JP2009251796A - Document data sorting device, its method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a document data sorting device capable of performing sorting with a raised reproduction ratio when document data is sorted into prescribed categories, and to provide its method and a program. <P>SOLUTION: The document data sorting device includes: a dictionary generation attribute extracting part 4 for extracting an FT from known-category patent publications data 3 which are previously sorted into the categories "applied" and "not-applied"; a probability value calculation method selecting part 6 for selecting an expression to calculate a probability value E, based on the ratio of the category "applied" to the category "not-applied"; an inter-attribute-category probability value calculating part 5 for calculating the probability value E by using the selected expression so as to store the value in a probability value dictionary 7; an attribute extracting part 2 for extracting the FT of the sorted patent publications data 1; and a category determining part 8 for determining the category with respect to the sorted patent publications data 1 through the use of the FT of the sorted patent publications data 1 and the probability value dictionary 7, and outputting the determination result 9. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a document data sorting apparatus and method and program for automatically sorting document data into predetermined categories, and more particularly to a document data sorting apparatus and method and program for automatically sorting patent publications.

  Conventionally, there is Patent Document 1 as a technique related to automatic classification of patent publications. This technique automatically sorts based on the display of the applicant and the inventor, and the processing is as described in FIG. That is, the inventor information extraction unit 32 extracts inventor information from the input classified patent publication data 31, and the input classified patent publication data 31 is extracted by the inventor information extraction unit 32. Based on the inventor information, the patent publication classification unit 33 classifies the information and registers it in the inventor-specific patent publication DB 34. According to this method, patent publications can be classified by inventor. However, the inventor does not depend on the inventor's content and invention category, and the same inventor may make a plurality of different inventions. Further, the classified patent gazette data 31 is classified based on the presence or absence of the inventor information, and is not classified according to the technical contents described in the classified patent gazette data 31. Therefore, when looking at the classified results, the accuracy and recall were insufficient to efficiently perform secondary search and secondary investigation based on the technical contents. The precision and the recall here are as described in FIG. That is, the accuracy refers to the degree (A / A + B) of the related patent gazette included in the automatically classified patent gazette, and the recall rate does not leak the related patent gazette included in all patent gazettes. Describes the degree (A / A + C) indicating whether or not the extraction was successful.

Conventional automatic classification methods have a problem of improving accuracy, and there are various classification methods. Among them are support vector machines, decision trees, and decision lists. In the decision list, a dictionary is referenced to create a list of probability values corresponding to related attributes, and the category that gives the highest probability value in the list is used as the classification result.
For example, Non-Patent Document 1 describes a method of using a decision list for a problem homonym problem that extracts words suitable for sentence meaning from a set of words having the same hiragana notation.

In addition, Non-Patent Document 2 describes a method in which a uniform distribution and a beta distribution are approximately introduced as a prior distribution by using a decision list and using a Bayesian statistical method for calculating a probability value.
In addition, Patent Document 2 describes a method of using a difference between probabilistic perplexity depending on attributes and probabilistic perplexity not depending on attributes using a decision list.
JP 2006-58977 A JP 2001-266060 A IPSJ Journal, Vol. 39, no. 12 “Identification of homonyms using decision lists weighting evidence from compound words” (1998) Natural language processing, Vol. 9, no. 3. “Rule reliability estimation method for decision lists using Bayesian statistics” (2001)

  However, in patent document 1, although it classifies automatically by the presence or absence of the item of an applicant, an inventor, etc., even if it is the same inventor, a several different invention may be made, and the technique described in the patent gazette Since it is not classified based on the contents, the technical contents described in the patent gazette are not easily reflected in the classified results, and both the accuracy and recall are effective for efficient secondary search and secondary investigation based on the technical contents. It is insufficient. Patent Document 2 and Non-Patent Document 1 to Non-Patent Document 2 contribute to improving accuracy but do not contribute to improving recall. In general, accuracy improvement and reproducibility improvement are in a trade-off relationship, and it is not possible to expect improvement in reproducibility with conventional techniques aiming at accuracy improvement. Also, simply reducing the accuracy does not improve the reproducibility.

These methods aimed at improving accuracy are not suitable for investigations where a small number of search omissions is desirable. In particular, it was unsuitable for the secondary search of patent gazettes and the secondary search where the contents were scrutinized once again after classification.
Therefore, the present invention has been made paying attention to the above-mentioned conventional unsolved problems, and it is possible to perform document data classification that enables classification with an increased recall when document data is classified into predetermined categories. It is an object of the present invention to provide an apparatus and a method and program thereof.

  In order to solve the above-mentioned problem, a document data classification device according to claim 1 of the present invention for classifying document data into a predetermined category includes attribute extraction means for extracting attributes of inputted classified document data, A dictionary creating attribute extracting means for extracting attributes from known document data classified into a predetermined category, a category of known document data classified into the predetermined category, and an attribute extracted by the dictionary creating attribute extracting means; Between the attribute categories / probability value calculating means for calculating the probability value representing the relationship between the attribute categories, and between the attribute categories / probability value calculating means based on the ratio between the categories of the known document data divided into the predetermined categories Probability value calculation method selection means for selecting a method, and the probability values calculated by the attribute category / probability value calculation means are combined with the category and the attribute. A category determination unit that determines a category for the input classified document data using the probability value dictionary to be stored and the attribute extracted by the attribute extraction unit and the probability value dictionary, and the determined determination And a determination result output means for outputting a result.

According to the first aspect of the present invention, it is possible to select a calculation method that is expected to improve the reproduction rate by selecting a calculation method according to a ratio between categories of known document data that is divided into predetermined categories in advance. Thus, automatic classification of document data with a high reproduction rate becomes possible.
The document data sorting apparatus according to claim 2 is the document data sorting apparatus according to claim 1, wherein the probability value calculation method selection means calculates the probability value using (Expression 1) or (Expression 2). It is characterized by selecting the method to do.

  According to another aspect of the present invention, there is provided an attribute extracting unit for extracting an attribute of input classified document data, and a dictionary creating attribute for extracting an attribute from known document data previously classified into a predetermined category. A probability value representing the relationship between the extraction means, the category of the known document data classified into the predetermined category and the attribute extracted by the dictionary creation attribute means, using (Equation 3) A probability value calculating means; a probability value calculating parameter determining means for determining a parameter to be used in the attribute category / probability value calculating means based on a ratio between categories of the known document data divided into the predetermined categories; and the attribute A probability value dictionary that stores the probability value calculated by the inter-category / probability value calculating means in combination with the category and the attribute, and extracted by the attribute extracting means. A category determination unit that determines a category for the input classified document data using the attribute and the probability value dictionary, and a determination result output unit that outputs the determined determination result. Features.

According to the third aspect of the present invention, the calculation formula that is expected to improve the reproduction rate is determined by determining the parameter used for the calculation based on the ratio between the categories of the known document data that is divided into predetermined categories in advance. It becomes possible.
A document data sorting apparatus according to claim 4 is the document data sorting apparatus according to any one of claims 1 to 3, wherein the inputted classified document data is data of a patent publication, and the predetermined data The category belongs to a predetermined technical field, and the attributes extracted by the attribute extraction means and the dictionary creation attribute extraction means are F-term, IPC (International Patent Classification), FI, and Patent Gazette And at least one of the words appearing in the data.
According to the fourth aspect of the present invention, by using document data as patent gazette data, automatic classification of patent gazettes that are expected to have few search omissions becomes possible.

  Further, the document data classification method according to claim 5 includes an attribute extraction step of extracting attributes of inputted classified document data, and a dictionary creation attribute for extracting attributes from known document data previously classified into predetermined categories. A probability value representing a relationship between the extraction step and the category of the known document data and the attribute extracted in the dictionary creating attribute extraction step based on a ratio between categories of the known document data classified into the predetermined category; A probability value calculation method selection step for selecting a calculation method to be calculated; an attribute inter-category / probability value calculation step for calculating the probability value using the selected calculation method; and the calculated probability value as the category and A probability value dictionary storing step for storing in the probability value dictionary in combination with the attribute; the attribute of the classified document data and the probability value dictionary; There are, characterized by comprising a determining category determining step a category with respect to the classification document data the input, and a determination result output step of outputting the determined determination result.

According to the invention of claim 5, it is possible to select a calculation method that is expected to improve the reproduction rate by selecting the calculation method according to the ratio of the categories of the known document data previously classified into predetermined categories. Become.
The document data classification method according to claim 6 is the document data classification method according to claim 5, wherein the probability value is calculated using (Equation 1) or (Equation 2) in the probability value calculation method selection step. It is characterized by selecting the method to do.

  The document data classification method according to claim 7 is an attribute extraction step of extracting attributes of inputted classified document data in the document data classification method for classifying document data into a predetermined category, which is executed by a computer. A dictionary creating attribute extracting step for extracting attributes from known document data previously classified into predetermined categories; and a category of the known document data based on a ratio between categories of known document data classified into the predetermined categories; A parameter determining step for determining a parameter in (Equation 3) for calculating a probability value representing the relationship with the attribute extracted in the dictionary creating attribute extracting step, and the known document data classified into the predetermined category A probability value representing the relationship between the category and the dictionary attribute is calculated using (Equation 3). Using a value calculation step, a probability value dictionary storage step of storing the calculated probability value in the probability value dictionary in combination with the category and the attribute, the attribute of the classified document data and the probability value dictionary, A category determining step for determining a category for the inputted classified document data, and a determination result outputting step for outputting the determined determination result are provided.

According to the seventh aspect of the present invention, the calculation formula that is expected to improve the reproduction rate is determined by determining the parameter used for the calculation based on the ratio of the categories of the known document data that are classified into predetermined categories in advance. Is possible.
The program according to the invention of claim 8 is stored in the computer by downloading from the server or copying from the recording medium, thereby realizing the method described in any one of claims 5 to 7 by the computer. It becomes possible.

  According to the present invention, it is possible to select a calculation method that is expected to improve the reproduction rate by selecting a calculation method according to the ratio between categories of known document data that is divided into predetermined categories in advance, thereby improving the reproduction rate. Automatic classification of document data becomes possible.

Hereinafter, as a specific example of the embodiment of the present invention, a case where patent gazette data including the contents of one patent gazette is input will be described. FIG. 1 is a diagram illustrating a functional configuration of the document data sorting apparatus according to the present embodiment. This document data sorting apparatus includes an attribute extraction unit 2, a dictionary creation attribute extraction unit 4, an attribute category / probability value calculation unit 5, a probability value calculation method selection unit 6, a probability value dictionary 7, and a category determination unit 8. Yes. The probability value dictionary 7 is a database provided in a storage device (not shown) provided in the document data classification device, and includes an attribute extraction unit 2, a dictionary creation attribute extraction unit 4, an attribute category / probability value calculation unit 5, and a probability value. The calculation method selection unit 6 and the category determination unit 8 are functions realized by executing a program stored in a storage device by a CPU (not shown) included in the document data sorting device.
The classified patent publication data 1 as input is input to the document data sorting apparatus, and the attribute extraction unit 2 extracts, for example, F-term (FT: File Forming Term) which is attribute data of the publication data.

The category known patent gazette data 3 is data of a patent gazette to which a predetermined category is assigned in advance, and a large number are prepared. In the present embodiment, the category known patent publication data 3 including the same technical field as the technical field included in the classified patent publication data 1 includes, for example, among the categories of ○ and × (belonging or not belonging) as belonging ○ is given, and the category known patent gazette data 3 including different technical fields is given x as a category. In general, the larger the number of category known patent publication data 3 is, the easier it is to obtain a correct classification result. For this category known patent gazette data 3 as well, for example, F-term (FT), which is attribute data, is extracted by the dictionary creating attribute extracting unit 4 as described above.
Further, the relationship between the attribute extracted by the dictionary creating attribute extraction unit 4 and the known category is expressed by a probability value by the attribute category / probability value calculation unit 5. The probability value E is expressed by (Expression 4) or (Expression 5).

  P (category | FT) represents a conditional probability value in which the category assigned to the patent gazette is the category when the FT is given, and is a value calculated from the category known patent gazette data 3. Specifically, when the category is ○ (belongs to), P (○ | FT) = the number of patent publications having the category ○ / the number of patent publications having the FT, Means the ratio of patent gazettes with a circle. For example, if the number of patent gazettes having the FT is 100, and among them, the number of patent gazettes with a category of 30 is 30, P (◯ | FT) = 0.3.

  P (category) represents a probability value that the category assigned to the patent gazette is the category, regardless of the value of FT, and is a value calculated from the category known patent gazette data 3. Specifically, when the category is ○ (belongs to), P (○) = number of patent publications with category ○ / number of all patent publications, and the ratio of patent publications with category ○ regardless of FT. means. For example, if the number of all patent gazettes whose categories are known is 1000, of which 30 is the number of patent gazettes with a category of ○, P (◯) = 0.03.

Here, the ¬ category represents a complementary set of the category in the category set. Specifically, ¬ means x.
That is, the probability value E in (Expression 4) is normalized by using the difference between the probability value that depends on FT and the probability value that does not depend on FT.
The probability value calculation method selection unit 6 uses (Expression 4) and (Expression 5) as expressions for calculating the probability value E based on the ratio of the category (◯, ×) given to the category known patent publication data 3. Choose to use either.

For example, if the number of patent publications with a category of 30 is 30 and the number of patent publications with a category of 970 is 970, the ratio of x is much higher than ○, so the probability value calculation method selection unit 6 Then, (Formula 4) is selected, and the inter-attribute category / probability value calculator 5 is instructed.
The probability value E calculated by the attribute category / probability value calculation unit 5 is registered in the probability value dictionary 7.
The FT extracted by the attribute extraction unit 2 from the classified patent publication data 1 is input to the category determination unit 8. The category determination unit 8 refers to the probability value dictionary 7, creates a list of probability values related to the FT, and selects the category (◯ or ×) having the highest probability value among them. The selected item is output as the determination result 9.

  FIG. 2 shows an example of the result of automatic classification of patent gazettes using the conventional method and the method of this embodiment. Specifically, the results shown in FIG. 2 indicate whether or not a patent publication belonging to a certain technical field is related to whether or not it is related to a predetermined technology of interest set by the person in charge of A to G. It is the automatic classification result of the patent gazette that is ◯ when classified by. In addition, since the persons in charge A to G perform evaluation from different viewpoints, each patent publication is given a category of XX from different viewpoints.

In the figure, DL indicates the method of Non-Patent Document 1. DL-beta shows the method of Non-Patent Document 2 and uses a beta distribution as a prior distribution.
From FIG. 2, it can be seen that when the method of the present invention is used, the recall rate is remarkably improved as compared with the case where other methods are used. This shows that the method of the present invention greatly improves the performance of automatic classification of patent publications in which it is important to eliminate search omissions as much as possible.

  Example 2 differs from Example 1 in the calculation formula for calculating the probability value E. The formula for calculating the probability value E uses (Formula 6). All other processing is equivalent.

  Here, α and β are parameters satisfying α + β = 1, and in the second embodiment, the ratio (r) of the category ○ in the category known patent publication is used as a variable (Formula 7).

  Here, E in (Equation 6) represents the ratio between the FT-dependent probability value and the FT-independent probability value when α = 1, and when β = 1, the FT-dependent category and the complement of the category Represents the ratio.

  The third embodiment differs from the first embodiment in the number of categories, and uses three categories of A, B, and C. All processes other than the determination process in the category determination unit 8 are the same. Whether it belongs to any category of A, B and C is processed in two stages. In the first stage, it is determined whether or not A (¬A = {B, C}), and in the second stage, it is determined whether or not B (¬B = C) with respect to ¬A.

  The fourth embodiment is different from the first embodiment in the number of categories, and uses three categories of A, B, and C. All processes other than the determination process in the category determination unit 8 are the same. Whether any of the categories A, B, and C belongs is processed using the probability value for each category and category. That is, using the probability values for the three categories and their complements, A, B, C, ¬A, ¬B, and ¬C, one level using the values calculated in (Equation 5) and (Equation 6). Judge with.

  As explained above, automatic classification of patent gazettes with high recall is possible by using the method of this embodiment, so it can be used for secondary searches and secondary investigations of patent gazettes that scrutinize the classification results. It becomes possible. The attributes of patent gazettes are not limited to F-term, but IPC (International Patent Classification), FI (IPC complete symbol (symbol up to subgroup) + 3-digit number and / or 1-digit alphabet, (Classification used for organizing search files in the Patent Office) or words appearing in the data of patent gazettes. Further, the document data is not limited to the data of the patent gazette, and may be any data including character information such as questionnaires, web texts, papers, books, and advertising papers.

It is a block diagram of the document data classification device in Example 1 of this invention. It is explanatory drawing of the result of having classified the patent gazette automatically using the conventional method and the method of Example 1, respectively. It is a figure for demonstrating the technique regarding the automatic classification | category of the patent gazette in the past. It is a figure for demonstrating the definition of a precision and a recall.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Category patent gazette data 2 Attribute extraction part 3 Category known patent gazette data 4 Dictionary creation attribute extraction part 5 Inter-attribute category / probability value calculation part 6 Probability value calculation method selection part 7 Probability value dictionary 8 Category determination part

Claims (8)

In a document data classification device that classifies document data into predetermined categories,
Attribute extraction means for extracting attributes of the input classified document data;
A dictionary creating attribute extracting means for extracting attributes from known document data that has been classified into predetermined categories in advance;
Between attribute categories / probability value calculating means for calculating a probability value representing a relationship between a category of known document data classified into the predetermined category and an attribute extracted by the attribute extracting means for creating a dictionary;
Probability value calculation method selection means for selecting a calculation method to be used in the attribute category / probability value calculation means based on a ratio between categories of known document data classified into the predetermined categories;
A probability value dictionary for storing the probability values calculated by the attribute category-to-category / probability value calculating means in combination with the categories and the attributes;
Using the attribute extracted by the attribute extraction means and the probability value dictionary, a category determination means for determining a category for the input classified document data;
And a determination result output means for outputting the determined determination result. 2. The document data sorting apparatus according to claim 1, wherein the probability value calculation method selection means selects a method for calculating the probability value using (Expression 1) or (Expression 2). Sorting device.

In a document data classification device that classifies document data into predetermined categories,
Attribute extraction means for extracting attributes of the input classified document data;
A dictionary creating attribute extracting means for extracting attributes from known document data that has been classified into predetermined categories in advance;
Attribute category-to-category / probability value calculating means for calculating a probability value representing the relationship between the category of the known document data classified into the predetermined category and the attribute extracted by the dictionary creating attribute means using (Equation 3) When,
A probability value calculation parameter determining means for determining a parameter to be used in the attribute category / probability value calculating means based on a ratio between categories of known document data classified into the predetermined category;
A probability value dictionary for storing the probability values calculated by the attribute category-to-category / probability value calculating means in combination with the categories and the attributes;
Using the attribute extracted by the attribute extraction means and the probability value dictionary, a category determination means for determining a category for the input classified document data;
And a determination result output means for outputting the determined determination result.

  4. The document data sorting apparatus according to claim 1, wherein the inputted document data to be classified is data of a patent publication, and the predetermined category is whether or not it belongs to a predetermined technical field. The attribute extracted by the attribute extraction means and the dictionary creation attribute extraction means is at least one of F-term, IPC (International Patent Classification), FI, and a word appearing in the data of the patent gazette. A document data sorting device characterized by being. In a document data classification method executed by a computer to classify document data into predetermined categories,
An attribute extraction step for extracting the attributes of the input classified document data;
An attribute extraction step for creating a dictionary that extracts attributes from known document data that has been previously classified into predetermined categories;
A calculation method for calculating a probability value representing a relationship between a category of the known document data and an attribute extracted in the attribute extracting step for creating a dictionary based on a ratio between categories of the known document data classified into the predetermined category. A probability value calculation method selection step of selecting
Between attribute categories and probability value calculating step for calculating the probability value using the selected calculation method;
A probability value dictionary storing step of storing the calculated probability value in a probability value dictionary in combination with the category and the attribute;
A category determination step of determining a category for the input classified document data using the attribute of the classified document data and the probability value dictionary;
And a determination result output step of outputting the determined determination result. 6. The document data classification method according to claim 5, wherein in the probability value calculation method selection step, a method for calculating the probability value is selected using (Expression 1) or (Expression 2). Classification method.

In a document data classification method executed by a computer to classify document data into predetermined categories,
An attribute extraction step for extracting the attributes of the input classified document data;
An attribute extraction step for creating a dictionary that extracts attributes from known document data that has been previously classified into predetermined categories;
Calculating a probability value representing a relationship between the category of the known document data and the attribute extracted in the attribute extracting step for creating a dictionary based on a ratio between the categories of the known document data classified into the predetermined category; A parameter determination step for determining parameters in (Equation 3);
An attribute category / probability value calculation step of calculating a probability value representing a relationship between a category of known document data classified into the predetermined category and the dictionary attribute, using (Equation 3);
A probability value dictionary storing step of storing the calculated probability value in the probability value dictionary in combination with the category and the attribute;
A category determination step of determining a category for the input classified document data using the attribute of the classified document data and the probability value dictionary;
And a determination result output step of outputting the determined determination result.

  The program for making a computer perform the method described in any one of Claims 5-7.

Images