System and method for generating queries

The system and method relate to data clustering systems and in particular to generating Boolean queries based on clusters.

Currently, there are a variety of systems that can cluster data based on search terms. K-means and bisecting k-means are examples of algorithms that can be used to cluster data. These methods of clustering tend to be very processing intensive and are difficult to implement where there are rapid changes in the data set that is to be clustered. These clustering methods are not used to generate Boolean queries that can search a data set faster and with fewer processing resources than using existing clustering methods.

For example, U.S. Pat. No. 5,862,519 defines a system for “blind” clustering. The system takes data and segments the data into clusters. This system does not generate Boolean queries based on the clustered data and does not run the Boolean queries against a data set.

U.S. Patent Application Publication 2002/0087579 discloses a system for clustering data into classifications. The system then shows the relationships between data. The relationships can be various types of relationships, including Boolean relationships. However, this system does not generate any Boolean queries based on the clustered classifications. Moreover, the system does not run Boolean queries against the clustered classifications.

The system and method are directed to solving these and other problems and disadvantages of the prior art. Each of a plurality of training documents comprises a training token(s). The plurality of training documents are clustered into a plurality of clusters based on at least one training token in the plurality of training documents. Each cluster contains at least one training document. A Boolean query(s) is generated for a cluster based on an occurrence of the at least one training token in a training document in the plurality of training documents. The system gets production data that comprises a plurality of production documents. Each of the plurality of production documents comprises a production token(s). The Boolean query(s) is then executed on the production data.

A further embodiment of the system gets the training data. The training data is cleaned. Salient token(s) are identified from the training token(s) in at least one of the plurality of training documents. The plurality of training documents are clustered into a plurality of clusters based on at least one training token or at least one salient token. Each cluster contains at least one training document. A Boolean query is generated for a cluster based on an occurrence of at least one salient token in a training document in the plurality of training documents. The system gets production data and executes the Boolean query on the production data.

These and other features and advantages of the system and method will become more apparent from considering the following description of an illustrative embodiment of the system and method together with the drawing, in which:

FIG. 1 is a block diagram of an illustrative system for generating and executing Boolean queries.

FIG. 2 is a flow chart of a method for generating and executing Boolean queries.

FIG. 3 is a flow chart of another method for generating and executing Boolean queries.

FIG. 4 is a flow chart of a method for weighting and ranking token(s)/salient token(s) in a cluster.

FIG. 5 is a diagram that illustrates a set of clusters and the distribution of training/salient tokens among them.

FIG. 6 is a table that illustrates the percentage of training documents that contain a training token/salient token in each cluster.

FIG. 7 is a table that illustrates the calculated token/cluster weight matrix for each cluster.

FIG. 8 is a table that illustrates the ranking of tokens for each cluster.

FIG. 9 is a flow chart of a method for generating Boolean queries using a Disjunctive Normal Form (DNF) algorithm.

FIG. 10 is a flow chart of a method for generating Boolean queries using an OR-BUT-NOT algorithm.

FIG. 11 is a flow chart of a method for analyzing the query results based on new training data, new production data, a request to modify a query, and/or a request to change the number of clusters.

FIG. 1 is a block diagram of an illustrative system 100 for generating and executing Boolean queries. The system 100 comprises a query management system 101 and a data source(s) 105 . The query management system 101 further comprises a data manager 102 , a clustering manager 103 , and a query manager 104 .

The data source(s) 105 can be any type of system capable of generating and/or receiving data. The data source 105 can be a call center, an email system, a social network, a web site, a video server, a database, and the like. The query management system 101 can be any type of system capable of processing data such as a server, a Personal Computer (PC), a Private Branch Exchange (PBX), and the like. Likewise, the data manager 102 , the clustering manager 103 , and the query manager 104 can be any type of system(s) capable of processing data.

The data in the data source(s) 105 can contain training data and/or production data. Training data and/or production data can come from multiple data sources 105 . Both the training data and the production data contain a plurality of documents. Each document of the plurality of documents in both the training data and the production data contains one or more tokens. The plurality of documents in both the training data and the production data can be any type of documents and/or information that can be compared such as emails, written documents, user input from a social network, video files, audio files, text transcripts of audio files, and the like. A document could be an individual email, groups of emails, a transcript of an individual call, a transcript of a group of calls, an entry in a social network, multiple entries in a social network, and the like.

Each document of both pluralities of training and production documents contains one or more tokens. A token can be any type of element that can be compared such as a text string, a unigram, a bi-gram, a trigram, identified video elements, identified audio elements, text strings generated from an audio file, and the like. For example, a token can be a video element, an identified video element, or any video object that may be identified and compared such as a car, a mountain, a person, and the like. Likewise, an audio element can be words from an audio file, a sound, and the like.

The data manager 102 gets training data from the data source(s) 105 . The data manager 102 can get the training data from a single or multiple data sources 105 . The training data contains a plurality of training documents. Each of the plurality of training documents contains at least one training token. The clustering manager 103 clusters the plurality of training documents into a plurality of clusters. The clustering can be based on a training token(s) in the plurality of training documents. For example, algorithms that can cluster documents based on token(s) include a k-means, a bisecting k-means, an agglomerative, or a divisive hierarchical clustering algorithm. A cluster contains at least one training document. A training document can be in more than one cluster.

The query manager 104 generates a Boolean query(s) for a cluster in the plurality of clusters based on an occurrence of the at least one training token in a clustered training document(s) in the cluster. The Boolean query(s) are generated for a cluster based on the occurrence or absence of the token in the plurality of training documents. The Boolean queries can be generated in a variety of ways such as a Disjunctive Normal Form (DNF) algorithm (see FIG. 9 ), an OR-BUT-NOT algorithm (see FIG. 10 ), and the like.

The data manager 102 gets production data from the data source(s) 105 . The data manager 102 can get the production data from the same data source 105 as the training data or from a different data source 105 . The production data may contain some or all of the training data. The data manager 102 can get the production data at any time, including before getting the training data. The production data contains a plurality of production documents. Each of the plurality of production documents contains at least one production token. The query manager then executes the Boolean query(s) on the plurality of production documents in the production data.

Boolean queries are executed on production tokens in the production data by determining if the token(s) in the Boolean query are in or not in the production documents. For example, if the generated Boolean query for Cluster 1 is Token 1 (the word “purchase”) and Token 2 (the word “appliance”), the results of the Boolean query (purchase AND appliance) will only return production documents that contain the production tokens (words) purchase and appliance. Likewise, if the query is Token 1 and NOT token 2 , the results of the Boolean query will return production documents that contain the production token purchase and do not contain the production token appliance. The production documents that are returned from the Boolean queries in this example are those that are most relevant to Cluster 1 . This is because the queries are generated from the training token(s) in cluster 1 .

FIG. 2 is a flow chart of a method for generating and executing Boolean queries. Illustratively, the query management system 101 , the data manager 102 , the clustering manager 103 , and the query manager 104 are a stored-program-controlled entity, such as a computer, which performs the methods of FIGS. 2-4 and 9 - 11 by executing a program stored in a storage medium, such as a memory or disk.

The process begins by getting 200 training data. The training data contains a plurality of training documents. Each of the plurality of training documents contains at least one training token. The plurality of training documents are clustered 201 into a plurality of clusters. The clustering is based on at least one training token in the plurality of training documents. Each cluster contains at least one training document. A Boolean query(s) is generated 202 for a cluster in the plurality of clusters based on an occurrence of a training token in the plurality of training documents. The process gets 203 production data. The production data contains a plurality of production documents. Each of the plurality of production documents contains at least one production token. The Boolean query(s) is executed 204 on the production documents in the production data.

FIG. 3 is a flow chart of an alternative method for generating and executing Boolean queries. The process begins by getting 300 training data. The training data contains a plurality of training documents. Each of the plurality of training documents contains at least one training token. The plurality of training documents in the training data are cleaned 301 . An example of cleaning 301 the plurality of training documents in the training data can be taking out word(s) (training token(s)) that occur very frequently in normal speech/text. These are words that don't carry useful information about the content of the text. Words such as “a”, “the”, “an”, “from” and “of” are types of frequent words that can be cleaned from the training data. Other examples of cleaning 301 the training data include deleting training tokens such as “um” and “uh” from a transcript of speech conversations for calls in a call center.

The process identifies 302 a salient token(s) from the training token(s) in the plurality of training documents. A salient token is a token that is relevant to the plurality of training documents to be analyzed. A salient token is a token that occurs more or less frequently in documents in the plurality of training documents in relation to documents in other document collections. Identifying 302 salient token(s) can be accomplished by comparing frequencies of individual words (unigrams), sequences of words (e.g., bi-grams, tri-grams), parts of words, spoken words, video objects, and the like in the plurality of training documents with a generic-domain corpus. A generic-domain corpus is a base line of how frequently words/phrases/objects are used in a spoken/written language and/or video. For example, “Word Frequencies in Written and Spoken English” (Andrew Wilson, Geoffery Leech, Paul Rayson, ISBN 0582-32007-0, Prentice Hall, 2001), which is incorporated herein by reference, defines an example generic-domain corpus. Identifying salient words (i.e. words that are characteristic of our corpus) can be based on known algorithms such as Relative Frequency Ratio (RFR). For example, “Fred J. Damerau. Generating and evaluating domain-oriented multi-word terms from texts. Information Processing and Management 29(4): 433-447, 1993.” discloses the use of RFR.

The plurality of training documents are clustered 303 into a plurality of clusters based on at least one training token in the plurality of training tokens and/or at least one salient token. A Boolean query(s) is generated 304 for a cluster based on occurrence of the salient token(s) in a training document in the plurality of training documents. The process gets 305 production data. The production data contains a plurality of production documents. Each of the plurality of production documents contains at least one production token. The Boolean query(s) is executed 306 on the plurality of production documents in the production data.

FIG. 4 is a flow chart of a method for weighting and ranking training token(s)/salient token(s) in a cluster. FIG. 4 is a flow diagram of steps that go between step 201 and step 202 in FIG. 2 or between step 303 and step 304 in FIG. 3 . After clustering 201 or 303 the plurality of training documents into a plurality of clusters based on the training token(s)/salient token(s), the process calculates 400 a token/cluster weight matrix for the training tokens/salient token(s) (see FIG. 7 for an example token/cluster weight matrix). The training token(s)/salient token(s) are ranked 401 for each cluster (see FIG. 8 for an example of ranked tokens for each cluster). A list of the top N token(s)/salient token(s) is selected 402 . N is a positive integer. N is usually determined based on the size and number of training documents. The process then goes to step 202 or step 304 .

FIG. 5 is a diagram that illustrates a set of clusters 500 - 502 and the distribution of training/salient tokens among them. FIG. 5 is an illustrative example of clusters 500 - 502 that are generated in steps 201 and 303 . FIG. 5 contains three clusters: cluster one 500 , cluster two 501 , and cluster three 502 . Training documents are placed in clusters using known clustering algorithms. In this example, cluster one 500 has training documents that contain at least one of the training tokens/salient tokens T 1 through T 4 . Cluster two 501 has training documents that contain training tokens/salient tokens T 3 or T 4 . Cluster three 502 has training documents that contain training tokens/salient tokens T 4 or T 5 . Training tokens/salient tokens T 1 -T 5 can be training tokens or they can be salient tokens depending on the method used.

FIG. 6 is a table 600 that illustrates the percentage of training documents that contain a training token/salient token in each cluster 500 - 502 . FIG. 6 is a table representation of the clusters 500 - 502 shown in FIG. 5 . Each column represents a cluster 500 - 502 . Each row represents a training token/salient token. The percentage number represents the percentage of training documents in the cluster 500 - 502 that contain the training token/salient token.

In this example, Token 1 (T 1 ) is in 50% of the training documents in cluster one 500 . T 1 is not in any of the training documents in cluster two 501 or cluster three 502 . Token 2 (T 2 ) is in 30% of the training documents in cluster one 500 . T 2 is not in any of the training documents in cluster two 501 or cluster three 502 . Token 3 (T 3 ) is in 20% of the training documents in cluster one 500 . T 3 is in 50% of the training documents in cluster two 501 . T 3 is not in any of the training documents in cluster three 502 . Token 4 (T 4 ) is in 10% of the training documents in cluster one 500 . T 4 is in 15% of the training documents in cluster two 501 . T 4 is in 30% of the training documents in cluster three 502 . Token 5 (T 5 ) is not in any of the training documents in cluster one 500 or cluster two 501 . T 5 is in 10% of the training documents in cluster three 502 .

FIG. 7 is a table 700 that illustrates the calculated 400 token/cluster weight matrix for each cluster 500 - 502 . The following equation (Equation 1) is used to calculate 400 the token/cluster weight matrix for each cluster using the data from Table 600 .

The weight (t,C j ) is the weight of token t for cluster j. DF (t,C j ) is the number of documents in cluster j that contain token t. N Cj is the total number of documents in cluster j. M (m) is the number of clusters. Equation 1 weights a token in a cluster by taking into account the descriptive and the discriminative power of the token for that cluster. Intuitively, a token is weighted higher if it appears in many documents in the cluster, but in only a few documents in the other clusters.

Using Equation 1, the weight of T 1 for cluster one 500 would be calculated 400 as: 0.5/(1+0+0), or 0.5. The weight of T 2 for cluster one 500 would be calculated 400 as: 0.3/(1+0+0), or 0.3. The weight of T 3 for cluster one 500 would be calculated 400 as: 0.2/(1+0.5+0), or 0.133. The weight of T 4 for cluster one 500 would be calculated 400 as: 0.1/(1+0.15+0.3), or 0.069. The weight for T 5 is 0. Likewise, the values for cluster two 501 and cluster three 502 are calculated 400 using Equation 1 and are shown in Table 700 .

FIG. 8 is a table 800 that illustrates the ranking 401 of token(s)/salient token(s) for each cluster 500 - 502 . The following equation (Equation 2) is used to rank 401 the training token(s)/salient token(s) for each cluster 500 - 502 .

T (t) is the best token for cluster j->max weight difference with the cluster with the highest weight. For each cluster only the N top N tokens with a positive final weight (using Equation 2) are chosen. Using Equation 2, the ranking 401 of the token(s)/salient token(s) is derived as follows:

T 1 C 1=min{(0.5−0),(0.5−0)}=0.5

C 2=min{(0−0.5),(0−0)}=−0.5

C 3=min{(0−0),(0−0.5)}=−0.5

T 2 C 1=min{(0.3−0),(0.3−0)}=0.3

C 2=min{(0−0.3),(0−0)}=−0.3

C 3=min{(0−0),(0−0.3)}=−0.3

T 3 C 1=min{(0.133−0.417),(0.133−0)}=−0.284

C 2=min{(0.417−0.133),(0.417−0)}=0.284

C 3=min{(0−0.133),(0−0.417)}=−0.417

Likewise, the ranking 401 of T 4 -T 5 is accomplished using Equation 2. The results of ranking 401 of T 1 -T 5 in clusters 500 - 502 is shown in Table 800 . The ranking 401 is organized from the highest number to the lowest number. A list of the top N number of training token(s)/salient token(s) are selected 402 for each cluster 500 - 502 . N is a positive integer.

FIG. 9 is a flow chart of a method for generating ( 202 or 304 ) Boolean queries using a Disjunctive Normal Form (DNF) algorithm. FIG. 9 is a flow diagram of either step 202 in FIG. 2 or step 304 in FIG. 3 . The process described in FIG. 9 can be used for each cluster 500 - 502 . The process begins by initializing 900 . Initializing 900 can include, but is not limited to, setting the query to an empty query and getting the top N training tokens/salient tokens (see Table 800 ) for the cluster 500 - 502 for which the Boolean query is being generated ( 202 or 304 ). The top N training tokens/salient tokens for each cluster and their negative (the NOT of the training token/salient token) are put into a list. The clause is set 901 to an empty clause. The clause is a holder for building a group of tokens that are ANDed together. The next best training token/salient token for the clause is selected 902 from the top N list of training tokens/salient tokens for the cluster.

The next best training token/salient token is selected 902 . For each token in the list of top N training token(s)/salient tokens(s) the next best training token/salient token is the one that maximizes the difference of true coverage and false coverage when ORed with the current query. True coverage is the number of correct documents retrieved. False coverage is the number in incorrect documents retrieved. The next best training token/salient token is selected by subtracting false coverage (Query ORed with the token) from the true coverage (Query ORed with the token).

If the recall improvement of ORing the training token/salient token is not greater than the minimum improvement in step 903 , the process goes to step 911 . Recall is the proportion of correct documents that are retrieved. Recall improvement is the recall of the query ORed with the training token/salient token minus the recall of the query. The minimum improvement is a threshold of improvement. If the recall improvement of ORing the training token/salient token is greater than the minimum improvement in step 903 , clause is set 904 to the training token/salient token and variable clause tokens is set to one. The variable clause tokens is the number of training tokens/salient tokens currently in the clause.

The process selects 905 the next best training token/salient token as described in step 902 . The next best training token/salient token is the one that maximizes the difference of true coverage and false coverage when ANDed with the current clause. This is the true coverage (query OR (clause AND token)) minus false coverage (query OR (clause AND token)). The process determines 906 if a precision improvement of ANDing the training token/salient token is greater than a recall drop. The recall drop and precision improvement can be calculated in various ways. For example, the recall drop can be the number of true documents that are missed by including the training token/salient token in the Boolean equation. The precision improvement can be the number of false documents that are eliminated by the Boolean query by including the training token/salient token (or NOT token) in the Boolean query. A true document is a document that is in the cluster (the document could be in other clusters). A false document is a document that is not in the cluster. The recall improvement is the true document(s) that are added relative to the true documents that are added by the training tokens/salient tokens that are already in the Boolean query.

As a second exemplary process, the recall drop can be calculated using the following process. The recall drop is the recall (query OR clause) minus recall (query OR (clause AND token)). The precision improvement is the precision (query OR (Clause AND token)) minus precision (query OR clause). If the precision improvement of ANDing the training token/salient token is not greater than the recall drop in step 906 , the process goes to step 910 .

If the precision improvement of ANDing the training token/salient token is greater than the recall drop in step 906 , the process sets 907 the clause equal to the clause ANDed with the training token/salient token. The process increments 908 the variable clause tokens.

The process determines 909 if the variable clause tokens is less than the variable max clause tokens. The variable max clause tokens is a predefined number that sets a limit on the number of training tokens/salient tokens that will be in a clause. If the variable clause tokens is less than max clause tokens in step 909 , the process goes to step 905 to select the next best training token/salient token. Otherwise, if the variable clause tokens is not less than the variable max clause tokens in step 909 , the query is set 910 equal to the query ORed with the clause. The variable num clauses is incremented 911 . The process determines 912 if the variable num clauses is less than the variable max clauses. Max clauses is a limit on the number of clauses allowed in the Boolean query. If the variable num clauses is less than max clauses, the process goes to step 901 so a new clause can be generated. Otherwise, if the variable num clauses is not less than the variable max clauses in step 912 , the process goes to step 203 or 305 .

As an example of using the method of FIG. 9 , assume the five ranked training tokens/salient tokens for cluster one 500 from Table 800 are the number (N) of ranked training tokens/salient tokens. Assume the following values: min improvement=10%, max clause tokens=2, and max clauses=2. As discussed above, the list of N training tokens/salient tokens (and token NOT) can be determined by subtracting the false coverage (negative training documents) from the precision improvement (positive training documents). Using FIG. 6 , the following calculations would occur for cluster one 500 .

The above calculations are based on percentages of documents. However, other methods can be used. For example, calculations based on the number of documents can be used. T 1 is calculated based the true coverage (50% of positive documents added) minus the false coverage (0% of negative documents added), or 0.5. NOT T 1 is calculated based on a true coverage of 50% (NOT T 1 adds 50% of the training documents) minus the false coverage of 100% (NOT T 1 adds 100% of the training documents in both C 2 and C 3 ), or −0.5. The same calculations are used for T 2 . T 3 is calculated based on the true coverage of 20% minus the false coverage of 50%. NOT T 3 is calculated based on the true coverage of 80% minus the false coverage of 100% (i.e. 100% of the training documents in cluster three 502 ). Likewise, the values for T 4 -T 5 are calculated using the same method. Note that the values calculated for T 4 and NOT T 4 contain a range of values because there can be overlap from C 2 to C 3 (i.e., the same training document may be in both clusters).

After initializing 900 and setting 901 the clause to an empty clause, the process selects 902 the highest ranked training token/salient token NOT T 5 . The process determines 903 that the recall improvement for NOT T 5 (100%) is greater than the minimum improvement (10%). The clause is set 904 to NOT T 5 and clause tokens is set to one. The next best training token/salient token (T 1 ) is selected 905 . The process determines 906 that the precision improvement for T 1 (100%) is greater than the recall drop for T 1 (50%). The clause (NOT T 5 ) is set 907 to the current clause value (NOT T 5 ) ANDed with the current token T 1 producing a clause value of (NOT T 5 AND T 1 ). Clause tokens is incremented 908 to two. The process determines 909 that clause tokens ( 2 ) is not less than max clause tokens ( 2 ). The query (NOT T 5 AND T 1 ) is set 910 to the query ORed with the clause ((NOT T 5 AND T 1 ) OR). Num clauses is incremented 911 to one. The process determines 912 that num clauses ( 1 ) is less than max clauses ( 2 ).

The process sets 901 the clause to an empty clause. The process selects 902 the next best training token/salient token T 2 . The process determines 903 that the recall improvement for T 2 is (30%) is greater than the minimum improvement (10%) for T 2 . The clause is set 904 to T 2 and clause tokens is set to one. The next best training token/salient token (NOT T 4 ) is selected 905 . The process determines that the precision improvement for NOT T 4 50% (assume 50% of the retrieved documents are true) is greater than the recall drop for NOT T 4 (10%). The clause is set 907 to clause (T 2 ) ANDed with the training token/salient token NOT T 4 (T 2 AND NOT T 4 ). Clause tokens is incremented 908 to two. The process determines 909 that clause tokens ( 2 ) is not less than max clause tokens ( 2 ). The query is set 910 to the query (NOT T 5 AND T 1 ) ORed with the clause (T 2 AND NOT T 4 ), or ((NOT T 5 AND T 1 ) OR (T 2 AND NOT T 4 )). Num clauses is incremented 911 to two. The process determines that num clauses ( 2 ) is not less than max clauses ( 2 ). The process goes to step 203 or 305 . The generated Boolean query is ((NOT T 5 AND T 1 ) OR (T 2 AND NOT T 4 )).

The example above discloses one way of determining 906 if the precision improvement is greater than the recall drop. In the example above, the precision improvement is based on the selected next best token from step 905 . However, the precision improvement in step 905 can be calculated in real time based on the current query.

FIG. 10 is a flow chart of a method for generating ( 202 or 304 ) Boolean queries using an OR-BUT-NOT algorithm. FIG. 10 is a flow diagram of either step 202 in FIG. 2 or step 304 in FIG. 3 . The process begins by initializing 1000 . Initializing 1000 can include, but is not limited to setting the query to an empty query, setting the variable num OR tokens to zero, setting the variable num AND not token to zero, and getting the top N training tokens/salient tokens (see FIG. 8 ) for the cluster 500 - 502 that the Boolean query is being generated ( 202 or 304 ). The next best positive training token/salient token is selected 1001 .

The process determines 1002 if the recall improvement is greater than the minimum improvement for the selected positive training token/salient token. If the recall improvement is not greater than the minimum improvement in step 1002 , the process goes to step 1004 . Otherwise, the query is set 1003 to the query ORed with the selected positive training token/salient token. The variable num OR tokens is incremented 1004 . Num OR tokens is the number of training tokens/salient tokens that have been processed from the list of N training tokens/salient tokens. The process determines 1005 if the variable num OR tokens is less than max OR tokens. Max OR tokens is the maximum number of training tokens/salient tokens that will be ORed in the generated Boolean query. If the num OR tokens is less than max OR tokens in step 1005 , the process goes to step 1001 to select the next best positive training token/salient token.

Otherwise, the next best negative training token/salient token is selected 1006 . The process determines 1007 if a precision improvement is greater than a recall drop for the selected negative training token/salient token. If the precision improvement is not greater than the recall drop in step 1007 , the process goes to step 1009 . Otherwise, if the precision improvement is greater than the recall drop in step 1007 , the query is set 1008 to the query ANDed with the selected next best negative training token/salient token. The variable num AND NOT token is incremented 1009 . Num AND NOT token is the number of negative training tokens/salient tokens that have been processed. The process determines 1010 if num AND NOT token is less than max AND NOT token. Max AND NOT tokens is the maximum number of NOT training tokens/salient tokens that will be processed. If num AND NOT tokens is less than max AND NOT tokens, the process goes to step 1006 and selects the next best negative training token/salient token. Otherwise, the process goes to step 203 or 305 .

As an example of using the method of FIG. 10 , assume the five ranked training tokens/salient tokens for cluster one 500 from Table 800 are the number (N) of ranked tokens. Assume the following values: min improvement=10%, max OR tokens=2, and max AND NOT tokens=2. As discussed above, the list of N training tokens/salient tokens (and token NOT) is determined by subtracting the false coverage (negative training documents) from the true coverage (positive training documents). Using FIG. 6 , the following calculations would occur for cluster one 500 for the positive and negative training tokens/salient tokens.

After initializing 1000 , the next best positive training token/salient token T 1 is selected 1001 . The process determines 1002 that the recall improvement for T 1 (50%) is greater than the minimum improvement (10%). The query is set 1003 to the query ORed selected positive token (T 1 ). Num OR tokens is incremented to one. The process determines 1005 that num OR tokens ( 1 ) is less than max OR tokens ( 2 ). The next best positive token T 2 is selected 1001 . The recall improvement for T 2 is the number of new training documents that will be added based on including T 2 in the Boolean query. The recall improvement for T 2 could be from 0% to 30% depending on how many new documents will be added by including T 2 . For example, if T 2 is not in any of the training document that contain T 1 , the recall improvement for T 2 would be 30%. If half of the documents that contain T 2 also contain T 1 , then the recall improvement for T 2 would be 15% because adding T 2 to the Boolean query would add 15% new training documents. For this example, assume that the recall improvement is 15%. The process determines 1002 that the recall improvement for T 2 (15%) is greater than the minimum recall improvement (10%). The query is set 1003 to the query ORed with the selected next best positive token (T 1 OR T 2 ). Num OR tokens is incremented 1004 to two. The process determines 1005 that num OR tokens ( 2 ) is not less than max OR tokens ( 2 ).

The next best negative token (NOT T 5 ) is selected 1006 . The process determines 1007 that the precision improvement of NOT T 5 50% (assume 50% of the retrieved documents are true) is greater than the recall drop (0%). The query is set 1008 to the query AND NOT T 5 , or ((T 1 OR T 2 ) AND NOT T 5 ). The variable num AND NOT tokens is incremented 1009 to one. The process determines 1010 that num AND NOT token ( 1 ) is less than max AND NOT tokens ( 2 ). The next best negative training token/salient token (NOT T 4 ) is selected 1006 . The process determines 1007 that the precision improvement for NOT T 4 50% (assume 50% of the retrieved documents are true) is greater than the recall drop (10%). The query is set 1008 to the query AND NOT T 4 , or (T 1 OR T 2 ) AND NOT T 5 AND NOT T 4 . The variable num AND NOT token is incremented 1009 to two. The process determines 1010 that num AND NOT token ( 2 ) is not less than max AND NOT tokens ( 2 ). The process goes to step 2003 or 305 . The generated query is (T 1 OR T 2 ) AND NOT T 5 AND NOT T 4 .

The example above discloses one way of determining 1007 if the precision improvement is greater than the recall drop. In the example above, the precision improvement is based on the selected next best negative token from step 1006 . However, the precision improvement in step 1007 can be calculated in real time based on the current query.

FIG. 11 is a flow chart of a method for analyzing the query results based on new training data, new production data, a request to modify a query, and/or a request to change the number of clusters. The process begins after executing 204 or 306 the Boolean query(s) on the plurality of production documents in the production data. The results of the Boolean query(s) are analyzed 1100 to determine if the Boolean query(s) are effective. The analyzing 1100 could be a manual or an automated process.

For example, if a manager determined that the types of calls in a call center have changed, then a new set of training data and production data may be required to generate the appropriate Boolean query(s) for the current state of the call center. There is a potential that some of the Boolean query(s) that were generated are not effective in selecting production documents that are pertinent to the cluster. The manager may want to delete one or more of the generated Boolean queries in order to get production documents that better match important topics in the production data. The number of clusters that are generated in steps 201 and 303 can be changed based on identifying the training documents in one or more clusters that have been previously generated and rerunning the process of FIG. 2 or 3 using the training documents in the cluster as the training data.

After analyzing 1100 the results of the Boolean queries, the process determines 1101 if there is new training data. If there is new training data in step 1101 , the process goes to step 200 or 300 . Otherwise, the process determines 1102 if a one or more of the Boolean queries needs to be modified. If the Boolean queries need to be modified, the process goes to step 204 or 306 . Otherwise, the process determines 1103 if the number of clusters needs to be changed. If the number of clusters needs to be changed, the process goes to step 201 or 303 . Otherwise, the process determines 1104 if there is new production data or new production data and one or more of the Boolean queries need to be modified. If there is new production data or new production data and one or more of the Boolean queries need to be modified, the process goes to step 203 or 305 . Otherwise, the process is done 1105 .

Of course, various changes and modifications to the illustrative embodiment described above will be apparent to those skilled in the art. For example, per clustering data (manually or automatically) to only add new training data that is from a certain class or adding key words to the queries, even though the terms don't show up in the training data. These changes and modifications can be made without departing from the spirit and the scope of the system and method and without diminishing its attendant advantages. It is therefore intended that such changes and modifications be covered by the following claims except insofar as limited by the prior art. In addition, in the claims, when referring to an element in the singular (e.g. a production token), the element should also be viewed to include the one or more of the element (e.g. one or more production tokens).