Retrieving and ranking of documents from database description

[0002] The present invention relates to retrieving and/or ranking of documents in a large database, and more particularly relates to a method, a computer system, and a program product for retrieving and ranking the documents in a very large database by dimension reduction of a document matrix using a covariance matrix.

[0003] Recently as database systems handle increasingly large amounts of data, such as, for example, news data, client information, stock data, etc, it becomes increasingly difficult for users of such databases to search for desired information quickly and effectively, with sufficient accuracy. Timely, accurate, and inexpensive detection of new topics and/or events from large databases may provide very valuable information for many types of businesses including, for example, stock control, futures and options trading, news agencies which may desire to quickly dispatch a reporter without needing to maintain a number of reporters posted worldwide, and businesses based on the internet or other fast paced actions which need to know of major and new information about competitors in order to succeed.

[0004] Conventionally, detection and tracking of new events in an enormous database is expensive, elaborate, and time consuming work, because generally a searcher of the database needs to hire extra persons for monitoring tasks.

[0005] Recent detection and tracking methods used for search engines usually use a vector model for data in the database in order to cluster the data. These conventional methods generally construct a vector q (kwd1, kwd2, . . . , kwdN) corresponding to the data in the database. The vector q is defined as the vector having the dimension equal to numbers of attributes, such as kwd1, kwd2, . . . kwdN which are attributed to the data. The most commonly used attributes are keywords, i.e., single keywords, phrases, names of person(s), place(s). Usually, a binary model is used to create the vector q mathematically in which the kwd1 is replaced by 0 when the data do not include the kwd1, and the kwd1 is replaced by 1 when the data include the kwd1. Sometimes, a weight factor is combined with the binary model to improve the accuracy of the search. Such weight factor includes, for example, the number of times the keywords occur in the data.

[0006] FIG. 1(a) and FIG. 1(b) show typical methods for diagonalization of a document matrix D which is comprised of the above described vectors where the matrix D is assumed to be an n-by-n symmetric definite positive matrix. As shown, the n-by-n matrix D may be diagonalized by two representative methods depending on the size of the matrix D. When n is relatively small in the n-by-n matrix D represented at 20, the method used may typically be Householder bidiagonalization and the matrix D is transformed to the bidaiagonalized form as shown at 22 in FIG. 1(a) followed by zero chasing of the bidiagonalized elements at 24 to construct the matrix Vr consisting of the eigenvectors of the matrix D at 26.

[0007] In FIG. 1(b) another method for the diagonalization is described, and the diagonalization method shown in FIG. 1(b) as represented at 30 may be effective when the number n of the n-by-n matrix D is large or medium. The diagonalization process first executes Lanczos tridiagonalization as shown in FIG. 1(b) at 32 followed by Sturm Sequencing at 34 to determine the eigenvalues wherein “r” denotes the rank of the reduced document matrix. The process next executes Inverse Iteration at 36 so as to determine the i-th eigenvectors associated to the eigenvalues previously found as shown in FIG. 1(b) as shown at 38.

[0008] In so far as the size of the database is still acceptable for application of precise and elaborate methods to complete computation of the eigenvectors of the document matrix D, the conventional methods are quite effective to retrieve and to rank the documents in the database. However, in a very large database, the computation time for retrieving and ranking of the documents is sometimes too long for a user of a search engine. There are also limitations for the resources of computer systems, such as CPU performance and memory capacities needed for completing the computation.

[0009] Therefore, there are needs for a system implemented with a novel method for stably retrieving and ranking the documents in very large databases in an inexpensive, automatic manner within acceptable computation time.

[0010] Some statistical approaches have been proposed using algorithms for information retrieval based on vector space models (see, for example, Baeza-Yates, R., Riberio-Neto, B., “Modern Information Retrieval”, Addition-Wesley, NY, 1999, and Manning, C. Schutze, H., “Foundations of Statistical Natural Language Processing”, MIT Press, Cambridge, Mass., 1999).

[0011] Salton, G. et al., “The SMART Retrieval System—Experiments in Automatic Document Processing”, Prentice-Hall, Englewood Cliffs, N.J., 1971, have reviewed the vector space model. They modeled the documents using vectors in which each coordinate of the vectors represents an attribute of the vectors, e.g., a keyword. In binary models of the vector, a coordinate takes on the value unity when the corresponding attribute is present in the documents and zero when the attribute is absent from the document. More sophisticated document vector models take into account weighting of the keyword such as frequency and location of appearance, e.g., in the title, section header, or abstract.

[0012] Queries are also modeled as vectors in the same manner as described for the documents. For a given user input query, the relevancy of a particular document is computed by determining the “distance” between the query and each of the document vectors. Although a number of different kinds of norms may be used to determine the “distance” between the query vector and the document vector, the angle between the query and the document vector derived from a scalar product is used as the most common procedure to determine the distance therebetween.

[0013] U.S. Pat. No. 4,839,853 issued to Deerwester et al., entitled “Computer information retrieval using latent semantic structure”, and Deerwester et. al., “Indexing by latent semantic analysis”, Journal of the American Society for Information Science, Vol. 41, No. 6, 1990, pp. 391-407 disclose a unique method for retrieving the document from the database. The disclosed procedure is roughly reviewed as follows;

[0014] Step 1: Vector space modeling of documents and their attributes.

[0015] In latent semantic indexing, or LSI, the documents are modeled by vectors in the same way as in Salton's vector space model. In the LSI method, the relationship between the query and the documents in the database are represented by an m-by-n matrix MN, the entries are represented by mn (i, j), i.e.,

[0016] MN=[mn(i,j)].

[0017] In other words, the rows of the matrix MN are vectors which represent each document in the database.

[0018] Step 2: Reducing the Dimension of the Ranking Problem via Singular Value Decomposition.

[0019] The next step of the LSI method executes singular value decomposition, or SVD of the matrix MN. Noises in the matrix MN are reduced by constructing a modified matrix A_kfrom the k-th largest singular values _i, wherein i=1, 2, 3, . . . , k, . . . and their corresponding eigenvectors are derived from the following relation;

[0020] MN_k=U_kΣ_kV_k^T,

[0021] wherein Σ is a diagonal matrix with monotonically decreasing diagonal elements of _i. The matrices U_kand V_kare the matrices whose columns are left and right singular vectors of the k-th largest singular values of the matrix MN.

[0022] Step 3: Query Processing.

[0023] Processing of the query in LSI-based Information Retrieval comprises two further steps: (1) query projection followed by (2) matching. In the query projection step, input queries are mapped to pseudo-documents in the reduced query-document space by the matrix U_k, and then are weighted by the corresponding singular values _ifrom the reduced rank and singular matrix Σ. This process may be described mathematically as follows;

[0024] q−^hat{q}=q^TU_kΣ_k^{−1}

[0025] wherein q represent the original query vector, ^hat{q} represents a pseudo-document vector, q^Trepresents the transpose of q, and {−1} represents the inverse operator. In the second step, similarities between the pseudo-document ^hat{q} and the documents in the reduced term document space V_k^Tare computed using any one of many similar measures.

[0026] Although there are many conventional methods for retrieving and ranking the document as described above, the inventors of the present invention have long sought a novel method for retrieving and ranking the documents in very large databases effectively and quickly, with sufficient accuracy.

[0027] The present invention was essentially made by finding that the eigenvectors of the covariance matrix K having the largest eigenvalue represents the most predominant feature, and the eigenvector of the covariance matrix having the second largest eigenvalue represents the second most significant feature, and so on. Therefore, it is effective to use a certain small set of the eigenvectors of the covariance matrix for dimension reduction of the document matrix D.

[0028] In the present invention, to meet a user input query, the dimension of the document matrix D is reduced as follows:

[0029] (1) compute the j-th largest eigenvalues of the covariance matrix K and their corresponding eigenvectors v (Dj) first

[0030] └d(j); j=1,2,3, . . . ,

[0031] (2) compute the k-th dimensional subspace for documents d(i), which spanned by the k eigenvectors corresponding to the j-th largest eigenvalues of D as follows; $d\left(i\right)=\underset{\left(i,j\right)}{\mathrm{sum}}c\left(i,j\right)v\left(Dj\right),$

[0032] where i and j denote the respective indexes for documents and eigenvectors, and c denotes corresponding coefficients,

[0033] (3) project the user input query vector onto the k-th dimensional subspace defined by the eigenvectors which correspond to the j-th largest eigenvalues, and

[0034] (4) rank the relevancy of each document with respect to the user-input query by computing the distance therebetween.

[0035] Therefore, according to a first aspect of the present invention, a method for retrieving and/or ranking documents in a database, documents being added to said database, and including attribute data may be provided. The method comprises steps of;

[0036] providing a document matrix derived from said documents, said matrix including numerical elements derived from said attribute data;

[0037] providing a covariance matrix derived from said document matrix;

[0038] executing singular value decomposition of said covariance matrix so as to obtain the following formula;

[0039] K=V·Σ·V^T,

[0040] wherein K represents said covariance matrix, V represents the matrix consisting of eigenvectors, Σ represents a diagonal matrix, and V^Trepresents a transpose of the matrix V;

[0041] K represents said covariance matrix, V represents the matrix consisting of eigenvectors, © represents a diagonal matrix, and V^Trepresents a transpose of the matrix V;

[0042] reducing a dimension of said matrix V using predetermined numbers of eigenvectors included in said matrix V, said eigenvectors including an eigenvector corresponding to the largest singular value;

[0043] reducing a dimension of said document matrix using said dimension reduced matrix V; and

[0044] retrieving and/or ranking said documents in said database by computing the scalar product of said dimension reduced document matrix and a query vector.

[0045] According to the first aspect of the present invention, said attributes include at least one keyword and a time stamp.

[0046] According to the first aspect of the present invention, said covariance matrix may be computed by the following formula;

[0047] K=B−X_bar·X_bar^T

[0048] wherein K represents the covariance matrix, B represents a momentum matrix, X_barrepresents a mean vector and X_bar^Trepresents a transpose thereof.

[0049] According to the first aspect of the present invention, said predetermined numbers may be 15-25% of the total of the eigenvectors of said covariance matrix.

[0050] According to the first aspect of the present invention, the method further includes a switching step, from dimension reduction using said document matrix directly to dimension reduction using said covariance matrix, depending on predetermined computation time such that said dimension reduction using said covariance matrix is executed when said dimension reduction of said document matrix using eigenvectors thereof computed from said document matrix is not completed within said predetermined computation time.

[0051] According to a second aspect of the present invention, a computer system for executing a method for retrieving and/or ranking documents in a database, documents being added to said database, and including attribute data, may be provided. The computer system executes the method comprising steps of;

[0052] providing a document matrix from said documents, said matrix including numerical elements derived from said attribute data;

[0053] providing a covariance matrix derived from said document matrix;

[0054] executing singular value decomposition of said covariance matrix so as to obtain the following formula;

[0055] K=V·Σ·V^T,

[0056] wherein K represents said covariance matrix, V represents the matrix consisting of eigenvectors, Σ represents a diagonal matrix, and V^Trepresents a transpose of the matrix V;

[0057] reducing a dimension of said matrix V using predetermined numbers of eigenvectors included in said matrix V, said eigenvectors including an eigenvector corresponding to the largest singular value;

[0058] reducing a dimension of said document matrix using said dimension reduced matrix V; and

[0059] retrieving and/or ranking said documents in said database by computing the scalar product of said dimension reduced document matrix and a query vector.

[0060] According to the second aspect of the present invention, said attributes include at least one keyword and a time stamp.

[0061] According to the second aspect of the present invention, said covariance matrix may be computed by the following formula;

[0062] K=B−X_bar·X_bar^T

[0063] wherein K represents a covariance matrix, B represents a momentum matrix, X_barrepresents a mean vector and X_bar^Trepresents a transpose thereof

[0064] According to the second aspect of the present invention, said predetermined numbers are 15-25% of the total of the eigenvectors of said covariance matrix.

[0065] According to the second aspect of the present invention, said method further may include a switching step, from dimension reduction using said document matrix directly to dimension reduction using said covariance matrix, depending on predetermined computation time so that said dimension reduction using said covariance matrix is executed when said dimension reduction of said document matrix using eigenvectors thereof computed from said document matrix is not completed within said predetermined computation time.

[0066] According to a third aspect of the present invention, a program product including a computer readable computer program for executing a method for retrieving and/or ranking documents in a database, documents being added to said database, and including attribute data, may be provided. The method comprises steps of:

[0067] providing a document matrix derived from said documents, said matrix including numerical elements derived from said attribute data;

[0068] providing a covariance matrix from said document matrix;

[0069] executing singular value decomposition of said covariance matrix so as to obtain the following formula;

[0070] K=V·Σ·V^T

[0071] wherein K represents said covariance matrix, V represents the matrix consisting of eigenvectors, Σ represents a diagonal matrix, and V^Trepresents a transpose off the matrix V;

[0072] reducing a dimension of said matrix V using predetermined numbers of eigenvectors included in said matrix V said eigenvectors including an eigenvector corresponding to the largest singular value;

[0073] reducing a dimension of said document matrix using said dimension reduced matrix V; and

[0074] retrieving and/or ranking said documents in said database by computing the scalar product of said dimension reduced document matrix and a query vector.

[0075] According to the third aspect of the present invention, said covariance matrix may be computed by the following formula;

[0076] K=B−X_bar·X_bar^T

[0077] wherein K represents the covariance matrix, B represents a momentum matrix, X_barrepresents a mean vector and X_bar^Trepresents a transpose thereof

[0078] According to the third aspect of the present invention, said predetermined numbers may be 15-25% of the total of the eigenvectors of said covariance matrix.

[0079] According to the third aspect of the present invention, said method may further include a switching step, from dimension reduction using said document matrix directly to dimension reduction using said covariance matrix, depending on predetermined computation time so that said dimension reduction using said covariance matrix is executed when said dimension reduction of said document matrix using eigenvectors thereof computed from said document matrix is not completed within said predetermined computation time.

[0080] The present invention will be further understood by explaining the following non-limiting embodiments of the present invention along with drawings thereof wherein the drawings are described as follows:

[0081] FIG. 1(a) and FIG. 1(b) illustrate representative methods conventionally used to diagonalize matrixes.

[0082] FIG. 2 is a flowchart of a method according to the present invention.

[0083] FIG. 3 is a schematic construction of a document matrix.

[0084] FIG. 4(a) and FIG. 4(b) illustrate schematic procedures for forming the document matrix, and for formatting thereof, respectively.

[0085] FIG. 5 is a flowchart for computing a covariance matrix.

[0086] FIG. 6(a) and FIG. 6(b) illustrate schematic constructions of the transpose of the document matrix and a mean vector, respectively.

[0087] FIG. 7(a) and FIG. 7(b) illustrate a schematic procedure for dimension reduction using the covariance matrix.

[0088] FIG. 8(a) and FIG. 8(b) illustrate a detailed procedure for dimension reduction using the covariance matrix according to the present invention.

[0089] FIG. 9 is a flowchart of another embodiment of the method according the present invention.

[0090] FIG. 10 shows a representative computer system according to the present invention.

[0091] 1. General Procedure for Retrieving and Ranking of Documents

[0092] FIG. 2 is a flowchart of the method according to the present invention. The method starts at step 201, and proceeds to step 202 and creates the document matrix D (m-by-n matrix) from the keywords included in the documents. It may be possible to use time stamps simultaneously for creating the document matrix D such as time, date, month, year, and any combination thereof.

[0093] The method then proceeds to step 203 and calculates mean vectors X_barof the document vectors. The method proceeds to step 204 and computes the momentum matrix B−D^T·D/n, wherein B denotes the momentum matrix, and D^Tdenotes the transpose of the document matrix D. The method proceeds to step 205 and then computes the covariance matrix K according to the following formula:

[0094] K=B−X_bar·X_bar^T,

[0095] wherein X_bar^Tdenotes the transpose of the mean vector X_bar.

[0096] The method according to the present invention thereafter proceeds to step 206 and executes singular value decomposition of the covariance matrix K as follows;

[0097] K=V·Σ·V^T,

[0098] where the rank of the covariance matrix K, i.e., rank (K), is r.

[0099] The method then proceeds to step 207 and executes dimension reduction of the matrix V such that a predetermined numbers k of the eigenvectors corresponding to the eigenvectors having the largest top 15-25% singular value may be included so as to create the dimension reduced matrix V_k. The method thereafter proceeds to step 208 and executes reduction of the document matrix using the dimension reduced V_kin order to provide the dimension reduced document matrix, i.e., the document subspace used to conduct retrieving and ranking of the document with respect to the query vector such as the Doc/Kwd query search, New Event Detection and Tracking as described in step 209. Hereafter, the essential steps of the present invention will be discussed in detail.

[0100] 2. Creation of Document Matrix

[0101] FIG. 3 shows an example of the document matrix D. The matrix D comprises rows from document 1 (doc 1) to document n (doc n) which include elements derived from the keywords (kwd 1, . . . , kwd n) included in the particular document. The numbers of documents and the number of keywords are not limited in the present invention, and depend on the documents and size of the database. In FIG. 3, the elements of the document matrix D are represented by the numerals 1. However, other positive real numbers may be used as when weighting factors are used to create the document matrix D.

[0102] In FIG. 4(a) and FIG. 4(b), an actual procedure for forming the document matrix is shown. In FIG. 4(a), a document written under SGML format is assumed. The method of the present invention generates keywords based on the document with which retrieval and ranking are executed and then converts the format of the document into another format, such as, for example, shown in FIG. 4 (b), suitable for use in the method according to the present invention. Formats of the documents are not limited to SGML, and other formats may be used in the present invention.

[0103] A procedure of the attributes generation in FIG. 4(a) is described. For example, attributes are considered to be keywords. Keywords generation may be performed as follows:

[0104] (1) Extract words with capital letter

[0105] (2) Ordering

[0106] (3) Calculate number of occurrence; n

[0107] (4) Remove word if n >Max or n< Min,

[0108] (5) Remove stop-words (e.g., The, A, And, There)

[0109] wherein Max denotes a predetermined value for maximum occurrence per keyword, and Min denotes a predetermined value for minimum occurrence per keyword. The process listed in (4) may often be effective to improve accuracy. There is no substantial limitation on the order of executing the above procedures, and the order of the above process may be determined considering system conditions in use, and programming facilities. This is one example of a keyword generation procedure, and there may be many other procedures which it is possible to use in the present invention.

[0110] After generating the keywords and converting the SGML format, the document matrix thus built is shown in FIG. 3. A sample pseudo code for creating the document vector/matrix by the binary models without using a weighting factor and/or function is as follows:

[0111] REM: No Weighting factor and/or function

[0112] If kwd (j) appears in doc (i)

[0113] Then M(i,j)=1

[0114] Otherwise M(i,j)=0

[0115] A similar procedure may be applied to the time stamps when the time stamps are simultaneously used.

[0116] The present invention may use a weighting factor and/or a weighting function with respect to both the keywords and the time stamps when the document matrix D is created. The weight factor and/or the weight function for the keyword W_Kmay include the number of times of occurrence of the keywords in the document, a position of the keyword in the document, whether or not the keyword is capitalized, but is not limited thereto. A weighting factor and/or weighting function W_Tfor the time stamp may also be applied to obtain the time/date stamp as well as the keyword according to the present invention.

[0117] 3. Creation of the Covariance Matrix and the Dimension Reduction of Matrix V

[0118] The creation of the covariance matrix comprises generally three steps as shown in FIG. 5, that is, step 502 for computing mean vectors X_bar, step 503 for computing the momentum matrix, and step 504 for computing the covariance matrix. FIGS. 6(a) and 6(b) show the details of the procedures described in FIG. 5. The mean vectors, X_bar, are computed by adding the elements in each of the rows of the transpose of the document matrix D as shown in FIG. 6(a) and dividing the sum of the elements by the document number, i.e., n. The construction of the mean vector X_bar, is shown in FIG. 6(b), where the transpose of the document matrix D^Thas n-by-m elements and X_barcomprises only one column vector consisting of the mean values of the elements in the same row of A^T.

[0119] In step 503, the momentum matrix B is calculated by the following formula:

[0120] B=D^T·D/n

[0121] wherein D denote the document matrix and the D^Tis the transpose thereof. Next the procedure proceeds to step 504 and computes the covariance matrix K which may be computed by the following formula using the mean vector X_barand the momentum matrix B:

[0122] K=B−X_bar·X_bar^T

[0123] The resulting covariance matrix K is a positive definite m-by-m structure. Thus, singular value decomposition thereof is straightforwardly computed by a conventional method. The structure of the covariance matrix K after the singular value decomposition is shown in FIG. 7(a) where the covariance matrix K is represented by the following formula:

[0124] K=V·Σ·V^T

[0125] where V is a matrix comprising eigenvectors, Σ is a diagonal matrix, and V^Tis the transpose of the matrix V. In FIG. 7, the eigenvectors are represented by dashed lines.

[0126] The dimension reduction of the matrix V may be performed such that predetermined numbers k, as shown in FIG. 7(b), of the eigenvectors including the eigenvectors corresponding to the largest singular value, is selected to construct k-by-m matrix V_k. According to the present invention, the selection of the eigenvectors may be performed in various ways, as long as the eigenvector corresponding to the largest top singular value may be included. There is no substantial limitation on the numeric value k. However, the integer value k may preferably be set to about 15-25% of the total number of the eigenvectors so that retrieving and ranking of the documents in the database may be significantly improved. When the integer value k is too small, accuracy of the search may decrease, and when the integer value k is too large, the advantage of the present invention may be lost.

[0127] 4. Dimension Reduction of the Document Matrix

[0128] Next the method according to the present invention executes dimension reduction of the document matrix using the matrix V_k. The dimension reduction of the document matrix is shown in FIG. 8(a) and FIG. 8(b). The dimension reduced matrix ^hatD of the document matrix ^hatD, is simply computed by producing the document matrix D and the matrix V_kas shown in FIG. 8(a). It may be possible to add some weighting to the dimension reduced matrix ^hatD using the weighting matrix with k-by-k elements as shown in FIG. 8(b). The computed matrix ^hatD has k-by-n elements, and comprises relatively significant features associated with the keywords. Therefore, the retrieving and ranking of the documents in the database may be significantly improved in response to the input query provided by a user of a search engine.

[0129] FIG. 9 shows a flowchart of another embodiment for retrieving and ranking the documents in the very large database. The embodiment shown in FIG. 9 switches the dimension reduction method from a conventional method in which the document matrix D is subjected to the direct dimension reduction procedure which requires long computational time and huge hardware resources. In the embodiment described in FIG. 9, the step for determining computational time for the dimension reduction 904 is provided so as to determine whether or not the dimension reducing procedure according to the present invention is suitable for the documents in the subject database.

[0130] As shown in FIG. 9, the second embodiment of the present invention begins at step 901, and proceeds to step 902 to execute creation of the document matrix using the binary model. The procedure further proceeds to step 903 for executing a conventional dimension reduction using the document matrix directly. The second embodiment measures the computational time of the dimension reduction of the document matrix by suitable means such as a timer at step 904. If the computation time exceeds a predetermined time T (No), then the procedure diverts to step 905 for switching the method for dimension reduction to the method according to the present invention and the method proceeds to step 906 for executing KWD/query search etc. If, at step 904, the computation time does not exceed the predetermined time T, (Yes), the procedure proceeds to step 906 and then KWD/query search etc. may be executed so as to provide the search result.

[0131] 5. Computer System

[0132] Referring to FIG. 10, a representative embodiment of the computer system according to the present invention is described. The computer system according to the present invention may include a stand alone computer system, a client-server system communicating through a LAN/WAN with any conventional protocols, or a computer system including communication through an internet infrabase. In FIG. 10, the representative computer system effective in the present invention is described using client-server systems.

[0133] The computer system shown in FIG. 10 comprises at least one client computer and a server host computer. The client computer and the server host computer communicate through a TCP/IP communication protocol. However, any other communication protocols may be used in the present invention. As described in FIG. 10, the client computer issues a request 1 to the server host computer to carry out retrieving and ranking of the documents stored in memory means of the server host computer.

[0134] The server host computer executes retrieving and ranking of the documents in the database depending on the request from the client computer. A result of the detection and/or tracking is then downloaded by the client computer from the server host computer through the server stub so as to be used by a user of the client computer. In FIG. 10, the server host computer is described as the Web server, but is not limited thereto; server hosts of any other types may be used in the present invention so long as the computer systems provides the above described function.

[0135] Hereunder, the present invention will be further explained by non-limiting examples thereof.

[0136] The method according to the present invention was examined using the Reuters-21578 news database including over 20,000 articles available at: :http://www.research.att.com/˜lewis. The documents in the database were retrieved by the method according to the present invention. A keyword list was generated by looking for words which contain an uppercase letter (capital letter) in any position in the keyword. Although the inventors used simple keyword models for the explained examples, there may be better models or keyword sets generated by other procedures. However, as described below, such simple models may provide sufficient results in retrieving and ranking the documents in the database. In keyword generation, words with too frequent or too infrequent appearances were eliminated from the keyword list to improve the accuracy.

[0137] The experiments were conducted with two different sets of retrieval and ranking using the method according to the present invention (examples) and the LSI (Latent Semantic Indexing) method for each query as comparable examples. The two different sets are as follows;

[0138] Set 1: A mini-database including 83 articles (documents) from the larger Reuters database was constructed to examine the method according to the present invention. The mini-database was small enough to know what the results to be retrieved by the input query should be.

[0139] Set 2: Full Reuters-21578 database was used to examine the method according to the present invention. The sample database including 83 documents is listed in table I.

[0140] In Example 1, two methods, i.e., the method according to the present invention and the LSI method are compared. The data set on which the examination was performed was the small database constructed from the Reuter-21578 news database. The experiment was started from a 152 dimensional space, and the number of dimensions was reduced in both methods to a 30 dimensional space.

[0141] In the experiment, one given document included in the database was input as a query for retrieving and ranking. The input queries, used to retrieve the documents, therefore correspond to the first result (1) in each list of the tables. For each query, the result obtained by the LSI method is first presented as comparable examples followed by the result obtained by the method according to the present invention which uses a covariance matrix as examples. The sign “XXX” placed in front of a retrieved document indicates that the document is irrelevant to the input query. The results are listed as sets ordered as follows:

[0142] sign of error; order of document retrieved; relevancy; date of documents; document number; title of documents. The results are shown in Table II-Table VI together with comments on the accuracy.

[0143] In the experiment using the second query, 2 documents were retrieved as errors in the method according to the present invention. However, the relevancy in the parenthesis showed significant cut off as described in Table III. Therefore, it is expected such errors may be eliminated by incorporating further refined keyword sets. In the experiment using the third query as described in Table IV documents 4, 7, and 11 were retrieved in the first rank. Documents 4, 7, and 11 are related to earthquakes which occurred in New Zealand. Although the method according to the present invention grouped together other documents without relation to the earthquake in New Zealand, the method according to the present invention may provide higher relative relevancy (shown in the parenthesis) together with the ranking thereof when compared with the LSI method. In the experiment described in Table VI, the gap between Macdonnel planes and other documents is well defined with high relevancy in the method according to the present invention, whereas the comparable examples shown in Table VI are interlaced with each other because of poor relevancy correlations.

[0144] In Example 2, the same experimental procedure was used to retrieve and rank the documents except for the database. The database used in Example 2 was the entire set included in the Reuter-21578 news database. The 7100 keywords were selected from the words included in the document “19870304 1534 leaf disease hits sri lanka rubber”. The results are shown from Table VII to Table X. The input query for each experiment was the first document listed in each of the corresponding Tables.

[0145] As shown in Table VII, the comparable example using the LSI method finds 1 document relevant to bahia cocoa, while the method according to the present invention retrieves 2 documents relating to bahia cocoa. The method according to the present invention further succeeded in finding another document related to cocoa.

[0146] As shown in Table II-Table X, the method according to the present invention exhibited better results for retrieving and ranking the documents than the results obtained by the LSI method, while some errors were observed. Thus, the present invention may provide novel and effective detection and tracking of the database. The method according to the present invention is also stable against the addition of new documents to the database, because the covariance matrix is used to reduce the dimension of the document matrix and only 15-20% of the largest i-th eigenvectors, which are not significantly sensitive to the addition of new documents to the database, are used. Therefore, once the covariance matrix is formed, many searches may be performed without elaborate and time consuming computation for singular value decomposition each time that search is performed, as long as the accuracy of the search is maintained, thereby significantly improving performance.

[0147] As described above, the present invention has been disclosed with respect to the specific embodiments thereof. However, a person skilled in the art may appreciate that various omissions, modifications, and other embodiments are possible within the scope of the present invention.

[0148] The present invention has been explained in detail with respect to the method for retrieving and ranking as well as detection and tracking. However, the present invention also contemplates a system for executing the method described herein, the method itself, and a program product within which the program for executing the method according to the present invention may be stored such as for example, optical, magnetic, or electro-magnetic media, used in the computer system of FIG. 10. The true scope can be determined only by the claims appended.