JP2000311173A - Device and method for retrieving similar document - Google Patents

Abstract

PROBLEM TO BE SOLVED: To retrieve a similar document which should be extracted with high precision without depending upon only the similarity by finding the probability that a document to be retrieved is a document which should be extracted by calculating the similarity between a retrieval key document and each document to be retrieved and outputting the document having the largest probability. SOLUTION: A similarity calculation part 206 calculates by a vector space method how much the retrieval key document in a retrieval key document storage buffer part 251 and a document to be retrieved in a retrieval object document storage buffer part 252 are similar to each other. A similarity probability conversion part 207 takes table data out of a similarity probability conversion table storage buffer part 256 according to the similarity and the field of the object document to be retrieved and calculates the corresponding similarity. A similar document candidate sorting part 208 sorts similar document candidates by the likelihood. A similar document candidate output part 209 calculates the probability that more than one correct answer is included in similar document candidates in order from the similar document candidate having the largest likelihood and displays the similar document candidates at a display part.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to an electronic document data search apparatus, and more particularly to a similar document search apparatus and a similar document search method for automatically searching for document data similar to a certain document data using the search key as a search key. About.

[0002]

2. Description of the Related Art In recent years, a large amount of digitized document data has been distributed, and for the purpose of performing automatic classification, a document resembling a document specified in a document database (hereinafter referred to as a search key document). A system for automatically retrieving documents has been put to practical use. As a general technique, a technique is known in which the similarity between a search key document and each search target document is calculated using the number of common words and the vector space method, and the similarity is output as similar documents in descending order of similarity. I have. Further, for example, as disclosed in Japanese Patent Application Laid-Open No. H11-73415, the statistical distribution of each similarity value (average of similarity values, etc.) is obtained in order to obtain a more reliable document as a similar document as a search result. There is a technique of outputting a document that satisfies a condition set by a user based on the statistical distribution as a similar document as a similar document. In short, this technology simply uses a set of similarity values and shows a method of extracting what should be regarded as similar documents in the distribution.

[0003]

The conventional similar document search apparatus uses a set of similarity values, and shows a method of extracting what should be regarded as similar documents in the distribution. We propose a method that is more accurate than using it. However, there is no consideration that the accuracy rate is not constant even for the same similarity value. That is, even if the similarity value is the same, the probability that the correct answer is included in the similarity value is not always the same. For example, in a search target document, there are two documents that happen to be given the same similarity value. However, one document may be correct and the other may be incorrect. In this way, if the similarity value is the same but there is a difference in the probability of being the correct answer, the above-mentioned technology treats the same similarity value as the same, so that a similar document with high accuracy that distinguishes the two. Cannot be extracted.

In addition, as for the number of documents to be output as similar, it is often sufficient that the output document contains one correct answer or a certain small number of specific answers. On the other hand, in the case of the above-mentioned technology, the distinction is made based on a condition such as “a search target document having a similarity of twice or more the average similarity is set as a search result”. When the determination is made only by the similarity in this manner, the probability that the correct answer is included in the output document varies widely, and there is a case where no correct answer is included in the output document or a case where the correct answer is output even more than necessary. Occurs. This is inefficient as a search operation.

SUMMARY OF THE INVENTION The present invention has been made in order to solve the above-described problem, and is directed to a similar document search apparatus and a similar document search method capable of searching for a similar document to be extracted with high accuracy without relying only on similarity. It is intended to be provided. The present invention also provides a similar document search device that uses a classification and a similarity to which a search target document belongs, reflects a difference in the accuracy rate of the similarity between the classes, and can search for a more appropriate similar document with high accuracy. It aims to provide a similar document search method. A further object of the present invention is to provide a similar document search apparatus and a similar document search method that allow a document to be extracted to be included in an output document as a result of a search with a desired probability.

[0006]

According to an embodiment of the present invention, there is provided a similar document search apparatus for searching a document similar to a search key document from a plurality of search target documents. A similarity calculating unit for calculating a similarity between the key document and each of the search target documents; and a probability that the search target document is a document to be extracted according to the similarity calculated by the similarity calculation unit. Means to seek,
Output means for outputting at least the document having the highest probability obtained by this means.
With this configuration, a document having a high probability of being extracted for each similarity is output, so that a similar document to be extracted can be searched with high accuracy.

Further, according to a second aspect of the present invention, there is provided a similar document search apparatus for searching a document similar to a search key document from a plurality of search target documents including classification information. A similarity calculating means for calculating a similarity between a search key document and each of the search target documents; and a search target document based on the similarity obtained by the similarity calculation means and the classification of the search target document. It is characterized by comprising means for calculating the probability that a document is a document to be extracted, and output means for outputting documents to be searched in order from the one with the highest probability obtained by this means. With this configuration, it is possible to search for a more appropriate similar document with high accuracy by reflecting the difference in the accuracy rate of the similarity between the classes.

According to another aspect of the present invention, there is provided a similar document search apparatus for searching a document similar to a search key document from a plurality of search target documents. And similarity calculating means for calculating the degree of similarity with each of the search target documents, and means for calculating the probability that the search target document is a document to be extracted according to the similarity calculated by the similarity calculation means And output means for outputting the search target documents in order from the one with the highest probability obtained by this means, and the search target documents output by the output means are output as the documents to be extracted in the search target documents. Means for until the probability of being included reaches a predetermined value or more. With such a configuration, a document to be extracted can be included in the output document with a desired probability.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a hardware configuration of a similar document search device according to an embodiment of the present invention.
This device is constructed as one function on a computer having a general architecture.

As shown in FIG. 1, the present apparatus comprises a control device 1, an input device 2 having a keyboard, a pointing device, a scanner, a display device 3 for displaying search results of similar documents, and an external storage device 4. Is done. The external storage device 4 stores a document information storage unit 4a for storing a large number of document information and a table storage unit 4b for storing a table for deriving accuracy from similarity. DVD (Di
digital video disc) device and the like.

FIG. 2 shows the configuration of the control device 1 in the present apparatus. The control device 1 has a CPU, a ROM, and a RAM. In FIG. 2, a portion performed by the CPU and the ROM is functionally represented as a program unit 200, and a RAM is functionally represented as a buffer unit 250. . The program unit 200 includes an initialization unit 201, a search key document input unit 20
4. Search target document reading unit 205, similarity calculation unit 20
6. Similarity probability conversion unit 207, similar document candidate sorting unit 2
08, similar document candidate output unit 209, output candidate number determination condition setting unit 210, correct / incorrect answer result input unit 211, similarity probability conversion table reading unit 212, and similarity probability conversion table writing unit 213. Have.

The buffer unit 250 includes a search key document storage buffer unit 251, a search target document storage buffer unit 252,
Similarity calculation result storage buffer unit 253, similar document candidate storage buffer unit 254, output candidate number determination condition storage buffer unit 255, similarity probability conversion table storage buffer unit 25
6 and an area 7 of the probability accumulation value storage buffer 257. The initialization unit 201 clears data in each buffer unit in the buffer unit 250 and writes a similarity accuracy conversion table from the table storage unit 4b to the similarity accuracy conversion table storage buffer unit 256.

The search key document input unit 204 causes the search key document storage buffer unit 251 to store the information of the search key document input by the user using the input device 2. The search target document reading unit 205 receives the search target document information from the document information storage unit 4a, and stores the search target document storage buffer unit 2.
52. The similarity calculation unit 206 calculates the similarity from the search key document stored in the search key document storage buffer unit 251 and the search target document stored in the search target document storage buffer unit 252 by the vector space method. Are stored in the similarity calculation result storage buffer unit 253. Here, the similarity may be calculated based on the number of common words instead of the vector space method.

The similarity probability conversion unit 207 extracts the data of the table stored in the similarity probability conversion table storage buffer unit 256 from the similarity calculated by the similarity calculation unit 206 and the field of the search target document, and After calculating the degree of accuracy, the similar document candidate is stored in the similar document candidate storage buffer unit 254. The accuracy is a value obtained by dividing the number of correct answers by a specific number. The positive number is the number of times that should be extracted as a search result, while the specific number is the number of times simply specified from the similarity and the classification. That is, the certainty is a probability that a target sentence is a correct answer to be hit, and is given for each similarity and each classification.

The similar document candidate sorting unit 208 sorts the similar document candidates stored in the similar document candidate storage buffer unit 254 with certainty, and stores them in the similar document candidate storage buffer unit 254 again. The similar document candidate output unit 209 calculates the probability that one or more correct documents are included in the similar document candidates stored in the similar document candidate storage buffer unit 254 in descending order of the likelihood. The calculated probability is stored in the output candidate number determination condition storage buffer 25.
The similar document candidates up to the output candidate number determination condition stored in No. 5 are output for display on the display device 3. The output candidate number determination condition is a probability that one or more correct answers are included in the output similar document candidates.

The output candidate number determination condition setting unit 210 stores the output candidate number determination condition input by the user via the input device 2 in the candidate number determination condition storage buffer unit 255. The correct answer / incorrect answer result input unit 211 is a correct answer to be extracted as a search result for each of the similarity calculation result stored in the similarity calculation result storage buffer unit 256 and the classification or field name given to the similar document. The data in the similarity / probability conversion table storage buffer unit 256 is updated / corrected in accordance with the information as to whether or not the similar document candidate should be extracted as a result. The information on whether the answer is correct is input by the user via the input device 2.

Similarity / accuracy conversion table reading section 212
Receives the similarity probability conversion table stored in the table storage unit 4b of the external storage device 4 and stores it in the similarity probability conversion table storage buffer unit 256. This similarity probability conversion table is a table for obtaining the probability from the similarity and the field of the search target document. The similarity probability conversion table writing unit 213 overwrites and stores the similarity probability conversion table stored in the similarity probability conversion table storage buffer unit 256 in the table storage unit 4b.

Next, the operation of the similar document search apparatus according to this embodiment will be described. The operation described here is based on the CP of the control device 1.
U executes by using a program in a ROM and a storage area in a RAM. This embodiment mainly includes a first step and a second step. The first step is to prepare a similarity probability conversion table using sample data and store it in the table storage unit 4b as a preparation stage for similar document search so that the similar document search device can perform similar document search. Step to let. The second step is a step of searching for similar documents using the created similarity probability conversion table. First, the first step of creating the similarity / degree of accuracy conversion table will be described. FIG. 3 is a flowchart showing the creation procedure.

First, the user uses the input device 2 to store data of a search target document to be searched in the document information storage unit 4a. At the time of this storage, a document ID for identifying the document and a field indicating the classification of the document are assigned to the search target document (step 300). Subsequently, the initialization unit 201 initializes all buffers (step 30).
1). When the user inputs a search key document to the search key document input unit 204 via the input device 2, the search key document is stored in the search key document storage buffer unit 251.
(Step 302). The search key document is, for example, a text document having contents as shown in FIG. 4, and is input by a keyboard or a scanner.

Next, the retrieval target document reading section 205 reads a plurality of retrieval target documents from the document information storage section 4a of the external storage device 4 to which the data of the retrieval target document has been inputted in step 300, and stores the retrieval target documents. Buffer unit 252
(Step 303). The search target document is, for example, as shown in FIG. 6, and as described above, the body, the document ID for identifying the document, and the field information indicating the classification of the document are added.

The similarity calculation unit 206 compares the search key document stored in the search key document storage buffer unit 251 with the text of the search target document stored in the search target document storage buffer unit 252, and uses the vector space method. To calculate the similarity. Here, a similarity calculation method using the vector space method will be briefly described. First, each of the search key document and the search target document is divided into words using morphological analysis. Next, the number of appearances of the same word appearing in the search key document is calculated, and similarly, the number of appearances of the same word appearing in the search target document is calculated. The words appearing in these documents and the number of appearances are extracted. For example, the word “document” appears three times in the search key document and the word “print” appears once, and the word “document” also appears in the search target document.
Suppose that the word "print" appears twice. In this case, the vectors of the search key document and the search target document are created using the number of “documents” on the x-axis and the number of “prints” on the y-axis. The vector of the search key document is (3,
1) The vector of the search target document is (5, 2). Then, COS θ is obtained by setting an angle formed by both vectors to θ.
The value obtained by normalizing this value with a predetermined formula is the similarity. Incidentally, the smaller the value of θ is, the higher the similarity is. This is performed for all words. The similarity calculation unit 206 stores the obtained similarity in the similarity calculation result storage buffer unit 253 together with the document ID of the search target document and the field information indicating the classification of the document. Similarity calculation result storage buffer unit 2
FIG. 6 shows an example of the data stored in the memory 53. The first data stored has the document ID “593”.
3 ", the field name is" publishing ", and the similarity is" 0.378 ".

After the similarity is stored, it is determined whether or not there remains a document to be searched for which the similarity has not been calculated (step 305). Steps 303 and 304
Is performed. On the other hand, if it is determined that there is no other document to be searched, the process proceeds to step 306. In step 306,
In step 304, the similarity calculation result storage buffer unit 253
The similarity calculation result stored in the similar document candidate sorting unit 2
At 08, the data is sorted by similarity and stored again in the similarity calculation result storage buffer unit 253.

Thereafter, the similar document candidate output unit 209 outputs
The display device 4 sorts similar document candidates in descending order of similarity.
(Step 307). The output count at this time is
There are a method of specifying each time by the user, a method of setting the number of cases in advance by the user, and outputting according to the number of cases. In this case, the display device 4 displays the similar document candidates for the top three candidates by pairing the document ID and the degree of similarity, for example, as shown in FIG.

Next, data indicating whether or not the similar document candidate output in step 307 is actually similar to the retrieval key document and should be extracted is determined by the input device 2 by the user.
By operation, it is input via the correct / incorrect answer result input unit 211 (step 308). The answer is correct if it should be extracted, otherwise it is incorrect. At this input,
The correct answer / incorrect answer result input unit 211 updates the corresponding part of the similarity accuracy conversion table in the similarity accuracy conversion table storage buffer unit 256 as follows (step 30).
9). FIG. 8 shows an example of the conversion table. Each row is divided by similarity, each column is divided by field,
Each field is divided into the number of correct answers and the specific number. In step 308, for example, when the user determines that the top-level search target document “document ID = 75268” is similar, and inputs data to that effect, the correct / incorrect answer result input unit 211 outputs the “ Domain = Printing "and" Similarity =
Then, the corresponding column A is specified from “0.39”, and “1” is added to both the number of correct answers “3” and the specific number “5” in that column. In addition, when the user determines that the second candidate search target document “document ID = 5933” is not similar, and inputs data to that effect, the correct / incorrect answer result input unit 211 outputs the “field = The corresponding column B is specified based on “publishing” and “similarity = 0.378”, and “1” is added only to the specific number “4” while the number of correct answers “2” in the column B remains unchanged. The same data processing is performed for the third-rank candidate. By this processing, the accuracy (= number of correct answers / specific number) of the corresponding column before and after the input is “field = printing” and “similarity = 0.38-0.
"39" has improved from 3/5 = 0.6 to 4/6 = 0.67, whereas "field = publishing" and "similarity = 0.37"
０．３0.38 ”drops from 2/4 = 0.5 to 2/5 = 0.4. In other words, this process is for setting an appropriate accuracy rate corresponding to the degree of similarity for each field, and when there is a difference in the accuracy rate for each field even with the same similarity,
This is to make it possible to reflect it in similar document extraction. When this is viewed in one field unit, the accuracy is determined corresponding to the similarity value.

It should be noted that, in addition to the method of the present embodiment, a method of determining whether or not the document is similar to the present embodiment is to input a document whose field is already specified but not registered as a search target document as a search key document. There is also a method of automatically determining whether or not the document is similar by comparing with a field assigned to a similar document output as a result of the search. Next, it is determined whether a search key document remains (step 31).
0) If it remains, return to step 302 and perform the same processing as described above for the search key document. If there is no other search key document, the process proceeds to step 311.

Finally, the similarity probability conversion table created in the preparation stage and stored in the similarity probability conversion table storage buffer unit 256 is written into the table storage unit 4b by the similarity probability conversion table writing unit 213 (step S1). 311) End the process. As shown in FIG. 9, the example of the created similarity / probability conversion table needs to be a table in which the number of correct answers and the specific number reach a predetermined practical level and can be used as the accuracy.

Next, the second step of searching for a similar document by using the similarity probability conversion table created as described above will be described. FIG. 10 is a flowchart showing the procedure. First, all buffers are cleared by the initialization unit 201, and the similarity / degree of accuracy conversion table reading unit 21 is read.
2, the similarity / accuracy conversion table storage buffer 25
6, the similarity / degree of accuracy conversion table is written from the table storage unit 4b of the external storage device 4 (step 401). The similarity / probability conversion table written here is the one created in the first step, or the one updated after the similar document search is executed. Here, it is assumed that the one created in the first step is written, and the similarity accuracy conversion table shown in FIG. 9 is written.

Subsequently, the output candidate number determination condition setting section 210
Is started. The output candidate number determination condition setting unit 210 receives from the user via the input device 2 a condition of a probability that one or more correct documents are included in the similar document candidates to be output,
It is stored (set) in the output candidate number determination condition storage buffer 255 (step 402). FIG. 11 shows that the probability that one or more correct answers are included in the output similar document candidate is 0.
It is a storage example when 97 is specified. Next, the search key document input unit 204 receives the search key document input by a user operation of the input device 2, and stores the search key document in the search key document storage buffer unit 251 (step 4).
03). As a specific example, it is assumed that a text document having the contents illustrated in FIG. 12 is stored as one of the search key documents.

Thereafter, the retrieval target document reading unit 205
Reads out a plurality of documents from the document information storage unit 4a and stores them in the search target document storage buffer unit 252 as search target documents (step 404). For example, a text document having contents as shown in FIG. Store as a document. The buffer part 2 stores the search key document and the search target document.
When stored in the search key document storage unit 50, the similarity calculation unit 206 compares the search key document stored in the search key document storage buffer unit 251 with the search target document stored in the search target document storage buffer unit 252. The similarity is calculated by the same method as in step 304, and is stored in the similarity calculation result storage buffer unit 253 as illustrated in FIG. 13 (step 40).
5).

After calculating the similarity with respect to the search target document, it is determined whether or not the search target document whose similarity has not been calculated remains (step 406).
Returning to step 04, the same processing as described above is performed on the search target document. If there is no other document to be searched, the process proceeds to the next process (step 407).

When the calculation of the similarity for all the search target documents is completed, the similarity calculation results stored in the similarity calculation result storage buffer unit 253 are calculated based on the similarity and the field assigned to the search target document. 14, the number of correct answers and the specific number in the corresponding column of the similarity probability conversion table stored in the similarity probability conversion table storage buffer unit 256 are extracted, and the accuracy (= correct number / specific number... Correct probability) is calculated. (Step 407). That is, step 405
"Document ID = 8652"
In the case of “field = computer” and “similarity = 0.371”, using the similarity accuracy conversion table in FIG. 9, the “computer” column of the row whose similarity is “0.37 to 0.38” From the specific number “327” of C and the number of correct answers “291”, the accuracy 291/327 = 0.890 is calculated. FIG. 14 shows an example of storing the converted similar document candidates. At the time of this similarity accuracy conversion, the similarity accuracy of the similarity accuracy conversion table (FIG. 9)
If the calculated similarity is more accurate than 0.01) in the example, the accuracy calculated from the values before and after the similarity accuracy conversion table is interpolated and used.

In step 407, the similar document candidates converted in the similarity probability and stored in the similar document candidate storage buffer unit 254 are sorted by certainty and stored again in the buffer unit 254 as shown in FIG. 15, for example (step S407). 40
8). Next, the process proceeds to the step of outputting, from the highest accuracy, candidates up to 0.97 which is the candidate number determination condition stored in the output candidate number determination condition storage buffer unit 255. First, the accumulated probability value storage buffer unit 257 is initialized to 1 (step 409). FIG. 16 shows an example of a state in which the accumulated accuracy buffer 257 is initialized. After initialization, the accumulated probability value storage buffer unit 257 is multiplied by a value obtained by subtracting the similarity document candidate probability from 1 to the accumulated probability value storage buffer unit 257.
7 (step 410) and output to the display device 3 via the similar document candidate 209 as a similar document candidate (step 411).

Subsequently, the input device 2 is fed back as to whether or not the similar document candidate output to the display device 3 is actually similar to the retrieval key document, and reflects it in the similarity probability conversion table. Acceptance from the operating user (step 412). The operation performed here according to the input data of similarity from the user is step 308,
And 309 are the same as the operations performed. In other words, the candidate number is not related to the specified number of the corresponding column (the field assigned to the similar document specified as the similarity) in the similarity probability conversion table stored in the similarity probability conversion table storage buffer unit 256 irrespective of whether or not it is similar. Is displayed, "1" is added, and when it is determined that they are similar, "1" is also added to the number of correct answers in the corresponding column (step 413).

The operations of steps 410 to 413 are repeated until the candidate number judgment condition stored in the output candidate number judgment condition storage buffer 255 is satisfied (step 414). Step 410 using FIGS. 15 and 17
A description will now be given of a specific example of the state of the similar document candidate and the change of the accumulated probability value storage buffer unit 257, and the operation of determining the number of candidates, in the steps from 414 to 414. First, since the certainty of the highest similar document candidate, that is, the document ID = 8652 has a certainty of 0.890, the cumulative certainty is 0.110 obtained by multiplying the initial value 1 by 1-0.890 = 0.110. Since 110 = 0.890 is smaller than the candidate number determination condition “0.97”, the process proceeds to the next candidate. The second similar document candidate, “Document ID = 24
Since the accuracy of “19” is 0.708, the accumulated accuracy is 0.11
0 is multiplied by 1-0.708 = 0.292 to obtain 0.03
2 and 1−0.032 = 0.968 is still smaller than the number-of-candidates determination condition “0.97”, so the process proceeds to the next candidate. The third similar document candidate, “Document ID = 39
924 ”is 0.508, so the cumulative accuracy value is 0.08.
Since 032 is multiplied by 1-0.508 = 0.492,
At 0.016, 1−0.016 = 0.0084 is larger than the candidate determination condition “0.97”, and it is determined that the condition is satisfied. The display on the display device 3 is based on these documents I
The process ends with only three candidates of D = 8652, 2419, 39924. FIG. 18 is an example of a similar document candidate display output when the similar document candidate is in the state shown in FIG. Here, the order in which these three candidates are displayed is not important. First, at least the point at which the most probable candidate is displayed and the probability that the correct answer is included in these three candidates are almost the desired value of 0.1. 0.984 near 97 ($ 1)
Is important.

When the update of the similarity / probability conversion table is completed, it is determined whether or not a search key document remains (step 4).
15) If it remains, return to step 403 and perform the processing of steps 403 to 414 for the search key document. If there is no other search key document, the process proceeds to step 416. Finally, the similarity / accuracy conversion table writing unit 21
3, the similarity / accuracy conversion table storage buffer unit 25
6 is written into the table storage unit 4b of the external storage device 4 (step 416), and the similar document search process ends.

It should be noted that although the processing after step 413 does not need to be performed for search, the processing after step 413 has an effect that the latest result can always be reflected in the similar document search. . In addition to performing the processing after step 412 immediately after outputting a similar document candidate, similar document output is collectively performed, and it is later determined whether or not those similar document output candidates are actually similar. The table may be updated.
The present invention is not limited to the above embodiments as long as it does not depart from the gist of the present invention. The present invention can be widely applied to a database search device, a document classification device, and the like.

[0037]

As described above in detail, according to the present invention, the correct answer rate based on the data accumulated for each calculated similarity is obtained, and a document having a high correct answer rate is extracted.
A similar document to be extracted can be searched with high accuracy without depending only on the degree of similarity. According to the present invention, since the data of the accuracy rate is accumulated for each field, it is possible to search for a more appropriate similar document with high accuracy by reflecting the difference in the accuracy rate of the similarity between the classifications. Further, according to the present invention, as a result of the search, a document to be extracted can be included in the output document with a desired probability.

[Brief description of the drawings]

FIG. 1 is a diagram showing a hardware configuration of a similar document search device according to an embodiment of the present invention.

FIG. 2 is a functional block diagram of a control device in the similar document search device of FIG. 1;

FIG. 3 is a diagram showing a procedure for creating a table for converting similarity of a search target document into accuracy;

FIG. 4 shows an example of a search key document when preparing a table.

FIG. 5 shows an example of a search target document when preparing a table.

FIG. 6 shows an example of a similarity calculation result of a search target document when preparing a table.

FIG. 7 shows an example of displaying similar document candidates when preparing a table.

FIG. 8 shows an example of a similarity accuracy conversion table in a table preparation stage.

FIG. 9 is an example of a similarity / probability conversion table used at the time of search;

FIG. 10 is a diagram showing a similar document search procedure.

FIG. 11 shows an example of an output condition determination condition at the time of a search.

FIG. 12 shows an example of a search key document at the time of search

FIG. 13 illustrates an example of a similarity calculation result of a search target document during a search.

FIG. 14 shows an example of the accuracy of similar document candidates

FIG. 15 shows examples of similar document candidates sorted by accuracy.

FIG. 16 shows an example of an initial value of a cumulative accuracy value.

FIG. 17 shows an example of a change in the cumulative accuracy value.

FIG. 18 shows an example of similar document candidate display at the time of search

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 control device 2 input device 3 display device 4 external storage device 4 a document information storage unit 4 b table storage unit 200 program unit 201 initialization unit 204 search key document input unit 205 read document to be searched Unit 206: similarity calculation unit 207: similarity probability conversion unit 208: similar document candidate sorting unit 209: similar document candidate output unit 210: output candidate number determination condition setting unit 211: correct / incorrect answer result input unit 212: similarity Probability conversion table reading unit 213 ... Similarity probability conversion table writing unit 250 ... Buffer unit 251 ... Search key document storage buffer unit 252 ... Search target document storage buffer unit 253 ... Similarity calculation result storage buffer unit 254 ... Similar document candidate storage buffer Unit 255: buffer unit for determining the number of output candidates judgment condition storage unit 256: conversion table for similarity and accuracy Paid buffer unit 257 ... likelihood accumulated value storage buffer unit

Continued on the front page (72) Inventor Tsuyoshi Matsukuma One, 3-3-1 Shinmachi, Ome-shi, Tokyo Inside Toshiba Computer Engineering Co., Ltd. (72) Inventor Yukio Nakamoto 3-3-1 Shinmachi, Ome-shi, Tokyo Toshiba Computer Engineering Inside (72) Inventor Takuya Nishina 3-3-1 Shinmachi, Ome-shi, Tokyo Toshiba Computer Engineering Co., Ltd. F-term (reference) 5B075 ND03 NK10 NK54 PP24 PQ02 PQ46 PR06 PR08 QM08 QS20 UU06

Claims (13)

[Claims] 1. A similarity calculating device for calculating a similarity between a search key document and each of the search target documents in a similar document search apparatus for searching a document similar to the search key document from a plurality of search target documents. Means for calculating the probability that the search target document is a document to be extracted according to the similarity calculated by the similarity calculating means, and outputting the document having the highest probability obtained by at least this means. A similar document search device comprising output means. 2. A similar document search device for searching a document similar to a search key document from a plurality of search target documents including classification information, wherein a similarity between the search key document and each of the search target documents is calculated. Means for calculating a probability that the search target document is a document to be extracted based on the similarity calculated by the similarity calculation means and the classification of the search target document; Output means for outputting the search target documents in order from the one with the highest probability obtained. 3. A similar document search apparatus for searching a document similar to a search key document from a plurality of search target documents including field information, calculating a similarity between the search key document and each of the search target documents. A similarity calculating unit; a probability converting unit that calculates a probability that the search target document is a document to be extracted based on the similarity obtained by the similarity calculation unit and the field of the search target document; Means for outputting documents to be searched in order from the one having the highest probability obtained by the means, and data indicating whether or not the document to be searched output by the output means is a document to be extracted as a result. A similar document search apparatus, comprising: input means; and means for correcting a method of obtaining a probability in the accuracy conversion means according to data input by the input means. 4. A similarity calculating device for calculating a similarity between a search key document and each of the search target documents in a similar document search apparatus for searching a document similar to the search key document from a plurality of search target documents. And first output means for outputting in descending order of similarity obtained by the similarity calculation means,
Input means for inputting data as to whether or not the search target document output from the output means is a document to be extracted as a result; and searching based on the data input by the input means in accordance with the degree of similarity. Creating means for creating a conversion table for determining the probability that the target document is a document to be extracted; searching means for searching for a similar document using the conversion table created by this means; And a second output unit for outputting the searched similar document. 5. A similarity calculating device for calculating a similarity between a search key document and each of the search target documents in a similar document search apparatus for searching a document similar to the search key document from a plurality of search target documents. Means for determining the probability that the search target document is a document to be extracted according to the similarity calculated by the similarity calculation means, and search target documents in descending order of the probability calculated by this means. Output means for outputting a search target document, and a determination means for determining that a probability that a document to be extracted is included in the output search target document has reached a predetermined value. A similar document search device, comprising: when the determining means determines that the predetermined value has been reached, terminating the output of the document from the output means. 6. A similar document search apparatus for searching a document similar to a search key document from a plurality of search target documents including classification information, wherein a similarity calculating a similarity between the search key document and each of the search target documents. The degree of similarity calculated by the degree of similarity calculation means and the classification of the search target document, and the number of times of extraction of the result of the search performed in the past and the extraction of the result of the search performed in the past are performed. A conversion unit that obtains a probability that the search target document is a document to be extracted from a ratio of the number of times determined to be correct; and an output unit that outputs at least the highest probability obtained by the conversion unit. A similar document search device characterized by the following. 7. A similar document search method for searching a document similar to a search key document from a plurality of search target documents,
A similarity between a search key document and each of the search target documents is calculated, and according to the calculated similarity, a probability that the search target document is a document to be extracted is obtained. A similar document search method characterized by output of a similar document. 8. A similar document search method for searching a document similar to a search key document from a plurality of search target documents including classification information, wherein a similarity between a search key document and each of the search target documents is calculated. Based on the similarity and the classification of the search target document, a probability that the search target document is a document to be extracted is obtained, and the search target documents are output in descending order of the obtained probability. Similar document search method. 9. A similar document search method for searching a document similar to a search key document from a plurality of search target documents including field information, calculating a similarity between the search key document and each of the search target documents. Based on the similarity and the field of the search target document, the probability that the search target document is a document to be extracted is obtained, and the search target document is output in descending order of the obtained probability. A similar document search method, comprising inputting data as to whether or not the retrieved document is a document to be extracted as a result, and correcting the probability calculation method according to the input data. 10. A similar document search method for searching a document similar to a search key document from among a plurality of search target documents, wherein a similarity between a search key document and each of the search target documents is determined.
A first step of searching for a similar document using a conversion table for converting the probability that the search target document is a document to be extracted, wherein the conversion table is created;
A second step of searching for a similar document using the conversion table. 11. A similar document search method for searching a document similar to a search key document from a plurality of search target documents, calculating a similarity between the search key document and each of the search target documents, and calculating the similarity. In accordance with the similarity obtained by the means, the probability that the document to be searched is a document to be extracted is obtained, and the documents to be searched are output in descending order of the probability obtained by this means. A similar document search method, wherein the output is terminated when the probability that a document to be extracted is included in the output search target document reaches a predetermined value. 12. A similar document search method for searching a document similar to a search key document from a plurality of search target documents including field information, calculating a similarity between a search key document and each of the search target documents. Corresponding to the calculated similarity and the field of the document to be searched, the ratio of the number of times that the result of the search performed in the past was determined to be correct to the number of times that the result of the search performed in the past was extracted is determined by A similar document search method characterized by determining the probability that a search target document is a document to be extracted, and outputting at least the determined highest probability document. 13. A similar document retrieval method for retrieving a document similar to a retrieval key document from a plurality of retrieval target documents including field information, wherein a similarity between a retrieval key document and each of said retrieval target documents is determined. A similar document is searched using a conversion table that is set for each field and similarity of the search target document and is converted into a probability that the search target document is a document to be extracted. And a second step of searching for a similar document using the conversion table. The first step determines the similarity between a search key document and each of the search target documents. Calculating, and inputting data indicating whether or not the search target document having the calculated similarity is to be extracted. According to the input data, the similarity and the field of the search target document are determined. Generating the divided conversion table, wherein the second step calculates a similarity between a search key document and each of the search target documents, and calculates the calculated similarity and the search target document. A step of obtaining a probability that the target document should be extracted using the conversion table in accordance with the classification of the document, and a step of outputting the document in the descending order of the probability of the search target document. Method.