JP2000035964A - Relevance calculating apparatus, storage medium storing relevance calculating program, and information retrieval system - Google Patents

Abstract

(57) [Summary] [Problem] To provide a relevance calculating apparatus capable of obtaining a relevance in consideration of a co-occurrence relationship while suppressing an increase in a processing load. In a relevance calculating apparatus that calculates the relevance of two pieces of document information by focusing on a plurality of attributes specified by one piece of document information, the relevance of two pieces of document information for each of a plurality of attributes is calculated. The partial relevance evaluation unit 111 that evaluates the partial relevance indicated, the integration unit 112 that integrates the evaluation values obtained for each attribute, and the plurality of attributes are commonly included in the other document information. Detecting means 113 for detecting a combination of attributes; detecting means 113
And a contribution adding unit 114 that adds a predetermined constant to the integration result by the integration unit 112 as a contribution of the degree of association due to the co-occurrence relation for each of the combinations of the plurality of attributes detected by.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a relevancy calculating apparatus for obtaining a relevancy between two specified document information in an information retrieval system for retrieving a desired document from a large-scale document database. Things. In recent years, the digitization of document information has been progressing, and a huge number of document information has been accumulated in a document database. In addition, a full-text search technology has been applied to an information retrieval system from a document database. It is becoming. For this reason, depending on the combination of keywords specified by the query command, a large number of documents may be output as search results by the information search system, and the degree of matching between the query command and each searched document may be determined. There is a need to rank search results based on the indicators shown. On the other hand, in document creation support systems, etc., as an index for automatically dividing large texts into paragraphs,
There is a need to determine the similarity between a part of a sentence of interest and another part. In order to obtain such an index, a technique for quickly and accurately evaluating the degree of relevance between two pieces of document information is required.

[0002]

2. Description of the Related Art In an information search system for ranking search results (hereinafter simply referred to as an information search system),
A set Q of at least one search character string included in the search formula specified by the user is set as one piece of document information, a document i in the document database is set as the other piece of document information, and the relevance Si of these pieces of document information is set Q Contribution W _ij for element j of
(See equation (1)).

[0003]

(Equation 1) Evaluation function f (i, j) for evaluating this contribution W _ij
Is the simplest example, as shown in equation (2).
Is assigned a numerical value “1” and a numerical value “0” according to whether or not has appeared in the document i.

(Equation 2) Various other evaluation functions f (i, j) have been proposed. Generally, a search character string j in a document i
In many cases, statistical information such as the appearance frequency (text frequency) tf _{ij of} the document and the number of appearing documents (document frequency) df _j including the search character string j in all the documents registered in the document database are used as parameters.

As an example of an evaluation function g (tf _ij , df _j ) using such statistical information as a parameter, as shown in equation (3), the reciprocal of the number of appearing documents df _j is the total number of documents N Multiplied by idf _j (inverse document frequenc
What is expressed using y) is simple.

(Equation 3) In actual calculations, the logarithm of each parameter is often used to match the information units.

For details of the evaluation function using statistical information as a parameter, see “Justin Zobel, Alistair Moff
at: Similarity Measures Explored, ACM SIGMOD Confere
nce, 1995 "and" Umino Satoshi: Principle of Word Weighting Based on Appearance Frequency Information, Libraryand information Science, N
o.26, 1988 ". FIG. 7 shows a configuration example of an information search system that performs ranking by applying the relevance calculation method. FIG. 8 is a flowchart showing the ranking processing operation.

In FIG. 7, a user enters a keyboard 401
Is input to the search control unit 412 via the reception processing unit 411, and is further input to the document database 4 by the database management system (DBMS) 402.
03 is provided for search processing.

The arithmetic processing unit 413 shown in FIG.
In response to an instruction from the search control unit 412, the partial relevance W _ij of the document i for the specified search character string j is calculated,
The integration processing unit 414 receives the obtained partial relevance W _ij from the arithmetic processing unit 413 and uses the buffer 415 to
It is configured to accumulate for each document ID corresponding to each document.

At this time, the database management system 40
2 searches the database 403 using the search character string j received from the search control unit 412 as a keyword, and sends a set of document IDs indicating documents including the search character string j to the arithmetic processing unit 413 via the search control unit 412. Just pass it. The arithmetic processing unit 413 receives the statistical parameters tf _ij and idf _j together with the document ID string including the search character string j as shown in step 501 of FIG. Using g (tf _ij , df _j ), a partial relevance W _ij is calculated for the specified document i (step 502).
What is necessary is just to send to the integration process part 414 shown in FIG.

In response, the integration processing unit 414 first determines whether or not the partial relevance of the document ID corresponding to the received partial relevance W _ij is already stored in the buffer 415 (step 503). . If affirmative determination in step 503, the relevant portions relevance W _ij newly computed portions relevance W _ij are held in the buffer 415
Is added (step 504), and in the case of a negative determination, the partial relevance W _ij corresponding to the corresponding document ID is newly stored in the buffer 415 (step 505).

Step 504 or step 50
After the end of Step 5, the search control unit 412 determines, in Step 506, all the documents i indicated by the document ID column.
It is determined whether the calculation of the partial relevance W _ij is completed. If the determination is negative, the process returns to step 502 to calculate the partial relevance W _ij for the next document i.

In this way, when the partial relevance is calculated for all the documents indicated by the document ID column, an affirmative determination is made in step 506, and in step 507,
The search control unit 412 determines whether or not the processing has been completed for all the search character strings specified by the search formula. In the case of a negative determination in step 507, the search control unit 412
May be performed by searching the database 403 using the new search character string as a keyword via the database management system 402, specifying the search character string in the arithmetic processing unit 413, and returning to step 501.

In this manner, steps 501 through 5 described above are repeated until the affirmative determination in step 507 is made.
07 is repeated to calculate the partial relevance W _ij of the document i for the search character string j, and the buffer 415
The relevance Si of each document can be obtained as the integration result held for each document ID. Therefore, the sort processing unit 416 shown in FIG. 7 sorts the document IDs based on the relevance Si stored in the buffer 415 according to the instruction from the search control unit 412 (step 50).
8) The ranking process may be terminated.

Thus, a number of documents extracted from the database 403 in accordance with the retrieval formula are ranked according to the relevance S, and the sorting result is displayed on the CRT display device (CRT) 405 via the display processing unit 404. By providing the processing, it can be provided to the user as an index when selecting a document considered important from a large number of documents obtained as a search result.

However, the relevance obtained by simply integrating statistical information such as the appearance frequency of each search character string as shown in Expression (3), and the user obtains the document obtained as a search result The degree of importance does not always match the degree of importance, which serves as an index when subjectively evaluating whether or not important. Therefore, based on statistical information, in order to determine the relevance of a value that approximates the importance of a document that is subjectively evaluated by a human, in addition to the appearance frequency of each search string, An evaluation formula has been proposed for obtaining a degree of relevance in consideration of a combination in which a search character string as an element appears together (hereinafter referred to as a co-occurrence relation) (Tetsu Takagi, Tsuyoshi Kitani) : Examination of document importance assignment using word co-occurrence relation, IPSJ research report, 96-FI-41, pp.61-68, 1996).

When the co-occurrence relation is considered, the relevance Si between the search formula Q and the document i is determined by the element j of the set J of the search character strings included in the document i as shown in formula (4). the sum of the sum of the partial relevance W _ij obtained by focusing, the sum of the above-described search string portions relevance W _iy obtained by focusing on the element y each set Y comprising a combination of elements of the set J of It is expressed as

[0016]

(Equation 4) Therefore, for example, a search expression Q specifying a search character string {a, b, c, d} is given, and the search character string a,
When b and c are included, if the evaluation function g (tf _ij , df _j ) shown in Expression (3) is used to calculate the partial relevance W _ij , the relevance between the document i and the search formula Q is obtained. Si is represented as in equation (5).

[0017]

(Equation 5) Here, since the presence or absence of the co-occurrence of the search character string in the document is considered to greatly affect the index when the user determines the importance of the document, as described above,
By evaluating the degree of relevance in consideration of co-occurrence relations, it can be expected that more effective ranking will be possible.

[0018]

As described above, when the relevance is obtained based on the statistical information about each search character string, the processing load required for the relevance calculation processing is small. Effective rankings may not be obtained. For example, when documents 1 to 4 shown in FIG. 9A are searched from the database in response to the input of a search expression specifying the search character string {hay fever, effect, medicine}, FIG. When calculated using the statistical parameters shown and the evaluation function shown in equation (3), the frequency of occurrence of the search character string {medicine} in document 4 is extremely high. The result that the degree of relevance between the search expression and the search expression is the highest is obtained.

A similar example is a document including a list of company names, such as a stock column. Therefore, when a full-text search is performed in a database that stores a large amount of newspaper articles and the like, a stock search is performed. A column or the like is searched as a document with a high rank, and it is not possible to provide a user with an effective index. On the other hand, if the relevance is evaluated in consideration of the co-occurrence relation, it is possible to evaluate a document including a plurality of search character strings higher than a document including only one search character string at a high frequency.

On the other hand, the processing load required for calculating the degree of relevance increases for the evaluation of the contribution of the co-occurrence relation, and the time required for the search processing becomes longer as a whole. The present invention relates to a relevance calculating device capable of obtaining a relevance in consideration of a co-occurrence relationship while suppressing an increase in processing load, a storage medium storing a relevance calculating program, and information provided with the relevance calculating device. The purpose is to provide a search system.

[0021]

FIG. 1 is a block diagram showing the principle of the relevance calculating apparatus according to the first and second aspects. According to a first aspect of the present invention, there is provided a relevance calculating apparatus for calculating relevance between two pieces of document information by focusing on a plurality of attributes specified by one piece of document information. A partial relevance evaluation unit 111 for evaluating the partial relevance indicating the relevance, an integration unit 112 for integrating the evaluation values obtained for each attribute, and a plurality of attributes commonly included in the other document information Detecting means 113 for detecting the combination of the attributes being detected, and detecting means 11
And a contribution adding means 114 for adding a predetermined constant to the integration result by the integration means 112 as a contribution of the degree of association by the co-occurrence relation for each combination of the plurality of attributes detected by step 3. .

According to the first aspect of the present invention, the contribution adding means 114 operates in accordance with the detection result of the detecting means 113, whereby the partial relations relating to the individual attributes obtained by the partial relevance evaluating means 111 and the integrating means 112 are obtained. By adding a predetermined constant to the sum of degrees as an evaluation value of each of the co-occurrence relations that have appeared, it is possible to obtain the degree of relevance considering the co-occurrence relations with a small amount of calculation.

According to a second aspect of the present invention, there is provided a relevance calculating apparatus for calculating relevance between two pieces of document information by focusing on a plurality of attributes specified by one piece of document information. Partial relevance evaluation means 111 for evaluating the partial relevance indicating the relevance of the document information, and integration means 112 for integrating the evaluation values obtained for each attribute
Detecting means 113 for detecting, from a plurality of attributes, a combination of attributes commonly included in the other document information.
And for each combination of the plurality of attributes detected by the detection unit 113, an approximate value is evaluated using the function of the partial relevance obtained by the partial relevance evaluation unit 111 for each attribute in the co-occurrence relationship. And a co-occurrence relation evaluation unit 115 for adding the approximate value to the result of integration by the unit 112.

According to a second aspect of the present invention, the co-occurrence relation evaluation means 115 operates according to the detection result of the detection means 113, so that the contribution due to the co-occurrence relation uses a function of the degree of partial relevance for each attribute. Is added to the sum of the partial relevance for each attribute obtained by the partial relevance evaluating means 111 and the integrating means 112. In this case, it is possible to evaluate the contribution of the co-occurrence relation in consideration of the statistical characteristics of each attribute in the co-occurrence relation, so that statistical information such as the frequency of occurrence of the co-occurrence relation is reflected in the degree of association. do it,
The degree of association can be evaluated more precisely.

According to a third aspect of the present invention, there is provided a storage medium storing a relevance calculating program for calculating a relevance between two pieces of document information by focusing on a plurality of attributes specified by one piece of document information. A partial relevance evaluation procedure for evaluating a partial relevance indicating the relevance of two pieces of document information for each, an integration procedure for integrating the evaluation values obtained for each attribute, and the other document from among a plurality of attributes A detection procedure for detecting a combination of attributes commonly included in the information and, for each combination of a plurality of attributes detected in the detection procedure, a partial relevance evaluation procedure for each attribute in a co-occurrence relationship The computer evaluates the approximate value using the function of the partial relevance, and causes the computer to execute a co-occurrence relation evaluation procedure of adding the approximate value to the integration result by the integration procedure. Characterized in that it records the program.

According to a third aspect of the present invention, a co-occurrence relation evaluation procedure is executed according to a detection result in the detection procedure.
The contribution by the co-occurrence relationship is evaluated using a function of the partial relevance for each attribute, and is added to the sum of the partial relevance for each attribute obtained by the partial relevance evaluation procedure and the integration procedure. In this case, it is possible to evaluate the contribution of the co-occurrence relation in consideration of the statistical characteristics of each attribute in the co-occurrence relation, so that statistical information such as the frequency of occurrence of the co-occurrence relation is reflected in the degree of association. Thus, the degree of association can be evaluated more precisely.

FIG. 2 is a block diagram showing the principle of an information retrieval system according to the present invention. According to a fourth aspect of the present invention, there is provided a document database system comprising:
In an information search system including a ranking processing unit 103 for assigning a rank according to the degree of association with a search formula to a set of documents searched by the search unit 102 from 01, the ranking processing unit 103 converts a plurality of search character strings into Receiving means for receiving and retrieving search character strings from the sorting means 121 in order, the sorting means 121 sorting and outputting the documents in the order in which the number of appearing documents indicating the frequency included in the documents stored in the document database 101 is small, A partial relevance calculating unit 122 that calculates a partial relevance indicating the relevance focusing on the corresponding search character string for each document and the search formula searched by the search unit 102 as including the search character string; Tallying means 1 for integrating the partial relevance obtained by the relevance calculating means 122 for each document
23 and each time the partial relevance calculating means 122 calculates the partial relevance of a new search character string, for each applicable document, either the search character string or the search character string for which the contribution has already been tabulated. A co-occurrence relation detecting means 124 for detecting a co-occurrence relation with the person;
24, each time a co-occurrence relationship is detected, the integration result obtained by the tallying unit 123 for the corresponding document is multiplied by a predetermined constant to obtain a tally result including the contribution by the detected co-occurrence relationship. And a ranking unit that ranks the documents whose relevance to the search formula is calculated based on the counting result of the counting unit and provides the search result to the output processing of the search result. And

According to the fourth aspect of the present invention, the partial relevance calculating means 122 is used in accordance with the order determined by the sorting means 121.
In addition, each time the co-occurrence relation is detected by the co-occurrence relation detecting means 124 in this process, the contribution adding means 125 is operated, so that a case where a predetermined model is established is established. According to the approximation formula, the degree of relevance including the contribution due to the co-occurrence relation can be calculated by a simple process and provided to the process by the ranking means 126.

[0029]

Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 3 shows a configuration of an information retrieval system to which the relevance calculating device according to claim 1 is applied. This information retrieval system adds an addition processing unit 211 to the conventional information retrieval system shown in FIG. 7, and the addition processing unit 211 responds to an instruction from the integration processing unit 414 to store a corresponding document in a buffer 415. A predetermined constant C1 is added to the degree of partial relevance as a contribution of the co-occurrence relation.

FIG. 4 is a flowchart showing a ranking processing operation in the information retrieval system. Similar to the processing of steps 501 and 502 shown in FIG. 8, the arithmetic processing unit 413 receives the document ID via the search control unit 412, and operates as the partial relevance evaluation unit 111 described in claim 1. For the document i indicated by the document ID, the partial relevance W _ij for the search character string j may be calculated (steps 301 and 302).

Upon receiving the partial relevance W _ij , the integration processing unit 414 determines whether the partial relevance W _ij for the document i is already stored in the buffer 415 as in step 503. Steps 301 to 303). Here, the fact that the partial relevance W _ij of the document i is already stored in the buffer 415 means that:
Indicates that there is a co-occurrence relation between the previously partially relevance W _ij search string j became calculation target of the calculation target since the search string and the current partial relevance W _ij of.

Therefore, in the case of an affirmative determination in step 303, the integration processing unit 414 instructs the addition processing unit 211 to add the contribution of the co-occurrence relationship, and accordingly, the addition processing unit 211 Then, the constant C1 is added to the partial relevance of the document i stored in (step 304), and the newly calculated partial relevance may be added similarly to step 504 (step 305).

On the other hand, in the case of a negative determination in step 303, the new partial relevance may be held as in step 505 (step 306). As described above, the integration processing unit 414 operates according to the determination result of step 303, and integrates the partial relevance for each document including the search character string j. Are implemented in parallel for each document.

Steps 306 and 3 described above
After the processing of step S05 is completed, steps 506 to 506 shown in FIG.
In the same manner as in step 507, it is determined whether the processing has been completed for all documents including the search character string j and all search character strings (step 307 and step 30).
8) In response to the affirmative determination in step 308, step 5
The same sort processing as in step 08 is performed (step 309).
What is necessary is just to end a process.

As described above, the functions of the detecting means 113 and the contribution adding means 114 described in claim 1 are realized by operating the integrating processing section 414 and the adding processing section 211 in accordance with the result of the determination in step 303. Then, in accordance with the detection of the co-occurrence relationship, a predetermined contribution can be added to the partial relevance of the corresponding document i. For example, when the above-described ranking process is applied to the documents 1 to 4 shown in FIG. 9 with the constant C1 being a numerical value “3”, the relevance and the contribution by the co-occurrence relationship obtained for each document are shown in Table 1. Show.

[0036]

[Table 1] In this case, for the document 4, the contribution due to the co-occurrence relation is not added at all, whereas for the document 2, the contribution due to the co-occurrence relation between the search character string {drug} and the search character string {effect} is not added. For documents 1 and 3, contributions due to co-occurrence relations for three search character strings including the search character string {hay fever} are added.

For this reason, document 1 is evaluated the highest, and then a large value is obtained as the degree of relevance of documents 3 and 2.
The relevance of the document 4 is evaluated to be the lowest. As described above, if an appropriate value is set as the contribution of the co-occurrence relation, the relevance between the search character string and the document is calculated in consideration of the co-occurrence relation by the simple processing as described above, and the search is performed. It can be provided to the user as a more effective index for selecting documents.

The value of the above-mentioned predetermined constant C1 is calculated, for example, by calculating the degree of relevance for various combinations of documents and search character strings, and ranking is performed. The result is closer to human subjective judgment. What is necessary is just to determine. The processing in step 302 corresponds to the partial relevance evaluation procedure described in claim 3.
Corresponds to a detection procedure, step 304 corresponds to a co-occurrence relation evaluation procedure, and steps 305 and 306 correspond to an integration procedure.

Therefore, if a storage medium storing a program for causing a computer to execute the processes of steps 301 to 309 described above is distributed, a wide range of users can be provided with a ranking result based on the degree of association in consideration of co-occurrence relationships. In addition, it is possible to support the operation of selecting a document having high importance from the retrieved documents. By the way, as described above, when the contribution by the co-occurrence relation is set to a predetermined constant, all the co-occurrence relations are evaluated as being equivalent, so that the statistical characteristics of the search string having the co-occurrence relation are evaluated. Was not reflected in the relevance at all.

Next, a description will be given of a method of calculating the degree of relevance in consideration of the statistical characteristics of the co-occurring search character strings. Here, when the search character strings a and b are included in the document i, a search expression that specifies these search character strings and the document i
Is related to the number of documents df _ab in which the search character string a and the search character string b appear together (hereinafter referred to as the number of co-occurring documents) and the occurrence of the search character strings a and b in the document i. Using the calculated frequency (hereinafter referred to as co-occurrence frequency) tf _iab ,
It is represented as

[0041]

(Equation 6) If the number of appearing documents dfa of the search character string a is smaller than the number of appearing documents dfb of the search character string b, the number of co-occurring documents df _ab for the search character strings a and b is, of course, Less than or equal to the number dfa. Above all, if the set J _a document containing the search string a is completely contained in the set J _b of documents that contain the search string b, the search string a, co-occurrence document number df _ab for b is , The number of appearing documents dfa in the search character string a.

Therefore, the above condition (J _a aJ _b )
If the holds, the co-occurrence frequency tf be substituted _iab in occurrence frequency tf _a of the search string a, partial relevance W _ia the search string a the contribution of the co-occurrence relationships included in the formula (6) And as a function of the partial relevance W _ij for each search string, as shown in equation (7),
The degree of association Si 'including the co-occurrence contribution can be represented.

[0043]

(Equation 7) Of course, when the above-mentioned condition is satisfied, it is a very limited special case, so that equation (7) cannot be generalized as it is. However, for example, as shown in Expression (8), the term including the contribution of the co-occurrence relation is expressed by multiplying the partial relevance of the search character string a by the constant C2, and the value of the constant C2 is appropriately adjusted. If it is determined, it can be expected that an approximate value of the degree of association Si including the contribution by the co-occurrence relation in the general case will be obtained.

[0044]

(Equation 8) Similarly, in the case where the search character strings a, b, and c are included in the document i, a method for obtaining an approximate value of the relevance Si of the search expression specifying the search character strings and the document i will be described. In this case, as a model for approximation, it is completely included in the set J _b document set J _a document containing the search string a contains the search string b, further comprising a search string b A document set _Jb whose document set Jb contains the search character string c
Consider the case where it is completely included in _c .

In the above-described model, the number of co-occurring documents df _ab , df _ac , for the subset of the search character string set {a, b, c} that includes the search character string a as a part thereof
df _abc is equal to the occurrence document number df _a search string a, The search string b, co-occurrence document number df _bc for c is equal to the appearance document number df _b search string b. Using this, by transforming equation (5), the above condition (J _a aJ _b ⊆
_Jc ) holds, the relevance Si 'of document i is
As shown in Expression (9), the search character string is expressed as a function of the partial relevance W _ij .

[0046]

(Equation 9) Therefore, the approximate value Si of the degree of association for the general case
Is expressed as shown in Expression (10) using the above-described constant C2 and the partial relevance W _ij of each search character string.

(Equation 10) Similarly, the number of appeared documents of the n-number of the search string term ₁ ~term _n
if _{df term1} ~df termn meets the relationship shown in equation (11), the approximate value Si of relevance in consideration of cooccurrence relation document i containing these search strings term ₁ ~term _n is formula
It can be represented by (12).

[0047]

[Equation 11]

(Equation 12) Here, in the model assumed to derive the equation (12), a set J of documents in which the search character strings term _{1 to} term _n appear
_{Considering that terms1 to Jtermn} satisfy the nested inclusion relationship shown in equation (13), equation (12) can be modified as shown in equation (14).

[0048]

(Equation 13)

[Equation 14] By doing so, in order to calculate the partial relevance in order from the search character string having the smaller number of appearing documents in accordance with the magnitude relation shown in Expression (11), in the process of calculating the relevance similar to the procedure shown in FIG. Each time a co-occurrence relationship is found in the same manner as in step 303, the calculated result of the partial relevance is multiplied by a constant C2 to obtain a search character string whose contribution is already calculated. It can be seen that a partial sum including the contribution due to the co-occurrence relationship with the character string can be obtained.

As described above, the contribution of the co-occurrence relationship between the processed search character string and the search character string currently focused on as a function of the partial relevance of the search character string which has been calculated so far. By expressing, it is possible to obtain the relevance in consideration of the statistical characteristics of the co-occurring search character string with a calculation amount substantially equal to that of the conventional relevance calculation processing.
FIG. 5 shows an embodiment of the information retrieval system according to claim 4.

In this case, as shown in FIG. 5, a relevance calculating device is provided with a multiplication processing unit 212 instead of the addition processing unit 211 shown in FIG.
Is in response to an instruction from the integration processing unit 414,
15 is multiplied by the constant C2 to the corresponding partial relevance,
The result of the multiplication may be held in the buffer 415.

In FIG. 5, the order determining unit 213
Is configured to receive a set Q of search character strings via the reception processing unit 411 and determine the calculation order of the degree of partial relevance based on the number of appearing documents corresponding to these search character strings. For example, in response to input of a set of search character strings, the order determination unit 213 requests the number of corresponding appearing documents from the database management system 402 corresponding to the search unit 102, and receives the smaller number of appearing documents in ascending order. The order of the search character strings may be rearranged, and this series of search character strings may be provided to the processing of the search control unit 412.

FIG. 6 is a flowchart showing the ranking processing operation. In this case, prior to the relevance calculation process, first, the above-described order determination unit 213 operates as the sorting unit 121 described in claim 4, and rearranges the search character strings into the order in which the partial relevance is to be calculated. (Step 311) In accordance with this order, the search control unit 412 specifies a search character string in the database management system 402 and searches for a corresponding document. What is necessary is just to send to the arithmetic processing part 413 as a calculation object of a degree.

In response, the processing similar to steps 301 to 303 shown in FIG. 4 is performed by the arithmetic processing unit 413 and the integration processing unit 414 (step 3).
12 to step 314). At this time, the arithmetic processing unit 413
Accordingly, the function of the partial relevance calculating means 122 described in claim 4 is fulfilled, and the function of the co-occurrence relation detecting means 124 is fulfilled by the integration processing section 414 performing the determination processing of step 314. Have been.

In response to the affirmative determination in step 314,
5. The contribution adding means 1 according to claim 4, wherein
25, and in step 315, the partial relevance W of the corresponding document i held in the buffer 415.
_ij is multiplied by the constant C2, and in step 316, the integration processing unit 414 adds the newly calculated partial relevance to obtain the processed search character string and the search character string j currently focused on. , A partial sum of the partial relevance including the contribution of the co-occurrence relation is obtained, and the partial relevance W corresponding to the document i is obtained.
_ij is held in the buffer 415.

On the other hand, in the case of a negative determination in step 314, the new partial relevance may be held as in step 306 (step 317). As described above, the accumulation processing unit 414 operates according to the determination result of step 314, and operates the value of the partial relevance of the corresponding document i in the buffer 415, thereby obtaining the totaling unit 123 described in claim 4. The function of has been fulfilled.

In this case, the process of calculating the partial relevance is performed in the order determined by the order determining unit 213, and in the process of calculating the partial relevance, the above-described step 314 is performed.
The function of the co-occurrence relation evaluation means 115 described in claim 2 is fulfilled by operating the multiplication processing section 212 in accordance with the determination result of Is evaluated as a function of the degree of partial relevance of each search character string in co-occurrence.

After the processing of step 317 and step 316 described above is completed, in the same manner as in steps 307 to 308 shown in FIG. 4, all the documents including the search character string j and all the search character strings are searched. It is determined whether the processing is completed (steps 318 to 3
19), in response to the affirmative determination in step 319, the sort processing unit 416 operates as the ranking means 126, and performs the same sort processing as in step 309 (step 32).
0), the process may be terminated.

As described above, by adding the processing of sorting the search character string and the processing of multiplying the constant by the detection of the co-occurrence relation to the conventional relevance calculation processing, the statistical information on the search character string is obtained. , The degree of relevance including the contribution of the co-occurrence relation can be calculated, and a rank corresponding to the degree of relevance can be assigned, thereby realizing the function of the ranking processing means 103 described in claim 4. Can be.

For this reason, the amount of calculation that is increased by the processing added is extremely small. On the other hand, the contribution by the co-occurrence relation in consideration of the statistical characteristics of the search character string to be co-occurred is evaluated. Thereby, the degree of relevance obtained for each document can be approximated to the degree of importance that humans subjectively evaluate. For example, for documents 1 to 4 shown in FIG. 9, Expression 3 is used as an expression for evaluating the degree of partial relevance, and a numerical value “2” is used as the value of a constant C2 used for calculating the contribution of the co-occurrence relationship.
Table 2 shows the result of calculating the degree of association by giving

[0060]

[Table 2] In this case, the contribution of the co-occurrence relation is evaluated using the frequency of appearance of the related search character string and the reciprocal of the number of appearance documents as weights. Therefore, rather than the co-occurrence relationship of search strings that appear in a relatively large number of documents, such as search strings {drugs} and search strings {effects}, they are more limited to search strings {hay fever} Emphasis is placed on the co-occurrence relationship of a search character string that appears only in a document, and a high evaluation value can be given according to the appearance frequency of the co-occurrence relationship that is regarded as important.

In the above-described example, the numerical value "2" is used as the value of the constant C2 used for the contribution of the co-occurrence relationship. However, a sufficient number of documents are obtained by the relevance calculating device shown in FIG. If an appropriate value of the constant C2 is obtained based on an experiment for collating the obtained ranking result with the ranking result by a human, it is expected that a ranking result closer to the evaluation result of the importance by a human will be obtained.

Further, if various evaluation formulas proposed in reference documents and the like are used as the evaluation formulas used in the arithmetic processing unit 413, the features of the evaluation formulas can be inherited.

[0063]

As described above, according to the first aspect of the present invention, the contribution of the co-occurrence relationship can be added to the evaluation result of the relevance by a very simple process. Can be obtained without significantly increasing the amount of calculation of.

Thus, for example, if the search documents are ranked based on the relevance, a more effective index can be given when the user selects the search documents according to the importance. According to the second and third aspects of the present invention, the weights corresponding to the statistical characteristics of each attribute are approximated by approximating the contribution by the co-occurrence relation by the function of the partial relevance obtained for the already processed attribute. Can be given to the contribution of the co-occurrence relation, so that the contribution by the co-occurrence relation can be evaluated more precisely.

In particular, the invention according to claim 4 calculates the partial relevance in an order corresponding to the number of documents appearing for each search character string, thereby obtaining a partial sentence sum of the partial relevance for the processed search character string. Is multiplied by a constant, it is possible to obtain an approximate value of the degree of relevance including the contribution due to the co-occurrence relationship precisely evaluated using the weight corresponding to the statistical feature of the search character string.

[Brief description of the drawings]

FIG. 1 is a principle block diagram of a relevance calculating apparatus according to claims 1 and 2;

FIG. 2 is a block diagram showing the principle of an information retrieval system according to claim 4;

FIG. 3 is a diagram showing a configuration of an information search system to which the relevance calculating device of claim 1 is applied.

FIG. 4 is a flowchart illustrating a ranking processing operation.

FIG. 5 is a diagram showing an embodiment of an information search system according to claim 4;

FIG. 6 is a flowchart illustrating a ranking processing operation.

FIG. 7 is a diagram showing a configuration example of an information search system to which a conventional relevance calculating device is applied.

FIG. 8 is a flowchart showing a conventional ranking processing operation.

FIG. 9 is a diagram illustrating an example of document information and a ranking result.

[Explanation of symbols]

 101, 403 Document database 102 Search means 103 Ranking processing means 111 Partial relevance evaluation means 112 Accumulation means 113 Detection means 114 Contribution addition means 115 Co-occurrence relation evaluation means 121 Alignment means 122 Partial relevance calculation means 123 Aggregation means 124 Co-occurrence Relationship detection means 125 Contribution addition means 126 Ranking means 211 Addition processing unit 212 Multiplication processing unit 213 Order determination unit 401 Keyboard 402 Database management system (DBMS) 404 Display processing unit 405 CRT display (CRT) 411 Reception processing unit 412 Search control Unit 413 arithmetic processing unit 414 integration processing unit 415 buffer 416 sort processing unit

Claims (4)

[Claims] 1. A relevance calculating apparatus for calculating relevance between two pieces of document information by focusing on a plurality of attributes specified by one piece of document information, wherein the relevance of two pieces of document information for each of the plurality of attributes is calculated. A partial relevance evaluating means for evaluating the partial relevance indicating the degree; an integrating means for integrating the evaluation values obtained for each attribute; and the common attribute included in the other document information from among the plurality of attributes. Detecting means for detecting a combination of attributes which are present, and for each combination of the plurality of attributes detected by the detecting means, a contribution of adding a predetermined constant to the integration result by the integration means as a contribution of the degree of association by the co-occurrence relationship. A relevance calculating apparatus comprising: a minute adding unit. 2. A relevance calculating apparatus for calculating relevance between two pieces of document information by focusing on a plurality of attributes specified by one piece of document information, wherein the relevance of the two pieces of document information for each of the plurality of attributes is calculated. A partial relevance evaluating means for evaluating the partial relevance indicating the degree; an integrating means for integrating the evaluation values obtained for each attribute; and the common attribute included in the other document information from among the plurality of attributes. Detecting means for detecting a combination of attributes that are present; and for each of the plurality of attribute combinations detected by the detecting means, a function of the partial relevance obtained by the partial relevance evaluating means for each attribute in co-occurrence. And a co-occurrence relation evaluating means for evaluating the approximate value using the sum and adding the approximate value to the result of integration by the integrating means. 3. A storage medium storing a relevance calculation program for calculating relevance between two pieces of document information by focusing on a plurality of attributes specified by one piece of document information, wherein 2 A partial relevance evaluation procedure for evaluating the partial relevance indicating the relevance of two pieces of document information; an integration procedure for integrating the evaluation values obtained for each attribute; and a common procedure for the other document information from the plurality of attributes. A detection procedure for detecting a combination of attributes included as a part, and for each of a combination of a plurality of attributes detected in the detection procedure, a part obtained in the partial relevance evaluation procedure for each attribute in a co-occurrence relationship. The computer executes a co-occurrence relation evaluation procedure of evaluating an approximate value using the function of the degree of association and adding the approximate value to the integrated result by the integrating procedure. A storage medium recording a program to. 4. A ranking processing means for, in response to input of a search expression including a plurality of search character strings, assigning a rank according to the degree of relevance to the search expression to a set of documents searched by a search means from a document database. The ranking processing means receives the plurality of search character strings and sorts them in ascending order of the number of appearing documents indicating the frequency included in the documents stored in the document database. A search unit that sequentially receives a search character string from the alignment unit, and focuses on a corresponding search character string for each document and the search expression searched by the search unit as including the search character string. A partial relevance calculating means for calculating a partial relevance indicating the relevance, and a partial relevance obtained by the partial relevance calculating means Each time the partial relevance of a new search character string is calculated by the totaling means for integrating for each document and the partial relevance calculating means, the search character string and the contribution have already been totaled for each applicable document. A co-occurrence relation detecting means for detecting a co-occurrence relation with any one of the searched character strings, and each time the co-occurrence relation is detected by the co-occurrence relation detecting means, Contribution adding means for multiplying the result by a predetermined constant to obtain a totaled result including the detected co-occurrence contribution, and calculating the degree of association with the search formula based on the totalized result by the totalizing means An information search system, comprising: a ranking unit that ranks the selected documents and provides the result to a search result output process.