JP6181890B2 - Literature analysis apparatus, literature analysis method and program - Google Patents

Description

  The present invention relates to a document analysis device, a document analysis method, and a program.

  For example, a process for assigning a classification code indicating a technical classification to a patent document or searching for another document having similar contents to the patent document is generally performed manually. For this reason, it is inevitable that a difference occurs in the accuracy of the grant and the accuracy of the search depending on the subjectivity and experience of the person in charge.

  Therefore, research is being conducted on techniques for automatically assigning classification codes and searching for similar documents using a computer. In this method, for example, an estimation dictionary described in advance using one Boolean expression based on a logical sum and logical product based on terms as the assignment condition or search condition, or one neighborhood expression consisting of a plurality of terms is created. By comparing the character string described in the document to be classified or searched and the above dictionary for estimation, the degree of similarity between the two is obtained, and whether or not the classification can be given according to the obtained similarity or whether the document is a similar document. A method of determining whether or not is being studied (see, for example, Patent Document 1 or Patent Document 2).

JP 2004-94583 A However, it is difficult to define complicated assignment conditions or search conditions without excess or deficiency in the estimation dictionary in which the above-described assignment conditions or search conditions are described using one Boolean algebraic expression or neighborhood expression. In addition, a large amount of processing time is required for comparison with the target document.

  The present invention has been made paying attention to the above circumstances, and its object is to make it possible to simply and accurately define the grant conditions or search conditions described in the estimation dictionary, thereby reducing the accuracy with less processing time. An object of the present invention is to provide a document analysis apparatus, a document analysis method, and a program that enable high classification and document search.

  In order to achieve the above object, according to a first aspect of the present invention, there is provided a document meaning in which a document analysis apparatus includes a basic element in each range obtained by dividing a search condition by a predetermined range in which a document is separated as a semantic expression. An estimation dictionary storage unit that stores estimation dictionary data described by an expression, and a predetermined term that breaks down document data related to a document to be searched into word units and generates a term as a semantic expression. A document term storage unit for storing each term divided as a unit as a document term data, a sentence meaning included in the estimation dictionary data stored in the estimation dictionary storage unit, and a document term storage unit A score calculation that compares terms contained in stored document terminology data for each range divided by a predetermined range that is separated as a semantic expression, and calculates a score that represents the degree of similarity And stage, on the basis of the calculated score, in which the search target documents is to be provided with a determining means for determining whether or not the search condition is satisfied literature.

  According to a second aspect of the present invention, the predetermined range defined as a semantic expression defined in each of the estimation dictionary data, the document term data, and the score calculation means is any one of a phrase, a sentence, and a paragraph. This is the range.

  According to a third aspect of the present invention, the estimation dictionary data includes a compound basic element obtained by hierarchically combining basic elements defined in each of the phrase, sentence, and paragraph in a sentence expression. The basic elements include a term and one or more phrase basic elements for a sentence basic element, and a term and one or more sentence basic elements or phrase elements for a paragraph basic element. Including a basic element, and the score calculation means compares a sentence meaning included in the dictionary data for estimation stored in the dictionary storage unit with a term included in the document term data stored in the document term storage unit For each phrase, sentence, and paragraph of the document to be searched, the term group included in the range is compared with the sentence meaning including the complex basic element, and the score is calculated based on the matching frequency. What to calculate A.

  According to a fourth aspect of the present invention, the estimation dictionary data defines a controlled word that is a generic term for a plurality of terms having the same meaning, describes the grammatical expression using the controlled word, and the score When the calculation means compares the meaning expression included in the estimation dictionary data stored in the estimation dictionary storage unit and the term included in the document term data stored in the document term storage unit, the calculation means Compare multiple terms defined as synonyms with controlled terms and literature term data, and if any one of the terms defined as synonyms is included in the literature term data, the condition for comparison match It is determined that it falls under.

  According to a fifth aspect of the present invention, the score calculation means divides the search target range of the search target document into a plurality of ranges in units of the first range having a group as a semantic expression, For each of the ranges, a first calculation means for calculating a first score based on a matching frequency between the term group included in the range and the literary expression, and among the search target ranges of the search target documents, For a second range wider than the first range, a second calculating means for calculating a second score based on the coincidence frequency between the term group included in the range and the sentence meaning; and the calculated first And a third calculation means for calculating a score representing a similarity between the estimation dictionary data and the document to be searched based on the score of the second and the calculated second score It is.

  According to a sixth aspect of the present invention, the third calculation means weights and adds the maximum value of the calculated first scores and the second score with a preset coefficient. The score representing the similarity between the estimation dictionary data and the document to be searched is calculated.

  In a seventh aspect of the present invention, when the dictionary data for estimation is composed of a classification code assignment condition described by a sentence expression including a plurality of sets of terms indicating a co-occurrence relationship as basic elements, When the determination unit determines that the document to be searched is a document satisfying the condition for assigning the classification code, the classification code is assigned to the document to be searched.

  According to an eighth aspect of the present invention, the estimation dictionary data is a description of search conditions for similar documents described in terms of a grammatical expression, and the determination means includes a search condition for the search target documents for the similar documents. When it is determined that the document satisfies the above, the search target document is selected as a similar document.

  According to a ninth aspect of the present invention, the literature analysis program causes a computer to function as each means included in the literature analysis apparatus.

  According to a tenth aspect of the present invention, the document analysis method is a document analysis method performed by a document analysis device including a computer and a memory, and the document analysis device is configured to delimit a search condition with a document as a semantic expression. A process of storing in the memory as estimation dictionary data described by a lexical expression including a basic element in each range divided as a unit of the document, and the document analysis device decomposes the document data related to the document to be searched into word units And storing the terminology data in the memory as the terminology data for each range divided as a unit of a predetermined range that is delimited as a semantic expression. The semantic expressions included in the estimated dictionary data and the terms included in the literature term data stored in the memory are separated as semantic expressions. A process of comparing each range divided by a predetermined range as a unit and calculating a score indicating the degree of similarity, and the document analysis device, based on the calculated score, the search target document is the search And determining whether or not the document satisfies the condition.

  In an eleventh aspect of the present invention, the estimation dictionary creating apparatus has a function for automatically creating estimation dictionary data. The estimation dictionary creating device receives a theme to be used to create the estimation dictionary data and a seed term representing characteristics of the theme by the automatic creation function of the estimation dictionary data, and a literature database A teacher document extracting means for extracting a positive example teacher document group that includes the seed term and corresponds to the theme, and a negative example teacher document group that includes the seed term and does not correspond to the theme, and the extracted The seed term and its co-occurrence word for each range obtained by dividing each of the documents included in each of the positive example and the negative example teacher documents as a unit of a predetermined range that is delimited as a semantic expression. A basic element candidate extraction means for extracting a combination of words including at least one combination as a candidate for the basic element, and extracted by the basic element candidate extraction means A basic element selection means for selecting a basic element candidate that has a probability that is included in the positive example teacher document group and is not included in the negative example teacher document group from a plurality of basic element candidates; An estimation dictionary data creating unit that creates a sentence expression representing the estimation dictionary data based on the basic element candidates selected by the basic element selection unit and stores the estimation dictionary data in the estimation dictionary storage unit; It is a thing.

  According to a twelfth aspect of the present invention, the basic element candidate extraction means defines at least sentences and paragraphs as a predetermined range that is delimited as a semantic expression of the document, and each of the extracted positive examples and negative examples is extracted. First, a combination of terms including at least one combination of the seed term and its co-occurrence word is extracted as a candidate for the basic element from each document included in the document group for each sentence and each paragraph. Candidate extraction means, a co-occurrence word having a co-occurrence relationship with a basic element extracted from the sentence is extracted from another paragraph not including the sentence, and the basic element candidate extracted from the sentence and the other And a second candidate extracting means for adding a combination of product forms with the co-occurrence words extracted from the above paragraph to the basic element candidates.

In a thirteenth aspect of the present invention, the basic element candidate extraction unit further includes an extraction restriction unit. The extraction limiting means includes a first accuracy representing the ratio of the number of positive example teacher documents including basic elements to the sum of the number of positive example teacher documents including the basic elements and the number of negative example teacher documents, and the all positive example teacher documents. A threshold value set in advance is held for each of the first recall ratios representing the ratio of the number of positive example teacher documents including basic elements to the number. The extraction restriction means restricts the basic element candidates extracted by the first and second candidate extraction means to basic element candidates that exceed the accuracy threshold and the recall threshold. It is.

In a fourteenth aspect of the present invention, the basic element selection unit is configured to calculate the basic element with respect to the sum of the number of positive example teacher documents and the number of negative example teacher documents including the basic element candidates extracted by the basic element candidate extraction unit. The basic element candidates extracted by the basic element candidate extraction unit for the second accuracy representing the ratio of the number of positive example teacher documents including the candidates and the total number of positive example teacher documents extracted by the teacher document extraction unit. A basic element for estimation is selected from a plurality of basic element candidates extracted by the basic element candidate extraction means based on the second recall ratio indicating the ratio of the number of positive example teacher documents including Is.

Fifteenth aspect of the invention, the basic element selection means, from among the candidates of a plurality of basic elements extracted by the primitive candidate extracting means, the second precision and the second recall highest First means for selecting basic element candidates as basic elements for estimation, and after the selection of the basic elements for estimation, from the positive example teacher document group and the negative example teacher document group extracted by the document extraction unit The selection including a plurality of basic element candidates extracted by the basic element candidate extraction unit is performed for the remaining positive example teacher document group and negative example teacher document group, excluding documents including the estimation basic element. second means recalculating the second precision and the second recall for candidates of each basic element except the basic elements for estimation is, a plurality of base extracted by the primitive candidate extracting means Elements of the From candidates of all basic elements except the basic element for estimating said selected one of the complement, for estimating a candidate of the second precision and the second highest recall the basic element after the recalculation The third means for additionally selecting as a basic element, and the processing of the second means and the third means, the remaining positive example teacher literature is less than a predetermined number, or the selected for estimation And a fourth means that is repeatedly executed until the number of basic element candidates excluding the basic elements is equal to or less than a predetermined number.

  According to a sixteenth aspect of the present invention, an estimation dictionary creation program causes a computer to function as each means provided in the estimation dictionary creation device of the document analysis apparatus.

  According to the first aspect of the present invention, the estimation dictionary data describes the search condition by a grammatical expression including a basic element in each range that is divided in units of a predetermined range that is divided as a semantic expression. Since the similarity between the dictionary data for estimation and the document term data is expressed as a score, it is possible to recognize the degree of satisfaction of the search condition, not just the determination of whether or not the search was successful. It becomes.

  According to the second aspect of the present invention, the predetermined range delimited as the semantic expression is any one of a phrase, a sentence, and a paragraph. Therefore, the estimation dictionary data and the literature term data are compared within the respective ranges of the phrase, sentence, and paragraph, thereby searching for a set of terms across the phrase, sentence, and paragraph. The similarity determination with high accuracy is possible.

  According to the third aspect of the present invention, the presumed dictionary data includes, in the grammatical expression, a composite basic element obtained by hierarchically combining basic elements defined within the respective ranges of phrases, sentences, and paragraphs, The compound basic element includes a term and one or more phrase basic elements for a sentence basic element, and a term and one or more sentence basic elements for a paragraph basic element. Or it contains the basic elements of the phrase. For this reason, it is possible to easily define a co-occurrence relationship between terms in a broad sense, and it is possible to express strict sentence meaning by the relationship between terms. Furthermore, the co-occurrence relationship can be grasped with the meaning of the document as a unit without being aware of the distance between the terms, and the description is easy to use and conforms to the meaning of the sentence.

  According to the fourth aspect of the present invention, the grammatical expression is described using controlled words that are generic terms for a plurality of terms having the same meaning. For this reason, a plurality of terms having the same meaning can be simply defined by controlled words, and it is possible to achieve a reduction in time and effort of creating a grammatical expression and simplification of the appearance of the literary expression.

  According to the fifth aspect of the present invention, the search target range of the document to be searched is divided into a plurality of ranges in units of the first range having a unit as a semantic expression, and the range is determined for each of the divided ranges. The first score is calculated based on the coincidence frequency between the term group and the word meaning included in the term group, and the term group and the word meaning expression included in the second range for the second range wider than the first range. A second score is calculated based on the frequency of coincidence with the score, and based on the calculated first score and second score, a score representing the similarity between the estimation dictionary data and the document to be searched is obtained. Calculated. For this reason, for example, similarity determination is performed by paying attention to both scores of the entire search target document (second range) and local range (first range). It can be carried out.

  According to the sixth aspect of the present invention, the maximum score and the second score among the calculated first scores are weighted and added at a predetermined ratio. For this reason, it is possible to optimally set the ratio of the influence of the entire document to the similarity and the influence of the local range of the document.

  According to the seventh aspect of the present invention, when it is determined that the search target document satisfies the search condition, a classification code is assigned to the search target document. Accordingly, it is possible to automatically assign a classification code to a document.

  According to the eighth aspect of the present invention, when it is determined that the search target document satisfies the search condition, the search target document is selected as a similar document. That is, it is possible to automatically search for similar documents.

  According to the ninth aspect of the present invention, the computer can function as each means provided in the document analysis apparatus by the document analysis program.

  According to the tenth aspect of the present invention, the estimation dictionary data describes the search condition by a grammatical expression including a basic element in each range divided by a predetermined range that is delimited by a document as a semantic expression. Since the similarity between the dictionary data for estimation and the document term data is expressed as a score, it is possible to recognize the degree of satisfaction of the search condition, not just the determination of whether or not the search was successful. It becomes.

  According to the eleventh aspect of the present invention, first, a positive example teacher document group including seed terms and corresponding to a theme and a negative example teacher document group including the seed terms and not corresponding to the theme are extracted from a document database. Then, a combination of phrases including at least one combination of a seed term and its co-occurrence word is extracted as a basic element candidate from these document groups. Then, from among the plurality of extracted basic element candidates, a basic element candidate whose probability of being included in the positive example teacher document group and not included in the negative example teacher document group is not less than a threshold is selected. A sentence expression including the basic element is created as estimation dictionary data. Therefore, it becomes possible to automatically create estimation dictionary data, which eliminates the need for a person to manually create an estimation dictionary, greatly reduces human workload, An early start of operation can be realized.

  According to the twelfth aspect of the present invention, when extracting basic element candidates, basic element candidates are extracted at least for each sentence and for each paragraph from each of the positive and negative teacher document groups. In addition, a co-occurrence word having a co-occurrence relationship with the basic element extracted from the above sentence is extracted from another paragraph not including the sentence, and the product form of the extracted basic element candidate and the co-occurrence word is used. Combinations are added to the basic element candidates. For this reason, candidates for basic elements are extracted for each document, at least for each sentence and paragraph, and even across paragraphs. Therefore, basic element candidates can be extracted without omission.

According to the thirteenth aspect of the present invention, the basic element candidates to be extracted are limited to those exceeding a preset first accuracy threshold and second recall ratio threshold. For this reason, the number of basic element candidates extracted can be effectively limited, and thereby the number of estimation dictionary data can be suppressed. Further, it is possible to reduce the processing load on the apparatus required for creating the estimation dictionary data, and to shorten the processing time.

According to the fourteenth aspect of the present invention, the basic element for estimation is selected in consideration of both the second accuracy of the basic element candidate and the second recall. For this reason, it is possible to extract necessary and sufficient basic elements as basic elements for estimation.

According to the fifteenth aspect of the present invention, when selecting the basic element for estimation based on the second accuracy and the second recall, the second element is selected every time one basic element is selected . The accuracy and the second recall are recalculated, and a basic element for estimation is selected based on the recalculated second accuracy and the second recall. For this reason, it becomes possible to effectively select the basic elements necessary for creating the estimation dictionary data without causing duplication or deficiency.

  According to the sixteenth aspect of the present invention, the computer can function as each means provided in the document analysis apparatus by the document analysis program.

  In other words, according to each aspect of the present invention, it is possible to simply and accurately define the grant conditions or search conditions described in the dictionary, thereby enabling high-accuracy classification and literature search in a short processing time. An apparatus, a literature analysis method, and a program can be provided.

The figure which shows the outline | summary of a system provided with the literature analysis apparatus which concerns on Example 1 of this invention. The figure which shows an example of the display screen used for preparation of the dictionary for presumption. The figure which shows the structure of the sentence meaning described in the dictionary for estimation. The flowchart which shows the procedure of the classification provision process performed by the literature analysis apparatus shown in FIG. 10 is a flowchart illustrating a document search processing procedure and processing contents performed by the document analysis system according to the second embodiment. 10 is a flowchart illustrating a document search processing procedure and processing contents performed by the document analysis system according to the third embodiment. The block diagram which shows the function structure of the dictionary creation apparatus for estimation used in the literature analyzer which concerns on Example 4 of this invention. 8 is a flowchart showing a creation processing procedure and processing contents of the estimation dictionary automatic creation device shown in FIG. 7. The flowchart which shows the process sequence of the basic element candidate extraction and storage process in the creation process procedure shown in FIG. FIG. 9 is a flowchart showing a processing procedure and processing contents when effective basic element selection / storage processing in the creation processing procedure shown in FIG. 8 is executed focusing on accuracy. FIG. 9 is a flowchart showing a processing procedure and processing contents when effective basic element selection / storage processing in the creation processing procedure shown in FIG. The figure which shows an example of the parameter taken in by the parameter taking process shown in FIG. The figure which shows an example of the list | wrist of the positive example teacher literature and the negative example teacher literature produced by the division | segmentation process of the teacher literature extraction process and positive example / negative example teacher literature shown in FIG. The figure which shows the outline | summary of the effective basic element selection process shown in FIG. The figure which shows an example of the basic element candidate extracted by the basic element candidate extraction process shown in FIG. The figure which shows an example of the basic element candidate updated in the selection process process of the effective basic element selection process shown in FIG. 10 and FIG.

Embodiments according to the present invention will be described below with reference to the drawings.
[Example 1]
FIG. 1 is a schematic configuration diagram of a document analysis system according to a first embodiment used when the present invention assigns a classification code to a patent document.
The document analysis system according to the first embodiment includes an estimation dictionary creation device 1, a document term table generation device 2, and an estimation device 3. These devices 1, 2 and 3 are respectively constituted by independent computers, but it is also possible to adopt a configuration in which the processing functions of the devices 1, 2 and 3 are given to one computer.

  The estimation dictionary creating apparatus 1 has an estimation dictionary editor 1A. The estimation dictionary editor 1A displays the estimation dictionary creation template 40 stored in the memory on the display unit, and the estimation dictionary based on the parameters and sentence expressions manually input by the estimation dictionary creator. It has a function of creating data P and storing the created estimation dictionary data P in the memory again.

  The estimation dictionary creating apparatus 1 has a function of reading a plurality of document data of sample documents prepared for creating an estimation dictionary and automatically creating estimation dictionary data using machine learning or the like. May be.

  The document terminology table generation device 2 reads the document data Q1, Q2, Q3,... To which the classification code is to be applied from an external document data storage device, and from the document data Q1, Q2, Q3,. Each has a function of generating document term table data R1, R2, R3,. The sentence meaning filter unit 2A divides the document data Q1, Q2, Q3,... Into word units by morphological analysis, and extracts and synthesizes necessary words and phrases from the divided words to thereby generate document term table data. R1, R2, R3,... Are generated.

  The estimation device 3 includes an estimation engine 3A and a difference extraction tool 3B.

  The estimation engine 3A fetches and stores the estimation dictionary data P and the literature term table data R1, R2, R3,... From the estimation dictionary creation device 1 and the literature term table generation device 2, respectively. In this state, the syntactic expression of the dictionary data for estimation P is compared with the terms in the literature term table data R1, R2, R3,. Based on the score calculation result, it is determined whether or not a classification code can be given. If yes, a classification code is given.

  The difference extraction tool 3B is used to increase the accuracy of the estimation dictionary data P. The difference extraction tool 3B collects the classification codes assigned to the literature group by the estimation engine 3A, and the assigned classification codes and those documents (sample literature) ) Is compared with a classification code given in advance by a person as a correct answer example, and if there is a difference in the assigned code between them, it is detected and the difference result is displayed. As a result, it is possible to overlook the application omission and unnecessary application of the estimation addition code to the sample document group, and it is possible to grasp the problems of the estimation dictionary as a whole.

(Operation)
Next, the classification giving operation by the system will be described.
Here, a case where a classification code is assigned to a patent document will be described as an example.

(1) Creation of Estimation Dictionary The estimation dictionary creation device 1 creates estimation dictionary data P under the control of the estimation dictionary editor 1A. The estimation dictionary data includes a code assignment definition table and a meta-rule that additionally instructs further code assignment using the code assignment result or instructs cancellation of code assignment. In addition, the code assignment definition table provides a function for defining controlled words and literary expressions in order to support the simple and simple creation and description of literary expressions.

  The estimation dictionary editor 1A first reads and displays the estimation dictionary creation template 40 stored in advance in the estimation dictionary creation memory. FIG. 2 shows a display example of the estimation dictionary creation template 40. A code assignment definition table is a template in which a sentence expression is entered.

  The estimation dictionary creation template 40 is provided with, for example, a plurality of parameters and a plurality of sentence input cells. Note that FIG. 2 shows a state in which parameters and a semantic expression have already been input for convenience of explanation, but no input is made in the initial state. This template can be used to define code assignment conditions for the required code.

(1-1) Parameters Parameters consist of “theme”, “code”, “group”, “grant reference value”, “weight”, and “paragraph type number”, and each is assigned to one column in the table. ing. In addition, “literary meaning”, “basic score of term”, “basic score of basic element”, and “basic score of literary expression” are also assigned to the table columns one by one. A plurality of sentence expressions are also assigned to the columns of the table one by one.

  “Theme” is the name of the field assigned to the technical field to be classified, and “Code” is the name of the code to be assigned, for example, F term and FI code defined by the Japanese Patent Office . The code can be described in a plurality of lines, but the line is changed particularly when the parameters are different, and the number in the code is described in each line in the “group” cell. “Granting reference value” specifies a threshold value for determining whether or not a code can be given. “Weight” designates the evaluation value magnification when calculating the matching score of the meaning expression described in the corresponding group. The “paragraph type number” is used to specify a search target range of patent documents in units of items: 1: “name of invention”, 2: “summary”, 3: “claim”, 4: “technical field”, 5: “Background / Prior art”, 6: “Problem”, 7: “Solution means”, 8: “Effect / action”, 9: “Explanation of the figure”, 10: “Example”, 11: “Reference sign” Is defined as follows.

(1-2) Written Expression The Written Expression uses a single term or a plurality of terms indicating a co-occurrence relationship as a basic element (hereinafter simply referred to as “basic element”) as a condition for assigning a classification. A basic element or a combination of a single term and a basic element is described by a Boolean expression.

  The basic element is the same as the phrase basic element that specifies multiple terms that exist in the same phrase as a set, and the sentence basic element that specifies multiple terms that exist in the same sentence as a pair There are three types of “paragraph basic elements” that specify a set of terms that exist within a paragraph. These elements are described using <>, {}, and [], respectively. Here, a phrase is a group of characters in a range delimited by either a punctuation mark (“,”) or a punctuation mark (“.”), And a sentence is within the range delimited by a punctuation mark (“.”). A group of characters, and a paragraph is a group of characters within a range delimited by an explicit or implied paragraph delimiter (such as indentation at the beginning of a line). A document contains one or more paragraphs, a paragraph contains one or more sentences, a sentence contains one or more phrases (if there are no punctuation in the sentence, the sentences and phrases are the same), and the phrase Includes one or more terms. Furthermore, these form a hierarchical structure of documents-paragraphs-sentences-phrases-terms with the documents at the top.

Examples of basic elements are shown below. That is, when specifying the co-occurrence relationships of the terms “position”, “GPS”, and “measurement”, the “phrase basic element”, “text basic element”, and “paragraph basic element” are <position, GPS, and measurement, respectively. >
{Location, GPS, Measurement}
[Position, GPS, Measurement]
Is described.

  By defining three basic elements, "phrase basic element", "sentence basic element" and "paragraph basic element", they are specified in the basic element only within each range of phrase, sentence and paragraph It can be searched whether a set of terms appears. Thereby, a set of terms is not searched (hit) across the boundaries of phrases, sentences, and paragraphs, and it is possible to specify sentence meaning with high accuracy.

The basic element can include a word order or a symbol designating a negative form. To specify the word order, for example, <position, GPS, measurement> c
In this way, “c” is added to the end of the basic element. In this case, it is specified that the terms “position”, “GPS”, and “measurement” should appear in the search range in this arrangement order.

To specify a negative form, for example, {position, GPS, measurement, #camera}
Add “#” to the beginning of the term to be negated. In the negative form, a hit condition is that a term designated as a negative form does not exist in each range of a phrase, sentence, or paragraph. In this example, “position”, “GPS”, and “measurement” exist within the range, and “camera” must not exist. "#" May be added to any number of terms in the basic element.

  Furthermore, when writing a grammatical expression, it is necessary to specify synonyms for each term without omission, and for this purpose, the controlled words b1 to bm can be defined as shown in FIG. The controlled words b1 to bm represent a plurality of synonyms c1 to cm. A controlled word dictionary that defines and stores a plurality of synonyms c1 to cm belonging to the controlled word for each controlled word b1 to bm is provided, and the controlled word is managed by this dictionary. Controlled words are expressed by adding “;” to the beginning of the sentence meaning.

  As an example, the term “computer” may use other terms such as “computer”, “computer”, and “PC”. So, if you define a controlled word “; PC” and specify a group of terms that have these similar meanings in it, by writing the controlled word “; PC” in the narrative, After that, the control side expands the controlled word into the form of the logical sum of the synonym groups defined therein. It is possible to reduce the time and effort of creating a sentence expression and simplify the appearance. The controlled word is registered in a controlled word dictionary (not shown) in association with a synonym group included in the controlled word. Controlled words registered in the controlled word dictionary can be described as an alternative to terms in the sentence meaning.

  It is also possible to describe a composite basic element having a multilayer structure (hierarchical structure) including the basic element in the basic element. In the compound basic element, for example, it is possible to describe a phrase basic element in a sentence basic element, describe a sentence basic element in a paragraph basic element, and further describe a phrase basic element in a sentence basic element Become.

A description example of a composite basic element that also serves as an example of a controlled word is shown below.
{<Stop, hour>c,; correction}
[; Navi end, {; Pedestrian, <Position,;Measurement> c}]
Here, the difference between the composite basic element “{<stop, time>c,; correction}” and the literal expression “<stop, time> c * {stop, time,; correction}” will be described.

  In the former compound basic element, the whole is surrounded by {} symbols indicating the basic element of the sentence, so <stop, hour> c and {stop, hour, correction} must appear in the sentence. Instruct. That is, only when “stop”, “time”, and “; correction” exist in the same sentence and “stop” and “time” are included in any phrase included in the sentence, The basic element is determined to have been hit.

  On the other hand, in the latter case, <stop, time> c and {stop, time, correction} may appear in different sentences. It means that each should be somewhere in the literature.

  Similarly, in [; Navi end, {; Pedestrian, <Position,; Measurement> c}], the whole is enclosed with a symbol [] indicating a paragraph basic element, and a symbol {} indicating a sentence basic element therein. And the symbol <> indicating a phrase basic element are included, so that the specified basic element satisfaction condition is satisfied in the same paragraph, in the same sentence, and in the same phrase in the sentence. Instructed that it should.

  A sentence expression is a Boolean algebraic expression in which one term or one basic element is used as a term, and one or more of these terms are combined. If there are multiple terms, the logical sum "+ ”Operator or logical product“ * ”. Furthermore, when there are a plurality of operators, the application order of the operators can be specified using the parenthesis operator “()” of the operation priority specification. A controlled word may be specified instead of a term in the basic element or instead of a single term.

  In addition, in the “F term” of the patent classification, for example, “AA04” is not given when “AA03” is given, and “BB03” is given when “AA10” and “AA15” are given. Is given. In order to deal with this rule, a high-order expression describing a grant condition between F terms called a meta rule may be defined.

  Meta rules are composed of conditional expressions and action expressions. The conditional expression is expressed by a sentence expression combining F terms. When the conditional expression is satisfied, the action expression is executed. In the action expression, code addition / deletion, message display, user intervention request, etc. are designated.

The meta-rule can be applied to the “FI code”. For example,
“If either“ AA10 ”and“ AA12 ”and“ BB03 ”and“ BB04 ”are assigned, give G06Fxxx @ A.”
Write a meta-rule. In this way, the same effect can be obtained for the FI code.

An example of the case where the meta rule is applied to the provision of the F term and the case of application to the provision of the FI code are shown below.
-Application example for F-term assignment Conditional expression; (AA03 + AA04 + # AA05) * AA08 = 1,
Action formula; AA10 = 1
The above conditional expression represents a condition that AA10 is given when either AA03 or AA04 is given or AA05 is not given and AA08 is given.

-Application example to FI code assignment If the conditional expression is (AA03 + AA04) * {BB01 to BB15> = 2},
The action formula is G01C21 / 00 @ p = 1.
In the above conditional expression, if either AA03 or AA04 is given and two or more codes are given to any of the codes from BB01 to BB15 (BB01, BB02, BB03,..., BB15), It represents the condition of giving G01C21 / 00 @ p.

(1-3) Estimating Dictionary Creation Procedure and Processing Contents The estimation dictionary creator first inputs the theme name and code to be created in the state where the estimation dictionary creation template 40 is displayed. Enter the group number, assignment reference value, weight, annotation, and paragraph type number for each. Next, the textual expressions corresponding to the code assignment conditions are described one by one in the textual expression input cell. The word expressions described in a plurality of cells in the same row are combined and processed by logical sum. Therefore, even if these plural sentence expressions are combined into one sentence expression and described in one cell, the effect is the same. However, if it is divided into a plurality of sentence expressions and described in cells, the sentence expressions can be corrected in cell units.

  As described in (1-2) above, the sentence meaning is based on a single element or a combination of multiple terms indicating a co-occurrence relationship for each phrase, sentence, and paragraph, and a combination of these basic elements. It is created by connecting with the operator. At that time, it is also possible to specify word order, negative form, use of controlled words, and description of complex basic elements.

  In order to improve the accuracy of the estimation dictionary data, in this embodiment, the following estimation dictionary creation procedure is executed. That is, first, the estimation dictionary creator carefully selects about 10 sample documents corresponding to the theme, and the classification dictionary that the estimation dictionary creator thinks is absolutely correct with his insight to these documents, This is the absolute reference.

  First, as described above, the first version of the estimation dictionary is created by inputting the parameters and the meaning expression, and then the estimation device 3 is actually classified into the sample document using the first version of the estimation dictionary. Give a code. Then, the classification code given by the estimation device 3 is compared with the absolute reference classification code to determine whether or not they match. The estimation dictionary data is changed, and the estimation dictionary correction operation is repeated until both classification codes match. When both match, this is the second version of the estimation dictionary.

  Next, after about 10 new sample documents are selected and absolute reference codes are given, classification codes are given to these sample documents by the estimation device 3 based on the second version of the estimation dictionary. Then, based on the difference between the assigned classification code and the absolute reference classification code, a sentence expression is further added or corrected to create a third version of the estimation dictionary.

  Thereafter, similarly, the estimation dictionary correction process is repeated, and when the difference between the classification code given by the estimation device 3 based on the estimation dictionary and the absolute reference classification code converges within a certain range, the estimation dictionary is used. Finish creating the dictionary. Note that the accuracy of the estimation dictionary is expected to decrease due to the advancement or change in the technical contents described in the patent literature. For this reason, it is desirable that the estimation dictionary is regularly maintained by the above-described method.

(2) Generation of document terminology table data of documents to be assigned For the purpose of assigning codes to documents, it is necessary to create a document terminology table for each document. The classification granter designates the number of one or more patent documents to be classified in the literature term table generation device 2. Then, the document term table generation device 2 first reads the document data Q1, Q2, Q3,... Of the designated patent document from a document data storage device (not shown). Subsequently, the sentence filter unit 2A divides the read document data Q1, Q2, Q3,... Into each document by word unit using morphological analysis, and a necessary phrase from the divided words is obtained. By extracting and synthesizing, literature term table data R1, R2, R3,... Are generated.

It supplements about the production | generation of literature glossary table data. If the word output by the morphological analysis is a noun, a compound noun is generated. For example, when a term such as “information processing dictionary” is included in the document data, it is divided into three words “information”, “processing”, and “dictionary” in the morphological analysis. These divided words are synthesized and restored to terms such as “information processing” and “information processing dictionary” and output to the literature term table data. Since a chemical formula such as “H ₂ SO ₄ ” is also divided and output into English letters and numbers in the morphological analysis, a process for restoring the original chemical formula is performed. Since verbs and adjectives change their endings, the final form is also output along with the changed word in the document data. For example, “run” is also generated for “run” and “run”.

  Furthermore, a term number, a phrase number, a sentence number, and a paragraph number, which are serial numbers from the beginning of the document, are also assigned to all the terms output in the document term table. Using these serial numbers, each term in the basic element appears in the same phrase / sentence / paragraph when comparing the phrase basic element, sentence basic element, paragraph basic element, etc. of the dictionary for estimation with the literature term table Judgment is made at high speed. Literature term table data R1, R2, R3,... Including such information is stored in a memory in the literature term table generation device 2.

(3) Estimation of classification code (3-1) Acquisition of dictionary data for estimation and literature term table The classification person in charge specifies a theme to be classified from the terminal 4 to the estimation device 3 and inputs a classification start command. . Then, the estimation device 3 starts the classification code estimation process as follows. FIG. 4 is a flowchart showing the processing procedure and processing contents.

  Here, a case where a plurality of classification codes defined in the estimation dictionary are sequentially assigned to a plurality of documents one by one will be described as an example. Note that the application of the meta-rule is performed for reviewing the code assignment result for the document when the assignment of all classification codes to the document is completed.

  When it is detected in step S11 that the classification start command has been input, first, in step S12, estimation dictionary data P corresponding to the specified theme is read from the estimation dictionary creation device 1. The estimation dictionary data P includes a code assignment definition table and a meta rule dictionary.

  At the same time, the estimation device 3 reads the document term table data R1, R2, R3,... From the document term table generation device 2 in step S12. A classification code is assigned to a document in units of documents. When a plurality of codes are defined in the estimation dictionary data P, code assignment is repeatedly performed for one document by the process described below. When assigning codes to a plurality of documents, this process is repeated for the number of documents. Hereinafter, a case where one code is assigned to one document will be described.

(3-2) Matching and Score Calculation The estimation device 3 reads the document terminology table data Rn of the document to be assigned in step S12, and sets the grant environment in step S13. It is assumed that a plurality of classification codes are defined in the estimation dictionary data, and a classification code to be assigned next is set in step S14. Next, in step S15, the estimation device 3 compares the term described in the document terminology table data Rn with the grammatical expression described in the estimation dictionary data, and calculates a score according to the degree of coincidence.

(3-3) Detailed Calculation Example of Score The score is calculated as follows.
That is, first, for each line (classification code group) of the dictionary data for estimation, the value of the coincidence frequency between the word meaning expression and the term described in the document terminology table data is obtained by the following formula, and the values of all the paragraphs are obtained. Is the score of the literature.

Wi * {log ₁₀ (N1 * f1 + N2 * f2 + N3 * f3)
Wi; weight coefficient N1; basic score of term (constant)
N2: Basic score of basic element (constant)
N3: Basic score of literal meaning (constant)
f1; normalized number of hits in the same group when focusing on a set of terms f2; normalized number of hits in the same group when focusing on a set of basic elements f3; focusing on a set of literal expressions Normalized number of hits in the same group When the scores of N1, N2, and N3 are constants determined according to the complexity of terms, basic elements, and literary expressions, respectively, the complexity is high. The score as a constant is increased.

  In addition, the normalization is defined as B / when “the average number of characters included in each paragraph of all documents included in the classification theme” is A and “the number of characters included in the paragraph of the document to be currently assigned” is B. This is calculated by multiplying the number of characters in the literature by using A as a normalization coefficient.

  That is, the description content to be searched for is expressed as a set of terms + a set of basic elements + a set of sentence expressions. Subsequently, the sum of the scores of all the groups with the same classification code is calculated for each paragraph, and the score of the entire document is calculated as the sum.

The above score calculation method will be described more specifically.
Now, it is assumed that a two-dimensional table (illustrated in FIG. 2) in which the grant conditions are described by a lexical expression is defined as follows.

Dictionary data group for estimation of F-term “DD19” (theme code “2F129”) Group 1;
Interpretation 1 “; Navigation device”
Standard value “0.6”
Weight “1”
Paragraph type number “1, 2, 3, 4, 6, 7, 8, 9”
Term basic score N1 "1"
Basic score N2 “1” of the basic element is a grammatical expression composed only of terms, and N2 is virtually meaningless, so any number may be used as long as it is 1 or more.
Basic score N3 “1” of the word meaning expression is a word meaning expression composed only of terms, and N3 is practically meaningless and may be any number as long as it is 1 or more.

Group 2;
Grammatical expression 1 “{<;route,;neighborhood;facility> c, {; facility,; search} c}”
Standard value “0.6”
Weight “2”
Paragraph type number “1, 2, 3, 4, 6, 7, 8”
Term basic score N1 "1"
Basic score of basic element N2 “2”
Basic score of literary expression "2"
Grammatical expression 2 “{<;route,; periphery, point> c, {point,; search} c}”
Standard value “0.6”
Weight “2”
Paragraph type number “1, 2, 3, 4, 6, 7, 8”
Term basic score N1 "1"
Basic score of basic element N2 “2”
The basic score of literary expression is “2”.

Further, assume that f1 to f3 are as follows.
Group 1
When attention is paid to a set of terms of the grammatical expression 1 in group 1, if the hit number h1 in the same group is “30” and the normalization coefficient is “0.8”, f1 is “24”.
When the number of hits h2 in the same group is “0 (zero)” and the normalization coefficient is “0.8” when focusing on the set of basic elements of the semantic expression 1 in group 1, f2 is “0 (zero). ) "
Since the semantic expression 1 in the group 1 is a conceptual expression composed only of terms, there is no set of basic elements, so the number of hits is “0 (zero)”.
Similarly, when focusing on the set of semantic expressions of semantic expression 1 in group 1, if the number of hits h3 in the same group is “0 (zero)” and the normalization coefficient is “0.8”, f3 is “0”. Since the meaning 1 in the group 1 is a meaning made up of only terms, there is no set of meanings, so the number of hits is “0 (zero)”.

Group 2
When attention is paid to a set of terms of the grammatical expression 1 in group 2, if the number of hits h1 in the same group is “5” and the normalization coefficient is “0.9”, f1 is “4.5”.
When focusing on the set of basic elements of the semantic expression 1 in group 2, if the number of hits h2 in the same group is “3” and the normalization coefficient is “0.9”, f2 is “2.7”.
Similarly, when focusing on the set of semantic expressions of semantic expression 1 in group 2, if the number of hits h3 in the same group is “2” and the normalization coefficient is “0.9”, f3 is “1. 8 ".

Group 2
When attention is paid to a set of terms in the meaning expression 2 in the group 2, the number of hits h1 in the same group is “4”, and the normalization coefficient is “0.9”, and f1 is “3.6”.
When focusing on the set of basic elements of the semantic expression 2 in group 2, if the number of hits h2 in the same group is “3” and the normalization coefficient is “0.9”, f2 is “2.7”.
Similarly, if the number of hits h3 in the same group when focusing on the set of semantic expressions of semantic expression 2 in group 2 is “2” and the normalization coefficient is “0.9”, f3 is “1. 8 ".

Therefore, each score is as follows.
Score of ideographic formula 1 of group 1:
1 (weight) × {log ₁₀ (1 (N1) × 24 (f1) +1 (N2) × 0 (f2) +1 (N3) × 0 (f3))} = 1.38
Score of grammatical expression 1 of group 2:
2 (weight) × {log ₁₀ (1 (N1) × 4.5 (f1) +2 (N2) × 2.7 (f2) +2 (N3) × 1.8 (f3))} = 2.26
Score of grammatical formula 2 of group 2:
2 (weight) × {log ₁₀ (1 (N1) × 3.6 (f1) +2 (N2) × 2.7 (f2) +2 (N3) × 1.8 (f3))} = 2.20.

Each of these scores and the reference value “0.6” are compared, and those above the reference value are extracted.
Note that “the base score Tt of the entire target document = the sum of the scores for each sentence expression”.

Therefore,
1.38 (score of group 1 sentence 1)
+ 2.26 (score for grammatical expression 1 of group 2)
+ 2.20 (score for group 2 sentence 2)
= 5.85 (Tt) ...
This is the score of the target document.

Then, the score T = αTt + βmaxTd (i) of the target document (target text) is calculated. here,
α, β; coefficient of weight Tt: base score of the entire target document
maxTd (i); the maximum score of the target paragraph type, for example, α = 1, β = 4, and assuming that maxTd (i) is the maximum of “10” in the paragraph type 10,
The score T of the target document (target text) is
1 (α) × 5.85 (Tt) +4 (β) × 10 (maxTd (10)) = 45.85

  Therefore, in this case, the score T of the target document (target text) is “45.85”. The score T calculated as described above is stored.

(3-4) Determination of Matching Score (Evaluation Value) When the score calculation is completed, the estimation device 3 reads the score T in step S16 and compares the score T with a preset threshold value. In step S17, it is determined whether or not the score T is equal to or greater than a threshold value. That is, it is determined whether or not the classification code should be assigned for each classification code with respect to the currently processed document.

(3-5) Assignment of Classification Code When the estimation apparatus 3 finds a classification code to be assigned, it assigns the classification code to the patent document to be processed in step S18. Specifically, the F term or FI code is associated with the number of the patent document to be granted. When the assignment of one classification code is completed, it is determined in step 19 whether or not all the classification code assignment processes have been completed. If not, the process returns to step 14 in order to perform the process for the next classification code.

  When the assignment of all the classification codes to the document is completed, the metarule dictionary is read out in step S20, and the assignment is not given to the finished document according to the code assignment result and the metarule to the document so far. New code is assigned or the assigned code is canceled.

  In step S21, the number of the patent document to be granted and the assigned classification code are output to the terminal 4 and displayed.

When the process of assigning all the classification codes defined in the dictionary data for estimation to one document is finished, it is determined in step S22 whether or not the assignment to all the documents is finished. The code assigning operation from S12 to the document is repeated. The process ends when the granting process for all documents is completed.
Note that it is also possible to output the document term table data of the second document group as the grant basis data after the grant is completed.

(effect)
As described above in detail, in the first embodiment, the estimation dictionary data P is assigned a classification code using at least one basic element that is a set of a plurality of terms indicating a co-occurrence relationship for each phrase, sentence, and paragraph. The sentence meaning describing the condition is stored. Then, the document terminology table data R1, R2, R3,... Generated from the document to be given is compared with the grammatical expressions described for each classification code in the estimation dictionary data P and basic elements as basic units, The degree of coincidence is calculated as a score, and when the total value of the scores is equal to or greater than a threshold, the classification code is assigned to the document to be given.

  Therefore, according to the first embodiment, the literal expression is described by using one term or a plurality of terms indicating a co-occurrence relationship as a basic element and using at least one of them as a basic element. It becomes possible to accurately reflect the satisfaction of the condition in the score. In addition, since the word meaning is defined for each phrase, sentence, and paragraph, when searching for the corresponding term from the literature term table data R1, R2, R3,. Done within range. For this reason, a set of terms is not searched across phrases, sentences, and paragraphs, and sentence meaning can be estimated with high accuracy.

  In the first embodiment, when the dictionary data P for estimation is created, the condition is divided into simple condition groups, and these condition groups are expressed as independent sentence meanings. One table is written in each cell of the table creation application. For this reason, complicated provision conditions can be simplified and expressed, thereby making it possible to easily create, change, and add conditional expressions.

  Furthermore, in Example 1, when the score is calculated, the score maxTd (i) of the paragraph having the highest score among all the paragraphs of the literature term table data R1, R2, R3,. Tt is obtained by adding weights of 1: 4.

  In general, a technique that reflects only the description content of the entire document in the score can roughly search for technical terms of the entire document, but can only determine a vague similarity. On the other hand, when the similarity is determined based only on the paragraph with the highest score, it is likely to be a determination result in which the forest is not seen by looking at the tree. However, according to the first embodiment, paying attention to the scores of all the paragraphs of the document to be granted and the paragraph with the highest score, the both scores are weighted and added according to their importance to determine the similarity. Therefore, the similarity determination with high accuracy can be performed.

    Furthermore, the basis of grant can also be displayed, and a person can easily determine the validity of the grant.

[Example 2]
In the first embodiment, the case where a classification code is assigned to a patent document has been described as an example. On the other hand, the second embodiment of the present invention relates to a case where a second document having a similar technical content to a specific first document is searched from a document set to be searched (referred to as “second document group”). The literature analysis system according to the above is used.

  The basic structure of the document analysis system according to the second embodiment is the same as that described in the first embodiment, and will be described with reference to FIG.

  FIG. 5 is a flowchart showing the procedure and processing contents of a document search process performed by the document analysis system according to Embodiment 2 of the present invention.

  The person in charge of the search uses the estimation dictionary editor 1A of the estimation dictionary creation apparatus 1 to specify predetermined items of the first document as a search source, for example, technical field, background art / prior art, examples, and claims. For each paragraph described, estimation dictionary data in which the search condition is described by a lexical expression is created. Similar to the first embodiment, the sentence meaning is based on one phrase or a plurality of terms showing a co-occurrence relationship for each phrase, sentence, and paragraph, and at least one of them is used as a sentence meaning expression. It is described. Also, word order and negative form can be specified in the sentence expression. If a classification code is assigned to the first document, a sentence meaning expression corresponding to the classification code may be created as it is or by processing the sentence meaning expression group.

  The estimation dictionary data P created as described above is stored in the memory in the estimation dictionary creation device 1. The estimation dictionary creation device 1 may be provided with a function of reading document data of a patent document as a search source and automatically creating the estimation dictionary data for search using machine learning or the like. The estimation device 3 reads the dictionary data for estimation of the first document stored in the memory of the dictionary creation device for estimation 1 in step S31.

  Next, in step 33, the estimation device 3 generates document term table data in each document of the second document group (document set to be searched) created by the document term table generation device 2. This literature term table data is the same as that used in the first embodiment, and the generation method is also the same. Further, new information for the purpose may be added to the literature term table data so that the concept search can be performed.

  Next, in step S33, the estimating apparatus 3 reads the document term table data R1, R2, R3,. Then, in step S34, it is compared with the sentence meaning described in each group of the estimation dictionary data P. As a result of this comparison, when a set of terms that match the basic elements described in the grammatical expression is found from the literature term table data, a score is calculated for each row according to the number of matches.

  As in the first embodiment, this score calculation method is first calculated as “a set of terms + a set of basic elements + a set of semantic expressions”, and the total score of all lines is calculated for each document and each paragraph. Calculate each. Then, a method of weighting and adding the score maxTd (i) of the paragraph having the highest score among all the paragraphs and the total value Tt of the scores of all the paragraphs at a ratio of 1: 4 is used. The score calculated for each second document is stored in association with the number of the second document.

  Then, the stored score is read, and this score is compared with a preset first threshold value. In step S36, it is determined whether or not the total score value is equal to or greater than the first threshold value. That is, it is determined whether the technical content similarity between the second document and the first document is greater than or equal to a threshold value. As a result of this determination, if a second document having a score representing the similarity equal to or higher than the first threshold is found, the second document is output to the terminal as a similar document in step S37.

  On the other hand, when the second document whose score representing the similarity is equal to or greater than the first threshold is not found, the search result output unit has a score representing the similarity in step S38 that is less than the first threshold. Is selected as a reference document. Then, the number of the second document selected as the reference document is output to the terminal.

  As described above in detail, in the second embodiment, when searching for a second document similar to the first document, a single term or a set of terms indicating a co-occurrence relationship based on the contents of the first document is used. A sentence expression describing a search condition is created using basic elements, and this is stored in an estimation dictionary. Then, the word meaning stored in the estimation dictionary is compared with the document term table data respectively generated from the plurality of second documents to be searched for in terms of terms, and the degree of matching is calculated as a score. When the total score is equal to or greater than the threshold, the second document is selected as a similar document.

  Therefore, according to the second embodiment, the sentence expression is described by using at least one of a plurality of terms indicating one term or a co-occurrence relationship as a basic element for each phrase, sentence, and paragraph. Thus, the degree of matching of terms between documents can be accurately reflected in the score. In addition, since the similarity between the estimation dictionary data and the document term table data is represented by a score, it is possible to determine the degree of similarity without being limited to determining whether the document is a similar document.

  In addition, since the word meaning is defined for each phrase, sentence, and paragraph, when searching for the corresponding term from the literature term table data R1, R2, R3,. Done within range. For this reason, a set of terms is not searched across phrases, sentences, and paragraphs, and sentence meaning can be estimated with high accuracy.

  In the second embodiment, when the estimation dictionary is created, the search condition is decomposed into a plurality of simple conditions, and these conditions are expressed as independent sentence meanings. One cell is described in each cell. For this reason, complicated search conditions can be expressed in a simplified manner, thereby making it possible to easily create, change, and add conditional expressions.

  Furthermore, in Example 2, when determining the similarity between the first document and the second document, the paragraph having the highest score is selected from all the paragraphs of the second document, and the score of the selected paragraph and the second score are selected. The total score of all paragraphs in the document is obtained by weighted addition at a ratio of 1: 4, and the similarity is determined based on the obtained score. For this reason, it is possible to perform the similarity determination with high accuracy in consideration of both the whole document and the paragraph in which the abstract is described in a well-balanced manner.

[Example 3]
The second embodiment is an example in which a document that is close to the content of the first document is searched from the second document group, but the first document is not assumed, and a document that simply describes a certain content is searched from the second document group. You may want to take it out. In this case, the content to be searched for is expressed using one or a plurality of grammatical expressions, and this is created as a definition table to be used as an estimation dictionary.

FIG. 6 is a flowchart showing the procedure and processing contents of a document search process performed by the document analysis system according to Embodiment 3 of the present invention.
The person in charge of the search uses the estimation dictionary editor 1A of the estimation dictionary creation device 1 to create estimation dictionary data that expresses the content to be searched for by means of grammatical expressions. Similar to the first embodiment, the sentence expression is described by using at least one of a single term or a plurality of terms indicating a co-occurrence relationship for each phrase, sentence, and paragraph as a basic element. is there. In addition, word order and negation can be specified in the sentence meaning, and controlled words and sentence meaning macros can be used.

  The estimation dictionary data P created as described above is stored in the memory in the estimation dictionary creation device 1. In step S40, the document term table data of the document 2 group to be searched is generated by the same method as in the first or second embodiment.

  Next, in step S41, the estimation apparatus 3 reads the estimation dictionary data 3A and the literature term table data R1, R2, R3,. In step S42, the grammatical expressions described in each group of the estimation dictionary data P are compared with each document term table data. In this comparison process, when a set of terms that match the basic elements described in the grammatical expression is found from the literature term table data, a score is calculated for each row according to the number of matches.

  As in the first embodiment, this score calculation method is first calculated as “a set of terms + a set of basic elements + a set of semantic expressions”, and the total score of all lines is calculated for each document and each paragraph. Calculate each. Then, a method of weighting and adding the score maxTd (i) of the paragraph having the highest score among all the paragraphs and the total value Tt of the scores of all the paragraphs at a ratio of 1: 4 is used. The score calculated for each second document is associated with the number of the second document and stored.

  Then, the stored score is read, and this score is compared with a preset first threshold value. In step S44, it is determined whether or not the total score value is equal to or greater than the first threshold value. That is, it is determined whether the technical content similarity between the second document and the search condition is equal to or greater than a threshold value.

  As a result of the determination, if a second document having a score representing similarity equal to or greater than the first threshold is found, the second document is output to the terminal 4 as a similar document in step S45.

  On the other hand, if the second document whose score representing the similarity is equal to or greater than the threshold is not found, the fact is displayed and the search is terminated.

  As described in detail above, in the third embodiment, when searching for a document that matches or is close to the intended content from the second group of similar documents, it is described by a grammatical expression describing the search condition, and this is stored in the estimation dictionary. Remember. Then, the semantic expression stored in the estimation dictionary is compared with the document terminology table data generated from the second document to be searched in terms of terms, and the degree of coincidence is calculated as a score. When the value is equal to or greater than the threshold value, the second document is selected as a similar document.

  Therefore, according to the third embodiment, the grammatical expression describes a basic element that is a set of a plurality of terms indicating a co-occurrence relationship as a basic element. Can be reflected in the score. In addition, since the similarity between the estimation dictionary data and the document term table data is represented by a score, it is possible to determine the degree of similarity without being limited to determining whether the document is a similar document.

  In addition, since the word meaning is defined for each phrase, sentence, and paragraph, when searching for the corresponding term from the literature term table data R1, R2, R3,. Done within range. For this reason, a set of terms is not searched across phrases, sentences, and paragraphs, and sentence meaning can be estimated with high accuracy.

Example 4
In the fourth embodiment of the present invention, an estimation dictionary creating apparatus is provided with an automatic creation function for estimating dictionary data. FIG. 7 is a block diagram showing a functional configuration of an estimation dictionary creating apparatus used in the document analysis apparatus according to Embodiment 4 of the present invention.

  The estimation dictionary creating apparatus 100 is composed of a server computer or a personal computer, and includes a control unit 110, a storage unit 120, and an input / output interface unit 130.

  The input / output interface unit 130 is connected to the console terminal 200, the document database 300, and the estimation device 3 via a local network such as a USB (Universal Serial Bus) cable, a LAN (Local Area Network), or a public network. The console terminal 200 is used by an operator to input parameters to the estimation dictionary creation device 100. The console terminal 200 can also be used to display the document analysis result obtained by the estimation device 3. The document database 300 includes a database that stores document data of patent documents, for example.

  The storage unit 120 is a non-volatile memory that can be written and read at any time, such as HDD (Hard Disc Drive) and SSD (Solid State Drive), and a volatile that can be written and read at any time, such as RAM (Random Access Memory). A memory is used, and a teacher document storage unit 121, a basic element candidate storage unit 122, and a basic element storage unit 123 are provided as storage areas necessary for carrying out the fourth embodiment.

  The teacher document storage unit 121 is used to store the document data of the teacher document that is the basis for creating the estimation dictionary data. The basic element candidate storage unit 122 is used for storing a candidate group of basic elements generated in the process of creating the estimation dictionary data. The basic element storage unit 123 is used to store basic element groups generated in the process of creating estimation dictionary data.

  The control unit 110 includes a processor and a working memory. As control functions necessary for carrying out the fourth embodiment, a parameter acquisition control unit 111, a teacher document extraction unit 112, a basic element candidate extraction unit 113, An effective basic element selection unit 114 and an estimation dictionary data creation unit 115 are provided. These control functions are realized by causing the processor to execute a program stored in a program memory (not shown).

  The parameter acquisition control unit 111 performs processing for taking in parameters input by the operator at the console terminal 200 via the input / output interface unit 130. The parameters include, for example, an F-term that represents a technical field for which an estimation dictionary is to be created, a seed term that represents a characteristic of the technical field, and various threshold values that are used in the process of creating the estimation dictionary data. It is.

  The teacher document extraction unit 112 extracts a corresponding teacher document from the document database 300 using the F-term and seed term input as the parameters as a key, and the extracted teacher document is set as a positive example teacher document set and a negative example teacher document set. And processing to be stored in the teacher document storage unit 121 separately. The positive example teacher document is a document that includes the above-mentioned terminology and is given the F term. On the other hand, the negative example teacher document is a document that includes the above-mentioned terminology and is given the F term.

  The basic element candidate extraction unit 113 extracts basic element candidates for each range of phrases, sentences, paragraphs, and documents from the positive and negative teacher document groups stored in the teacher document storage unit 121. The basic element candidate storage unit 122 stores the extracted basic element candidates. The basic element is a set of a single seed term and a term indicating a co-occurrence relationship with the seed term. Details of the basic element candidate extraction process will be described later.

  The effective basic element selection unit 114 is included in as many positive example teacher documents as possible from the basic element candidate group stored in the basic element candidate storage unit 122, and is included in negative example teacher documents as much as possible. A basic element candidate that does not exist is selected, and the selected basic element candidate is stored in the basic element storage unit 123 as a basic element for creating an estimation dictionary. Details of the selection process of the basic element for creating the estimation dictionary will be described later.

  The estimation dictionary data creation unit 115 reads the basic element group for creating the estimation dictionary from the basic element storage unit 123, describes the combination of the basic element groups as a grammatical expression (Boolean algebra), and estimates the dictionary data. Create And the process which outputs this dictionary data for estimation to the estimation apparatus 3 is performed.

(Operation)
Next, an automatic creation processing operation of the estimation dictionary by the estimation dictionary creation device 100 configured as described above will be described. FIG. 8 is a flowchart showing the overall processing procedure and processing contents.

(1) Parameter Acquisition The estimation dictionary creating apparatus 100 waits for a parameter input operation in step S100 under the control of the parameter acquisition control unit 111. In this state, when an operator inputs a parameter on the console terminal 200, the parameter acquisition control unit 111 captures the parameter via the input / output interface unit 130 in step S110, and notifies the teacher document extraction unit 112 of the captured parameter.

  The parameter is composed of a target F term, a seed term group, and various threshold values. FIG. 12 shows an example. As shown in FIG. 12A, the target F term is the name of the F term of the technology for which the dictionary for estimation is created. For example, the F term and FI code defined by the Japanese Patent Office correspond to this. To do.

  The seed term is represented by a term that is a single word or a combination thereof that is characteristic as a technique for which the estimation dictionary is created. For example, in the F term “BB19”, as shown in FIG. 12B, “pedometer”, “distance meter”, {walk, sensor},.

  As the threshold value, as shown in FIG. 12C, IDF as a co-occurrence term threshold value for each of “phrase”, “sentence”, “paragraph”, and “document”, which are ranges when extracting basic elements. And the number of documents, information specifying whether or not to use co-occurrence terms only as technical terms, accuracy thresholds, accuracy ticks, recall thresholds, document thresholds, and number of generated terms The upper limit value, the upper limit value of the processing time and the number of basic elements generated are specified. FIG. 12C illustrates a case where the display of each threshold corresponding to “phrase” is omitted.

  The IDF here is an abbreviation for Inverse Document Frequency, which is the reciprocal of the number of documents hit by a co-occurrence term. This is used as a weighting coefficient in order to reduce the influence of co-occurrence terms that appear in most documents. For example, when the IDF of the basic element of the paragraph is 1.2, the co-occurrence term hits only about 83% or less of documents in the literature group of the population. Assuming that there are 1000 documents in the population, the basic element of the paragraph hits 833 documents or less.

  The accuracy represents the ratio of the number of positive example teacher documents including basic elements to the total number of positive example teacher documents including basic elements and the number of negative example teacher documents. The recall rate represents the ratio of the number of positive example teacher documents including basic elements to the total number of extracted positive example teacher documents. The accuracy threshold is set to the lowest acceptable accuracy. Also, the recall threshold is set to the lowest acceptable recall rate.

Since the ratio between the number of positive example teacher documents and the number of negative example teacher documents differs for each F-term, the number of positive example teacher documents and the number of negative example teacher documents are normalized by the following equations.
Number of normalized documents = number of actual documents * normalization coefficient
Normalization coefficient of positive example teacher document = 100 / number of positive example teacher documents used for learning for the F term Negative item normalization coefficient = 100 / number of negative example teacher documents used for learning for the F term It is.

(2) Extraction of Teacher Literature When the above parameter specification is completed, the estimation dictionary creating apparatus 100 executes the teacher literature extraction process as follows under the control of the teacher literature extraction unit 112.
That is, first, in step S120, the document database 200 is searched using the seed terms designated as the parameters as a key, and all documents including the seed terms are extracted.

  In step S130, the document including the extracted seed term is divided into a document set corresponding to the F term and a document set not corresponding to the F term specified as the parameter. Then, the document set corresponding to the F term is set as a positive example teacher document set, and the document set not corresponding to the F term is stored in the teacher document storage unit 121 as a negative example teacher document set.

  At this time, the teacher document extraction unit 112 creates seed data for each F term, and lists data of each of the positive and negative example teacher documents, and also stores them in the teacher document storage unit 121. FIG. 13 shows an example of the list data of the positive example teacher document set and the negative example teacher document set.

  By outputting the list data to the console terminal 200 and displaying it, the operator can check the extraction result of the teacher document in association with the seed term.

(3) Extraction of Basic Element Candidates The estimation dictionary creating apparatus 100 then selects basic element candidates from each of the positive example and negative example teacher document sets in step S140 under the control of the basic element candidate extraction unit 113. The extraction process is executed as follows. FIG. 9 is a flowchart showing the processing procedure and processing contents.

(3-1) Extraction of Basic Element Candidate Focusing on Phrase First, in step S141, the basic element candidate extraction unit 113 sets “phrase” as a range for extracting a basic element candidate. Then, every time one document included in each teacher document set of positive examples and negative examples is read from the teacher document storage unit 121, a basic element candidate is extracted for each “phrase” and the extracted basics The element candidates are stored in the basic element candidate storage unit 122.

For example, first, basic element candidates (hereinafter referred to as two basic element candidates) consisting of combinations of seed terms and co-occurrence words are extracted. The following is an example.
-Basic element candidates in Section 2;
<Seed term, co-occurrence word 11>, <seed term, co-occurrence word 12>, ..., <seed term, co-occurrence word 1N>
Next, basic element candidates (hereinafter referred to as basic element candidates in 3 terms) composed of combinations of basic element candidates in 2 terms and co-occurrence words are extracted. An example is shown below.
・ Basic element candidates in Section 3;
<<< species term, co-occurrence word 11>, co-occurrence word 21>, <<< species term, co-occurrence word 12>, co-occurrence word 22>, ..., <<< species term, co-occurrence word 1N>, co-occurrence word 2N>
Similarly, for each “phrase”, the process of extracting the candidates for the basic elements of the fourth, fifth,.

Further, during the basic element candidate extraction process for each “phrase”, the basic element candidate extraction unit 113 monitors whether or not the extraction condition is satisfied in step S142. As an extraction condition, for example,
(1) The accuracy and recall of the extracted basic element candidates are less than the minimum precision and recall that were set in advance as parameters
(2) It is defined that even if a co-occurrence word is further added to the extracted basic element candidates, that is, even if the number of terms is further increased, the accuracy of the basic element candidates does not improve beyond a certain value. Note that the accuracy and recall are calculated based on the numbers of positive and negative example documents normalized as described above.
When either one of the above conditions is satisfied, the basic element candidate extraction unit 113 ends the basic element candidate extraction process for each “phrase” at that time.

(3-2) Extraction of Basic Element Candidate Focusing on Sentence In step S143, the basic element candidate extracting unit 113 next sets “sentence” including the above phrase as a range for extracting the basic element candidate. Each time a document included in each of the positive and negative teacher document sets is read from the teacher document storage unit 121, basic element candidates are extracted for each “sentence” and the extracted basics The element candidates are stored in the basic element candidate storage unit 122.

In the extraction process of basic element candidates for each sentence, the number of combination terms of basic elements or basic elements and co-occurrence words is increased by one as in the case of “phrase”, such as two terms, three terms,. Execute repeatedly. The following is an example.
-Basic element candidates in Section 2;
{Seed term, co-occurrence word 11}, {Seed term, co-occurrence word 12}, ..., {Seed term, co-occurrence word 1N}
・ Basic element candidates in Section 3;
{Seed term, <seed term, co-occurrence word 11>, co-occurrence word 21}, {seed term, <seed term, co-occurrence word 12>, co-occurrence word 22}, ..., {seed term, <seed term, Co-occurrence word 1N>, co-occurrence word 2N}.

  Even during the extraction process of basic element candidates for each “sentence”, the basic element candidate extraction unit 113 monitors whether or not the extraction condition is satisfied in step S144. The extraction conditions shown in (1) and (2) above are used. If the extraction condition is satisfied, the basic element candidate extraction unit 113 ends the extraction process of the basic element candidate for each “text” at that time.

(3-3) Extraction of Basic Element Candidate Focusing on Paragraph In step S145, the basic element candidate extraction unit 113 sets a “paragraph” including the above phrase and sentence as a range for extracting the basic element candidate. . Then, every time one document included in each teacher document set of positive examples and negative examples is read from the teacher document storage unit 121, a basic element candidate is extracted for each “paragraph” and extracted basic The element candidates are stored in the basic element candidate storage unit 122.

The basic element candidate extraction process for each sentence is the same as in the case of “phrase” and “sentence” above, the number of combined terms of basic elements and basic elements and co-occurrence words such as two terms, three terms,. Repeatedly incrementing by one. The following is an example.
-Basic element candidates in Section 2;
[{Seed term, co-occurrence word 11}, co-occurrence word d11], [{seed term, co-occurrence word 12}, co-occurrence word d12], ..., {{seed term, co-occurrence word 1N, co-occurrence word d1N ]}
・ Basic element candidates in Section 3;
[{Species term, co-occurrence word 11}, <seed term, co-occurrence word 11>, co-occurrence word d21], [{seed term, co-occurrence word 12}, <seed term, co-occurrence word 12>, co-occurrence Word d22], ..., [{species term, co-occurrence word 1N}, <seed term, co-occurrence word 1N>, co-occurrence word d2N].

  Even during the extraction process of basic element candidates for each “paragraph”, the basic element candidate extraction unit 113 monitors whether or not the extraction condition is satisfied in step S146. The extraction conditions shown in (1) and (2) above are used. If the extraction condition is satisfied, the basic element candidate extraction unit 113 ends the extraction process of the basic element candidate for each “paragraph” at that time.

(3-4) Extraction of Basic Element Candidate Focusing on Entire Document The basic element candidate extracting unit 113 next starts with a basic element candidate extracted for each kind term or each “sentence” in step S147 as a different paragraph. Extract co-occurrence words from Then, a basic element candidate is created by combining the basic element candidate extracted for each seed term or “sentence” and the co-occurrence word extracted from the different paragraphs in a product format. An example is shown below.
Species * Co-occurrence word N1
{Seed term, co-occurrence word 11} * Co-occurrence word N1.

  Even during the process of creating the “product format” basic element candidate, the basic element candidate extraction unit 113 monitors whether or not the extraction condition is satisfied in step S148. The extraction conditions shown in (1) and (2) above are used. If the extraction condition is satisfied, the basic element candidate extraction unit 113 ends the process of creating the “product format” basic element candidate at that time.

  FIG. 15 shows an example of basic element candidates created by the extraction process described above. As shown in the figure, each extracted basic element candidate is stored in the basic element candidate storage unit 122 in a state associated with the basic element ID, the basic element type, and the number of terms. Also, each basic element candidate is stored in association with accuracy, recall, F value, and the number of positive and negative teacher documents that appear. The F value is a value indicating the degree of perfect match between the accuracy and the recall (harmonic average), and is calculated by the formula of (2 × accuracy × reproducibility) / (accuracy + reproducibility). Note that FIG. 15 shows a state of sorting in descending order of accuracy and recall to aid understanding.

(4) Selection of effective basic elements for creating an estimation dictionary The basic element candidate group extracted by the basic element candidate extraction unit 113 is a threshold (allowable) for both accuracy and recall based on seed terms. Basic elements exceeding the lowest possible value are comprehensively extracted from the teacher literature. For this reason, the extracted basic element candidate group always includes basic element candidates that are effective in creating the dictionary data for estimation. At this time, it can be said that there are no basic element candidates other than the extracted basic element candidates remaining in the teacher document set.

  However, the extracted basic element candidate group includes a plurality of basic element candidates appearing in the same teacher document. When creating the estimation dictionary data, it appears in as many documents as possible (ideally all documents) included in the positive example teacher document set, and does not appear as much as possible in documents included in the negative example teacher document set. It is necessary to select the minimum number of basic element candidates (ideally zero).

  Therefore, the estimation dictionary creating device 100 creates estimation dictionary data from the basic element candidate group extracted by the basic element candidate extraction unit 113 under the control of the effective basic element selection unit 114 in step S150. The process of selecting a valid basic element is performed as follows. FIG. 10 is a flowchart showing the processing procedure and processing contents.

  That is, first, in step S151, the effective basic element selection unit 114 calculates the accuracy for each basic element candidate included in the basic element candidate group extracted by the basic element candidate extraction unit 113. Sort each basic element candidate above.

  Next, in step S152, the effective basic element selection unit 114 selects one basic element candidate having the highest accuracy and recall from the sorted basic element candidate group, and selects the selected basic element candidate as an estimation dictionary. Is stored in the basic element storage unit 123 as the first basic element that is most effective in creating the.

  Next, in step S153, the valid basic element selection unit 114 excludes each of the positive example and negative example teacher documents in which the selected basic element appears, and selects the selected basic element for the remaining teacher document set. The accuracy is recalculated for each remaining basic element candidate excluded. Then, the remaining basic element candidates are sorted again in descending order of the recalculated accuracy.

  In step S154, the valid basic element selection unit 114 determines whether a preset selection end condition is satisfied. For example, the end of selection is determined based on whether or not all the target positive example documents remain. Note that the selection end may be determined based on whether or not the basic element candidates remain.

  If it is determined in step S154 that there is still a target positive example document and an unselected basic element remains, the effective basic element selection unit 114 returns to step S152. Then, one of the basic element candidates with the highest accuracy is selected from the re-sorted basic element candidate group, and this selected basic element candidate is the second most effective in creating an estimation dictionary. The basic element is stored in the basic element storage unit 123.

  Thereafter, in the same manner, the effective basic element selection unit 114 repeatedly executes the processes in steps S152 to S154, thereby selecting the basic elements in the order of accuracy calculated again each time, and selecting the sequentially selected basic elements. The basic element is stored in the basic element storage unit 123 as the basic element having the third, fourth,.

  In step S154, when it is determined that all of the target positive example documents remain, or that the basic element candidates remain, the effective basic element selection unit 114 uses the basic data used to create the estimation dictionary data. The element selection process ends.

  Next, the estimation dictionary creating apparatus 100 executes a process of selecting a basic element effective in creating the estimation dictionary data from the set of basic elements selected as having high accuracy as follows. . FIG. 11 is a flowchart showing the processing procedure and processing contents.

  First, in step S155, the valid basic element selection unit 114 sorts the basic element candidates in the descending order of the recall rate from the set of basic elements selected as having high accuracy. FIG. 15 shows an example of basic element candidates after sorting.

  Next, in step S156, the valid basic element selection unit 114 selects one basic element candidate with the highest recall from the sorted basic element candidate group, and creates an estimation dictionary for the selected basic element candidate. In the basic element storage unit 123, the first basic element having the highest effectiveness is stored.

  Next, in step S157, the valid basic element selection unit 114 excludes each of the positive example and negative example teacher documents in which the selected basic element appears, and selects the selected basic element for the remaining teacher document set. Recalculate the recall for each remaining basic element candidate. Then, the remaining basic element candidates are re-sorted in descending order of the re-calculated recall rate.

  In step S158, the valid basic element selection unit 114 determines whether a preset selection end condition is satisfied. For example, the end of selection is determined based on whether or not all the target positive example documents remain. Note that the selection end may be determined based on whether or not the basic element candidates remain.

  If it is determined in step S158 that there is still a target positive example document and an unselected basic element remains, the effective basic element selection unit 114 returns to step S156. Then, one basic element candidate with the highest recall is selected from the re-sorted basic element candidate group, and the selected basic element candidate is the second most effective in creating the estimation dictionary. The basic element is stored in the basic element storage unit 123 as a high basic element.

  Thereafter, in the same manner, the effective basic element selection unit 114 repeatedly executes the processes in steps S156 to S158, thereby selecting the basic elements in the descending order of reproducibility calculated each time, and sequentially selecting the basic elements. Are stored in the basic element storage unit 123 as basic elements having the third, fourth,.

  In step S158, when it is determined that there is no remaining target document or no remaining basic element candidates, the effective basic element selection unit 114 uses the basic data used to create the estimation dictionary data. The element selection process ends.

  FIG. 16 shows an example of the basic element selected by the basic element selection process from the basic element candidate group shown in FIG. In the figure, “accuracy”, “reproducibility”, “F value”, “number of positive example documents”, and “number of negative example documents” are the basics before starting the selection process, that is, immediately after sorting in step S151. Indicates the value of the element candidate group. On the other hand, “recalculation accuracy”, “recalculation recall”, “recalculation F value”, “number of recalculated positive example documents”, and “number of recalculated negative example documents” indicate values after recalculation. .

  In the basic element selection process described above, basic element candidates are set with the “recalculation accuracy” as the first priority, the “recalculation recall” as the second priority, and the “accuracy” before the recalculation as the third priority. Make a selection. The minimum accuracy value set as the accuracy threshold is set to “0.950”, and the maximum number of basic element selections is set to “200”.

  FIG. 14 shows an outline of the effective basic element selection process described above. In the figure, 50 indicates a positive document set. In the figure, the second basic element and the third basic element selected each time a large number of positive example documents are searched using the first basic element selected for the first time and then the accuracy and recall are recalculated. The remaining documents in the positive example document set are sequentially searched by the element and the fourth basic element. By executing the selection process in this way, it is possible to reduce the duplication of a plurality of basic elements for the same positive example document, and it is possible to select only a necessary and sufficient number of basic elements. Become.

(5) Creation of Estimation Dictionary Data Next, in step S160, the estimation dictionary creation apparatus 100 is based on the basic elements stored in the basic element storage unit 123 under the control of the estimation dictionary data creation unit 115. Next, create a lexical expression represented by a Boolean algebra. Then, the created sentence meaning is transferred as estimation dictionary data from the input / output interface unit 130 to the estimation device 3 in step S170. In addition, since the configuration of the textual expression is the same as that described in the first embodiment, the description thereof is omitted here.

(effect)
As described in detail above, in the fourth embodiment, in the estimation dictionary creation apparatus 100, based on a code representing a field for which a dictionary is to be created and a seed term representing a feature of the field, a set of positive teacher document sets from the document database 300 And a negative example teacher document set, a basic element candidate group consisting of a combination of the above-mentioned terminology and a term including its co-occurrence word is extracted from each of the positive example and negative example teacher document sets, and the extracted basic element From the candidate group, based on accuracy and recall, the smallest number that appears in as many documents as possible in the positive example teacher document set and does not appear in the negative example teacher document set as much as possible. A basic element candidate is selected, a grammatical expression is created based on the selected basic element candidate, and this is used as dictionary data for estimation.

  Therefore, it is possible to automatically create estimation dictionary data, which eliminates the need for a person to manually create an estimation dictionary, greatly reduces human workload, and analyzes documents. Early operation start of the device can be realized.

  In addition, since the basic element candidates are extracted from the phrases, sentences, paragraphs, and the entire document of the teacher document, it is possible to extract the basic element candidates without omission. At that time, since the extraction of basic element candidates is limited to those that exceed the threshold values of accuracy and recall, the number of basic element candidates can be effectively limited. It is possible to reduce the processing load on the apparatus required to create the image and to shorten the processing time.

  In addition, when selecting an effective basic element from the basic element candidates, the accuracy and recall are recalculated each time one basic element is selected. Based on the recalculated accuracy and recall, The next basic element is selected. For this reason, it becomes possible to effectively select the basic elements necessary for creating the estimation dictionary data without causing duplication or deficiency.

[Other Examples]
In the fourth embodiment, basic element candidates are extracted for each “phrase”, “sentence”, “paragraph”, and “document”, but “phrase” may be omitted. By omitting the “phrase”, it is possible to greatly reduce and shorten the processing load and processing time of the apparatus required for the process of selecting basic element candidates. Also, by reducing the number of extracted basic element candidates, it is possible to reduce the load and processing time for selecting effective basic elements.

  The present invention is not limited to the above-described embodiments. For example, the functional configuration of the estimation dictionary creation device, the literature term table generation device, and the estimation device, the processing procedure, and the processing content of the invention are also summarized. Various modifications can be made without departing from the scope of the present invention.

  Further, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the above embodiments. For example, some components may be deleted from all the components shown in each embodiment. Furthermore, the constituent elements in different embodiments may be appropriately combined.

  DESCRIPTION OF SYMBOLS 1,100 ... Estimation dictionary creation apparatus, 1A ... Estimation dictionary editor, 2 ... Literature term table generation apparatus, 3 ... Estimation apparatus, 2A ... Meaning filter part, 3A ... Estimation engine, 3B ... Difference extraction tool, 110 ... Control unit, 120 ... storage unit, 130 ... input / output interface unit, 111 ... parameter acquisition control unit, 112 ... teacher document extraction unit, 113 ... basic element candidate extraction unit, 114 ... effective basic element selection unit, 115 ... estimation dictionary Data creation part 121 ... Teacher document storage part 122 ... Basic element candidate storage part 123 ... Basic element storage part 200 ... Console terminal 300 ... Document database

Claims (17)

An estimation dictionary storage unit that stores search conditions as estimation dictionary data described by a lexical expression including basic elements in each range divided as a unit of a predetermined range that is delimited as a semantic expression;
Document terms that are stored as document term data for each range obtained by decomposing the document data related to the document to be searched into word units and generating a term based on the term as a unit of a predetermined range that is delimited as a semantic expression A storage unit;
A predetermined range in which a sentence expression included in the estimation dictionary data stored in the estimation dictionary storage unit and a term included in the document term data stored in the document term storage unit are separated as semantic expressions A score calculation means for comparing each of the ranges separated by the unit and calculating a score representing the similarity,
A document analysis apparatus comprising: a determination unit that determines whether or not the document to be searched is a document that satisfies the search condition based on the calculated score.   2. The predetermined range defined as each of the estimation dictionary data, the document term data, and the score calculation means and delimited as a semantic expression of the document is any one of a phrase, a sentence, and a paragraph. The document analysis apparatus described in 1. The estimation dictionary data includes, in a sentence expression, a composite basic element obtained by hierarchically combining basic elements defined in each of the phrase, sentence, and paragraph,
The compound basic element includes a term and one or more phrase basic elements for a sentence basic element, and a term and one or more sentence basic elements for a paragraph basic element. Or contains the basic elements of the phrase,
The score calculation means is for comparing a sentence meaning included in the estimation dictionary data stored in the estimation dictionary storage unit and a term included in the document term data stored in the document term storage unit. And for each phrase, sentence, and paragraph of the document to be searched, a term group included in the range is compared with a sentence expression including the composite basic element, and the score is calculated based on the matching frequency. Item 3. The document analysis apparatus according to Item 2. The estimation dictionary data defines a controlled word that is a generic term for a plurality of terms having the same meaning, and describes the grammatical expression using the controlled word,
The score calculation means is for comparing a sentence meaning included in the estimation dictionary data stored in the estimation dictionary storage unit and a term included in the document term data stored in the document term storage unit. , Comparing a plurality of terms defined as synonyms by the controlled word with the literature term data, and if any one of the terms defined as synonyms is included in the literature term data The document analysis apparatus according to claim 1, wherein the document analysis apparatus determines that the condition is satisfied. The score calculation means includes
The search target range of the search target document is divided into a plurality of ranges in units of a first range having a group as a semantic expression, and each of the divided ranges includes the term group and the sentence included in the range. First calculating means for calculating a first score based on the coincidence frequency with the intention;
A second score based on the matching frequency between a term group included in the second range and the grammatical formula for a second range wider than the first range in the search target range of the search target document. A second calculating means for calculating
Third calculating means for calculating a score representing a degree of similarity between the estimation dictionary data and the document to be searched based on the calculated first score and the calculated second score; The document analysis device according to claim 1, comprising:   The third calculation means weights and adds the maximum value of the calculated first scores and the second score with a preset coefficient, thereby the estimation dictionary data and the search The document analysis apparatus according to claim 5, wherein a score representing a similarity with the target document is calculated. The estimation dictionary data consists of a classification code assignment condition described by a sentence expression,
7. The determination unit, when it is determined that the document to be searched is a document satisfying the condition for assigning the classification code, the classification code is assigned to the document to be searched. The literature analysis apparatus in any one of. The estimation dictionary data consists of a description of search conditions for similar documents in terms of grammatical expressions,
7. The determination unit according to claim 1, wherein the determination unit selects the search target document as a similar document when it is determined that the search target document is a document that satisfies the search condition of the similar document. Document analysis apparatus according to the above.   A document analysis program for causing a computer to function as each unit included in the document analysis apparatus according to claim 1. A literature analysis method performed by a literature analysis apparatus including a computer and a memory,
Storing the search condition in the memory as estimation dictionary data described by a lexical expression including a basic element in each range divided as a predetermined unit that is delimited as a semantic expression, the document analysis device;
The document terminology data for each range obtained by dividing the document data related to the document to be searched into word units by dividing the document data generated based on the document data into semantic units as a semantic range. And storing in the memory as
The document analysis device has a predetermined range in which a sentence meaning included in the dictionary data for estimation stored in the memory and a term included in the document term data stored in the memory are separated as semantic expressions. Comparing each range divided as a unit, calculating the score that represents the similarity,
A document analysis method comprising: a step of determining whether or not the document to be searched is a document satisfying the search condition based on the calculated score. Further comprising an estimation dictionary creation device for creating the estimation dictionary data;
The estimation dictionary creation device includes:
A code defined for each field for which the estimation dictionary data is to be created, and means for receiving input of seed terms representing characteristics of the field;
A document group including the seed term and corresponding to the code and a document group including the seed term and not corresponding to the code are extracted from the document database as a positive example teacher document group and a negative example teacher document group, respectively. A teacher document extraction means;
From each of the documents included in each of the extracted positive example and negative example teacher document groups, for each range divided in units of a predetermined range where the document is delimited as a semantic expression, the seed term and its Basic element candidate extraction means for extracting a combination of words including at least one combination with a co-occurrence word as a candidate for the basic element;
Out of a plurality of basic element candidates extracted by the basic element candidate extracting means, basic element candidates that are included in the positive example teacher document group and that are not included in the negative example teacher document group have a threshold value or more. A basic element selection means to select;
An estimation dictionary data creation unit that creates a semantic expression representing the estimation dictionary data based on the basic element candidates selected by the basic element selection unit, and stores the estimation dictionary data in the estimation dictionary storage unit; Item 2. The document analysis apparatus according to item 1. The basic element candidate extraction means includes:
Define at least sentences and paragraphs as a predetermined range that is delimited as a semantic expression of the documents, and for each sentence and paragraph from each of the documents included in each of the extracted positive and negative teacher document groups And a first candidate extracting means for extracting a combination of terms including at least one combination of the seed term and its co-occurrence word as a candidate for the basic element;
A co-occurrence word having a co-occurrence relationship with a basic element extracted from the sentence is extracted from another paragraph not including the sentence, and extracted from the basic element candidate and the other paragraph extracted from the sentence. The document analysis apparatus according to claim 11, further comprising: second candidate extraction means for adding a combination of product forms with co-occurrence words to the basic element candidates. The basic element candidate extraction means includes:
1st precision showing the ratio of the number of positive example teacher documents including the basic element to the sum of the number of positive example teacher documents including the basic element and the number of negative example teacher documents, and the basic element for the total number of positive example teacher documents A threshold set in advance for each of the first recall ratios representing the ratio of the number of positive example teacher documents is held, and the basic element candidates extracted by the first and second candidate extraction means are defined as the first candidate. Extraction limiting means for limiting to basic element candidates exceeding the accuracy threshold and the first recall threshold;
The document analysis apparatus according to claim 12, further comprising: The basic element selection means includes
A second ratio representing the ratio of the number of positive example teacher documents including the basic element candidates to the total number of positive example teacher documents including the basic element candidates extracted by the basic element candidate extraction unit ; A second reproduction ratio indicating accuracy and a ratio of the number of positive example teacher documents including the basic element candidates extracted by the basic element candidate extraction unit to the total number of positive example teacher documents extracted by the teacher document extraction unit; The document analysis apparatus according to claim 11, wherein a basic element for estimation is selected from a plurality of basic element candidates extracted by the basic element candidate extraction unit based on the basic element candidate extraction unit. The basic element selection means includes
From among the candidates of a plurality of basic elements extracted by the primitive candidate extraction means, first selecting a candidate of the second precision and the second recall the highest basic elements as the basic element for estimation Means,
After the selection of the basic element for estimation, the positive example teacher remaining by excluding the reference including the basic element for estimation from the positive example teacher document group and the negative example teacher document group extracted by the teacher document extraction unit. as target document group and negative cases teacher document group, wherein for each candidate base element except the basic element for estimating said selected among the candidate of the basic elements candidate plurality of basic elements extracted by the extraction means the A second means for recalculating the accuracy of 2 and the second recall;
Among the plurality of basic element candidates extracted by the basic element candidate extracting means, out of all the basic element candidates excluding the selected basic element for estimation, the second accuracy after the recalculation and the second A third means for additionally selecting a basic element candidate having the highest recall rate of 2 as a basic element for estimation;
The processing of the second means and the third means is carried out in such a way that the number of remaining positive example teacher documents becomes a predetermined number or less, or candidates for each basic element excluding the selected basic element for estimation are predetermined. The document analysis apparatus according to claim 14 , further comprising: a fourth unit that is repeatedly executed until the number reaches a number or less.   An estimation dictionary creation program for causing a computer to function as each means provided in the estimation dictionary creation device of the document analysis apparatus according to claim 11. The literature analysis device further comprises a step of creating the estimation dictionary data,
The process of creating the estimation dictionary data includes:
A code defined for each field for which the estimation dictionary data is to be created, and a process of accepting input of a seed term representing the characteristics of the field;
A document group including the seed term and corresponding to the code and a document group including the seed term and not corresponding to the code are extracted from the document database as a positive example teacher document group and a negative example teacher document group, respectively. Process,
From each of the documents included in each of the extracted positive example and negative example teacher document groups, for each range divided in units of a predetermined range where the document is delimited as a semantic expression, the seed term and its Extracting a combination of words including at least one combination with a co-occurrence word as a candidate for the basic element;
Selecting a candidate for a basic element that has a probability of being included in the positive example teacher document group and not included in the negative example teacher document group from a plurality of extracted basic element candidates;
The document analysis method according to claim 10, further comprising: creating a sentence expression representing the estimation dictionary data based on the selected basic element candidates and storing the sentence meaning expression in a memory.