occurrences if the two words were independent  Normalizes this comparison by the expected number  Also limited form focused on word co-occurrence 23 independent co-occurrences

Association Measure Summary 24 Favor high- frequency terms Favor low- frequency terms

Association Measures n Associated words are of little use for expanding the query “tropical fish” n Expansion based on whole query takes context into account  e.g., using Dice with term “tropical fish” gives the following highly associated words: goldfish, reptile, aquarium, coral, frog, exotic, stripe, regent, pet, wet n Impractical for all possible queries, other approaches used to achieve this effect 25

Other Approaches n Pseudo-relevance feedback  Expansion terms based on top retrieved documents for initial query n Context vectors  Represent words by the co-occurred words e.g., top 35 most strongly associated words for “aquarium” (using Dice’s coefficient):  Rank words for a query by ranking context vectors 26

Other Approaches n Query logs  Best source of information about queries and related terms Short pieces of text and click data  e.g., most frequent words in queries containing “tropical fish” from MSN log: stores, pictures, live, sale, types, clipart, blue, freshwater, aquarium, supplies  Query suggestion based on finding similar queries Group based on click data 27

Query Expansion on Web Searches n According to [Crabtree 07] automatic query expansion (on Web searches)  Enhances recall but not precision: expanding queries using 20 terms improves the recall of the result set, but hurts the precision due to query drift.  Helps identifying other relevant pages by identifying additional terms that occur on relevant pages  Relies on both local and global document analysis (considering documents in the result set or using a thesauri/properties of the entire corpus, respectively) as sources for refinement terms 28 [Crabtree 07] D. Crabtree et al. Exploiting Underrepresented Query Aspects for Automatic Query Expansion. In Proc. of Conf. on Knowledge Discovery and Data Mining

Relevance Feedback n User identifies relevant (and maybe non-relevant) documents in the initial result list n System modifies query using terms from those documents and re-ranks documents  Example of simple ML algorithm using training data  But, very little training data n Pseudo-relevance feedback just assumes top-ranked documents are relevant – no user input 29

Relevance Feedback Example Top 10 documents for “tropical fish” 30 Snippets

Relevance Feedback Example n If we assume top 10 are relevant, most frequent terms are (with frequency): a (926), td (535), href (495), http (357), width (345), com (343), nbsp (316), www (260), tr (239), htm (233), class (225), jpg (221)  Too many stopwords and HTML expressions n Use only snippets and remove stopwords tropical (26), fish (28), aquarium (8), freshwater (5), breeding (4), information (3), species (3), tank (2), Badman’s (2), page (2), hobby (2), forums (2) 31

Relevance Feedback n Both relevance and pseudo-relevance feedback are effective, but not used in many applications  Pseudo-relevance feedback has reliability issues, especially with queries that don’t retrieve many relevant documents n Some applications use relevance feedback  Filtering, “more like this” n Query suggestion more popular  May be less accurate, but can work if initial query fails 32

Context and Personalization n If a query has the same words as another query, results will be the same regardless of  Who submitted the query  Why the query was submitted  Where the query was submitted  What other queries were submitted in the same session n These other factors (the context) could have a significant impact on relevance  Difficult to incorporate into ranking 33

User Models n Generate user profiles based on documents that the person looks at  Such as Web pages visited, messages, or word processing documents on the desktop n Modify user’s queries using words from his profile n Generally not effective  Imprecise profiles, information needs can change significantly 34

Query Logs n Query logs provide important contextual information that can be used effectively n Context in this case is  Previous queries that are the same  Previous queries that are similar  Query sessions including the same query n Query history for individuals could be used for caching 35

Local Search n Location is context n Local search uses geographic information to modify the ranking of search results  Location derived from the query text  Location of the device where the query originated n Examples  “Underworld 3 cape cod”  “Underworld 3” from mobile device in Hyannis 36

Local Search n Identify the geographic region associated with Web pages  Use location metadata that has been manually added to the document  Or identify locations such as place names, city names, or country names in text n Identify the geographic region associated with the query  10-15% of queries contain some location reference n Rank Web pages using location information in addition to text and link-based features 37

Snippet Generation n Query-dependent document summary n Simple summarization approach  Rank each sentence in a document using a significance factor  Select the top sentences for the summary  First proposed by Luhn in 50’s 38

Sentence Selection n Significance factor for a sentence is calculated based on the occurrence of significant words  If f d,w is the frequency of word w in document d, then w is a significant word if it is not a stopword, i.e., a high-frequency word, and where s d is the number of sentences in document d  Text is bracketed by significant words (limit on number of non-significant words in bracket) 39

Sentence Selection n Significance factor for bracketed text spans is computed by (i) dividing the square of the number of significant words in the span by (ii) the total number of words n Example. Significance factor = 4 2 /7 = The limit set for non-significant words in a bracket is 4

Snippet Generation n Involves more features than just significance factor n Example. (A sentence-based approach) on a news story 1. Whether the sentence is a heading 2. Whether it is the 1 st or 2 nd line of the document 3. The total number of query terms occurring in the sentence 4. The number of unique query terms in the sentence 5. The longest contiguous run of query words in the sentence 6. A density measure of query words (i.e., significance factor) n Weighted combination of features used to rank sentences 41

Blog Snippets n According to [Parapar 10], in performing a Blog search  Considering reader’s comments improves the quality of post snippets and thus enhances users’ access to relevant post in a result list  Indexing comments, as well as the post itself, enhances the recall of a blog search, which is more important in blog than web search, since searchers are usually interested in all recent posts about a specific topic) 42 [Parapar 10] J. Parapar et al. Blog Snippets: A Comments-Biased Approach. In Proc. of Intl. Conf. on Research and Development in Information Retrieval

Clustering Results n Result lists often contain documents related to different aspects of the query topic n Clustering is used to group related documents to simplify browsing Example clusters for query “tropical fish” 43

Result List Example Top 10 docs for “tropical fish” 44

Clustering Results n Requirements n Efficiency  Must be specific to each query and are based on the top-ranked documents for that query  Typically based on snippets n Easy to understand  Can be difficult to assign good labels to groups  Monothetic vs. polythetic classification 45

Types of Classification n Monothetic  Every member of a class has the property that defines the class  Typical assumption made by users  Easy to understand n Polythetic  Members of classes share many properties but there is no single defining property  Most clustering algorithms (e.g., K-means) produce this type of output 46

Classification Example  Possible monothetic classification  { D 1,D 2 } (labeled using a) and { D 2,D 3 } (labeled e)  Possible polythetic classification  { D 2,D 3,D 4 }, D 1  Labels? 47 D 1 = {a, b, c} D 2 = {a, d, e} D 3 = {d, e, f, g} D 4 = {f, g}

Result Clusters n Simple algorithm  Group based on words in snippets n Refinements  Use phrases  Use more features Whether phrases occurred in titles or snippets Length of the phrase Collection frequency of the phrase Overlap of the resulting clusters 48