Query Operations: Automatic Local Analysis

Introduction Difficulty of formulating user queries –Insufficient knowledge of the collection –Insufficient knowledge of the retrieval environment Query reformulation –two basic steps query expansion –Expanding the original query with new terms term reweighting –Reweighting the terms in the expanded query

Automatic Relevance Feedback Basic idea –clustering: known relevant documents contain terms which can be used to describe a larger cluster of relevant documents. –obtain a description for a larger cluster of relevant documents automatically. identifying terms which are related to the query terms synonyms, stemming variations, terms which are close to the query terms in the text, etc. –global analysis vs. local analysis

Global vs. Local Analysis Global analysis –all documents in the collection are used to determine a global thesaurus-like structure which defines term relationships –this structure can be shown to the user who selects clusters or terms for query expansion Local analysis –the documents retrieved for a given query q are examined at query time to determine terms for query expansion –local clustering and local context analysis: without assistance from the user

Local Clustering Operate solely on the documents retrieved for the current query Valuable because term distributions are not uniform across topic areas –Distinguishing terms are different for different topics –Global techniques cannot take these differences into account. Requires significant run-time computation –Not for Web search engines due to cost –Useful in intranet environments and for specialized document collections

Local Clustering Initially use stemming to group terms –For stem s = polish –V(s) = {polish, polishing, polished} Definitions –q: query –D l : local document set (retrieved documents) –V l : vocabulary of D l –S l : set of distinct stems for V l Three types of clusters –association clusters –metric clusters –scalar clusters

Association Clusters Idea: Terms which co-occur frequently inside documents likely relate the same concept. Simple computation based on the frequency of co-occurrence of terms inside documents – correlation between the stems

Association Clusters Definitions –Matrix m has |S l | rows and |D l | columns m ij = f si,j (frequency of stem s i in document d j ) –Correlation c u,v is computed as: –Matrix s = mm t Unnormalized –s u,v = c u,v Normalized –s u,v = c u,v / (c u,u + c v,v – c u,v )

Association Clusters Selecting Clusters –Normally want stems for each query term –Need clusters to be small in order to retain focus –Select fixed size of clusters, n To expand query q –Construct cluster for each query term q Identify s q, the stem for query term q For stem s q, select the top n values s q,v –Union of all query term clusters is expanded query

Metric Clusters Two terms which are near one another are more likely to be correlated than two terms which occur far apart –factor in the distance between two terms in the computation of their correlation factor Same as Association Clusters except for the computation of c u,v

Metric Clusters Same as Association Clusters except for the computation of c u,v Correlation between the stems s u and s v Where r (k i, k j ) = distance between keywords in the same document This is unnormalized. Can be normalized.

Scalar Clusters Idea: two stems with similar neighborhoods have some synonymity relationship The relationship is indirect or induced by the neighborhood. Quantifying such neighborhood relationships –Arrange all correlation values s u,i in a vector –Arrange all correlation values s v,i in another vector –Compare these vectors through a scalar measure –The cosine of the angle between the two vectors is a popular scalar similarity measure.

Clustering Approaches In practice: –Metric clusters outperform association clusters –Using a combination of normalized and unnormalized correlation factors can be beneficial Unnormalized factors tend to group stems due to large frequencies Normalized factors tend to group stems which are more rare

Clustering Approaches Local approaches use the frequencies and correlations of terms and stems within the set of documents retrieved –These frequencies and correlations may not be representative of the overall collection –How is this good? How is this bad?

Query Operations: Automatic Global Analysis

Motivation Methods of local analysis extract information from local set of documents retrieved to expand the query An alternative is to expand the query using information from the whole set of documents Until the beginning of the 1990s these techniques failed to yield consistent improvements in retrieval performance Now, with moderns variants, sometimes based on a thesaurus, this perception has changed

Automatic Global Analysis There are two modern variants based on a thesaurus-like structure built using all documents in collection –Query Expansion based on a Similarity Thesaurus –Query Expansion based on a Statistical Thesaurus

Similarity Thesaurus The similarity thesaurus is based on term-to-term relationships rather than on a matrix of co- occurrence. –These relationships are not derived directly from co- occurrence of terms inside documents. –They are obtained by considering that the terms are concepts in a concept space. –In this concept space, each term is indexed by the documents in which it appears. Terms assume the original role of documents while documents are interpreted as indexing elements

Similarity Thesaurus vs. Vector Model The frequency factor: –In vector model f (i,j) = freq ( term k i in doc d j ) / freq ( most common term in d j ) –In similarity thesaurus f (i,j) = freq ( term k i in doc d j ) / freq ( doc where term k i appears most) –Normalized based on document where the term appears most. The inverse frequency factor: –In vector model Idf (i) = log (# of docs in collection / # of docs with term k i ) –In similarity thesaurus Itf (j) = log (# of terms in collection / # of terms in doc d j ) –Calculates how good a discriminator is this document?

Similarity Thesaurus Definitions: –t: number of terms in the collection –N: number of documents in the collection –f i,j : frequency of occurrence of the term k i in the document d j –t j : vocabulary of document d j –itf j : inverse term frequency for document d j Inverse term frequency for document d j For each term k i where w i,j is a weight associated between the term and the documents.

Similarity Thesaurus The relationship between two terms k u and k v is computed as a correlation factor c u,v given by The global similarity thesaurus is built through the computation of correlation factor C u,v for each pair of indexing terms [k u,k v ] in the collection The computation is expensive but only has to be computed once and can be updated incrementally

Query Expansion based on a Similarity Thesaurus Query expansion is done in three steps as follows:  Represent the query in the concept space used for representation of the index terms 2 Based on the global similarity thesaurus, compute a similarity sim(q,k v ) between each term k v correlated to the query terms and the whole query q. 3 Expand the query with the top r ranked terms according to sim(q,k v )

Statistical Thesaurus Global thesaurus is composed of classes which group correlated terms in the context of the whole collection –Such correlated terms can then be used to expand the original user query –These terms must be low frequency terms –However, it is difficult to cluster low frequency terms –To circumvent this problem, we cluster documents into classes instead and use the low frequency terms in these documents to define our thesaurus classes. –This algorithm must produce small and tight clusters.

Complete Link Algorithm Document clustering algorithm –Place each document in a distinct cluster. –Compute the similarity between all pairs of clusters. –Determine the pair of clusters [C u,C v ] with the highest inter-cluster similarity. –Merge the clusters C u and C v –Verify a stop criterion. If this criterion is not met then go back to step 2. –Return a hierarchy of clusters. Similarity between two clusters is defined as the minimum of similarities between all pair of inter-cluster documents –Use of minimum ensures small, focussed clusters

Generating the Thesaurus Given the document cluster hierarchy for the whole collection –Which clusters become classes? –Which terms represent classes? Answers based on three parameters specified by operator based on characteristics of the collection –TC: Threshold class –NDC: Number of documents in class –MIDF: Minimum inverse document frequency

Selecting Thesaurus Classes TC is the minimum similarity between two subclusters for the parent to be considered a class. –A high value makes classes smaller and more focussed. NDC is an upper limit on the size of clusters. –A low value of NDC restricts the selection to smaller, more focussed clusters

Picking Terms for Each Class Consider the set of documents in each class selected above Only the lower frequency terms are used for the thesaurus classes The parameter MIDF defines the minimum inverse document frequency for a term to represent the thesaurus class

Initializing TC, NDC, and MIDF TC depends on the collection –Inspection of the cluster hierarchy is almost always necessary for assisting with the setting of TC. –A high value of TC might yield classes with too few terms –A low value of TC yields too few classes NDC is easier to set once TC is set MIDF can be difficult to set

Query Expansion with Statisitcal Thesaurus Adding Terms: –Use the terms in same class at terms in query Weights of new terms can be based on both –the original query term weights (if any) and –on the degree to which a term represents the class of the query term

Conclusions Automatically generated thesaurus is a method to expand queries Thesaurus generation is expensive but it is executed only once Query expansion based on similarity thesaurus uses term frequencies to expand the query Query expansion based on statistical thesaurus uses document clustering and needs well defined parameters