JP2009098809A - Technical literature classification system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To extract a characteristic vector representing needs to automatically classify a technical literature similar to the characteristic vector by using pattern classification techniques such as automatic text classification and machine learning and to match the technical literature with a customer needs. <P>SOLUTION: A technical literature classification system comprises a document input means for learning used for inputting a needs text that is needs information to a VOC document including the needs information and a technical literature including a technical element for solving the needs information; a document characteristic information generation means for generating document characteristic information based on an appearance frequency in the VOC document of the input needs text; a classifier generation means for generating a characteristic pattern of a category of the VOC document; a document input means for classification used for inputting a technical text that is the technical element; a classification means for using an identification function to be executed for the input technical text; and a result output means for outputting the classified classification result. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

The present invention relates to a technical document classification system for matching technical documents with customer needs by utilizing customer needs information (VOC).

The research results of companies are accumulated in detail in technical literature such as patents, technical reports and academic papers.
In order to make it easy to search such technical literature efficiently, there has been a method of classifying and organizing at the keyword level related to the technical theme name and technology, but how to put the accumulated technology into business? Whether it can be quickly linked to the needs of companies remains a challenge.

The needs of companies can be grasped from VOC information. VOC is “Voice of t
"He Customer" is data that reflects customer views and needs (complaints, questionnaire results, comments, market research, etc.). It is recognized that the key to survival and growth is how many companies put out customer-centric principles in their corporate strategy, listen to customers' opinions and reflect them in corporate activities, and how to use them is important. At present, there are methods and tools that can automatically collect VOC information from customers (see, for example, Patent Document 1).

However, how to link the automatically collected data with the technology possessed by the development department is quite a problem for sales teams who do not know the technology in detail. As a conventional solution, there has been a proposal for a service to transfer the technical information released by the research department to other companies through an intermediary agency, and to transfer the technology of the research results based on the public contract and the implementation contract.
(For example, refer to Patent Document 2). However, as a drawback, there is a risk of leaking technology and needs outside the company through a third party. Furthermore, since procedures such as application, registration and contract are necessary, it takes time and money. This method is not suitable when matching your company's needs and technology. In order to quickly respond to the needs of companies, there is a demand for an automatic matching method and system that can find out the technology of the company and that does not require a lot of effort.
Japanese translation of PCT publication No. 2004-503013 JP 2004-54608 A

Market needs can be determined based on VOC information collected from corporate customer service departments. However, investigating what technologies can be used to solve a need is quite difficult for a sales force that is not tech savvy.

On the other hand, technologies possessed by companies are disclosed through technical literature such as patents, technical reports, and academic papers. For engineers with little contact with customers, it is difficult to grasp what needs the developed technology can be applied to.

In order to quickly respond to market needs, research results must be utilized quickly, and there is a need for a technology that automatically matches needs and technologies and connects them to engineers who understand the technology.

The present invention is for solving the above-mentioned problems, and by using a pattern classification technique such as automatic text classification and machine learning, a feature vector that can represent a need is extracted, and a technical literature similar to the feature vector is extracted. Automatic classification to match customer needs and technical literature. Furthermore, using human information such as employee DB, it is possible to link with know-how information on who knows the corresponding technology, and connect with people who have needs and people who can solve the needs.

In order to solve the above-mentioned problem, the first invention of the present invention provides a needs text as the needs information for a VOC document including the needs information and a technical document including a technical element for solving the needs information. Learning document input means for inputting, document feature information generating means for generating document feature information based on the appearance frequency of the need text input by the learning document input means in the VOC document, and category of the VOC document Classifier creation means for generating a feature pattern, classification document input means for inputting technical text as the technical element, and classification using an identification function executed for the technical text input by the classification document input means And a result output means for outputting the classification result classified by the classification means.

The second invention of the present invention further comprises data pre-processing means for pre-processing the data of the various input texts. The third invention of the present invention is the author of the technical texts to be input. Engineer information search means for searching for engineer information is provided.

The fourth invention of the present invention is characterized in that the data preprocessing means performs morphological analysis on the inputted various text data, complements abbreviations, and unifies synonyms. It is characterized by that.

The fifth aspect of the present invention is characterized in that the feature extraction means extracts features using principal component analysis.

According to the present invention, only a VOC text including customer needs and a technical text including technology are prepared, and the feature text is generated as learning data, and the feature pattern of each category (class of needs) is obtained. By generating and assigning the label of the corresponding needs to similar technical texts, the needs information and the technical information can be automatically matched, so that the needs can be quickly met.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. Each of the following devices can be implemented with either a hardware configuration or a combination configuration of hardware resources and software. As the software of the combined configuration, a program that is installed in advance on a computer of a corresponding device from a network or a storage medium and that realizes the function of the corresponding device is used.

[First Example]
The first embodiment will be described below with reference to the drawings. FIG. 1 shows the overall configuration of an embodiment of the present invention.

The technical document classification system 100 in this embodiment includes a learning document input unit 101, a classification document input unit 102, a data preprocessing unit 103, a feature generation unit 104, a feature extraction unit 105, a classifier creation unit 106, and a classification unit 107. And a result output unit 108.

Hereinafter, the operation of the technical document classification system 100 in the first embodiment will be described (see FIG. 3).
.

1. The learning document input unit 101 reads a needs text from the needs DB 10 and inputs it as a learning text. The learning text is, for example, VOC text such as customer complaints, questionnaire results, comments, and market research collected by the sales department (see FIG. 5). In the description of this embodiment, a questionnaire result is targeted as learning data.

The needs text is divided into several categories in advance (for example, “I want to create a patent map automatically”, “I want to improve search efficiency”, “I want to improve system security”, etc.). A certain number of texts are input as learning data for each category (FIG. 5).
reference).

2. The classification document input unit 102 reads the technical text from the technical document DB 11 and inputs it as the classification text (see FIG. 6). The classification text is a text of a technical document such as a patent, a technical report, or an academic paper. The entire text can be used as it is, but to reduce costs, only the part that represents the problem and technical points (in the case of a patent, the technical field, the problem to be solved by the invention, the effect of the invention, etc.) It may be used. In the description of this embodiment, a patent as technical data is targeted.

3. The data preprocessing unit 103 performs abbreviated sentence complement, synonym reference, and morpheme analysis on the input text using a natural language analysis technique, and divides the document at the morpheme level. A known method is used for these processes. Only the process of morphological analysis may be performed on the input text. In addition, abbreviation completion and synonym reference may be performed. In general, since the expression of the document is not unified, the classification effect may be deteriorated. For example,
In the example sentence shown in FIG. 7, there is a need for “I want to make it easy”, but since the subject is omitted, the feature cannot be extracted. Through the abbreviated sentence completion process, the abbreviated subject “Generate patent diagram” is found from the contents of the preceding and following sentences, and the sentence is complemented. Furthermore, fluctuations in expressions such as “patent diagram” and “patent map” have the same meaning, although they are written differently. In order to avoid this, a synonym dictionary is used to unify terms with the same meaning.

4). The feature creation unit 104 creates a morpheme list for all the texts delimited by the process (3.) of the data preprocessing unit 103. The appearance frequency of morphemes in the created morpheme list is obtained for each learning text and classification text, and a feature vector is generated (Fig.
8).

Instead of the appearance frequency of the morpheme as the component of the feature vector, use the result of the removal of the function word from the unnecessary word list, the use of the relative appearance frequency independent of the document length, the power transformation that improves the symmetry, etc. May be.

Even if the feature vector generated here is transmitted as it is to the classifier generation unit 106 and the classification unit 107, automatic classification can be performed. However, to improve classification accuracy and reduce costs,
The following feature extraction process (5.) can be used.

5). The feature extraction unit 105 automatically extracts features having a good classification effect for the feature vectors generated by the feature creation unit 104, and generates new feature vectors in the order of features having excellent separability.

In the description of this embodiment, principal component analysis is used as a feature extraction method, but otherwise, an LSI method (latent semantic indexing-LSI) that maps feature vectors to another space may be used.

The procedure for feature extraction by principal component analysis is shown below (see Figure 9).
Step 1: Calculate a mean vector Mhat and a sample covariance matrix Σhat from the set of feature vectors χ = {X1, X2Λ, XN} of the input document according to the following equations. Xl is a feature vector of the input document l.

[Equation 1]
Mhat = (1 / N) ΣX (X∈χ)
[Equation 2]
Σhat = (1 / N) Σ (X-Mhat) (X-Mhat) ^ T (X∈χ)
Here, N is the sample size (number of learning documents). T represents the transpose of the vector.

Step 2: Obtain the eigenvalue λi and eigenvector Φi that satisfy the following equation for the obtained sample covariance matrix.

[Equation 3]
ΣhatΦi = λiΦi (i = 1,2,, Λ, n)
Here, λ1 ≧ λ2 ≧ Λ ≧ Λn.

Step3: For the feature vector X, find the principal components for m (m ≦ n) eigenvalues from the largest.

[Equation 4]
zi = Φ {from i to T} X (i = 1,2, Λ, n)
Step 4: An m-dimensional feature vector having m principal components as elements is set as a new extracted feature vector.

The dimension number n of the feature vector before the feature extraction is the number of elements of the feature vector generated by the feature creation unit 104, and the dimension number m of the feature vector after the feature extraction is the number of principal components obtained by [Equation 4]. is there. In feature extraction by principal component analysis, the user can freely select the dimension number m of a new feature vector.

6). The classifier creating unit 106 uses the feature vector of the learning text generated by the feature extraction unit 105 to generate a pattern for each category.

In the description of this embodiment, the pattern of each category is the average vector [Equation 1] of the learning population of each category, but a covariance matrix or the like may be used [Equation 2].

7). The classification unit 107 calculates the distance (similarity) between the classification vector generated by the feature extraction unit 105 and the pattern vector of each category, and classifies the category into the category with the smallest distance (highest similarity). Then, it is registered in the classification result DB 12 and transmitted to the result output unit 108.

In the description of this embodiment, the simplest discriminant function-ycklid distance is used, but discriminant functions such as inner product, Dice coefficient, cosine coefficient, and linear discriminant function may be used. The (square) Euclidean distance is a scale for evaluating the distance between two vectors by the sum of squares of the difference between the feature amounts of the feature vectors. For the feature vector X of the input document, the Euclidean distance from the category l is obtained as follows.

[Equation 5]
gl (X) = (X-Mlhat) ^ T (X-Mlhat) = || X-Mlhat || ^ 2
Here, Mlhat is the average vector of the lth category.

As shown in the example of FIG. 11, the input text X is classified into category 2 having the smallest Euclidean distance from the center of each category learning group.

In the explanation of this embodiment, it is intended to classify input technical texts into predetermined needs categories. However, even if the needs text is not classified, it is input together with the technical text and the text is grouped by the clustering method. You can also use matching method. In the case of clustering classification, the classifier creating unit 106 is not necessary. A known method can be used for clustering.

8). The result display unit 108 outputs the results classified by the classification unit 107 in a form that is easy for the user to understand (for example, the format shown in FIG. 12).

As described above, according to the present invention, by classifying into the needs category in the technical literature,
Need information and technical information can be linked automatically and without any hassle.

[Second Example]
A second embodiment of the present invention will be described below. FIG. 2 shows the overall configuration of the second embodiment. The technical literature classification system in the present embodiment is composed of elements obtained by adding a know-how information search unit 209 to the constituent elements of the first embodiment.

Hereinafter, the operation of the technical literature classification system in the second embodiment will be described (see FIG. 4).

The process from the reading of learning data by the learning document input unit 201 until the result display unit 208 outputs the classification result to the user is the same as in the first embodiment.

The result display unit 208 displays the inventor of the patent together with the patent number (see FIG. 12). When the inventor's name is clicked, the inventor's name is transmitted to the know-how information search unit 209 as a search keyword. In the description of this embodiment, the inventor of the patent document is targeted.
In the case of other technical literature, the author, creator, etc. are targeted.

The know-how information search unit 209 uses the inventor's name transmitted from the result display unit 208,
Search the know-how database. The retrieved result is transmitted to the result display unit 208, and the result display unit 20
8 to the user (see FIG. 13). In the explanation of this embodiment, the know-how DB
Targets the employee information DB, but other personal information DBs may be used.

In this way, in addition to the effect of the first embodiment, the person who can solve the needs and the person who can solve the needs can be connected, and the cooperation between the sales team and the research team of the company can be supported more efficiently.

As described above, according to each embodiment of the present invention, only a VOC text including customer needs and a technical text including technology are prepared. The feature information of the category) is generated, and the label of the corresponding needs is assigned to the similar technical text, so that the needs information and the technical information can be automatically matched, so that the needs can be quickly met.

Furthermore, a high-precision classification effect can be obtained by processing such as text preprocessing, feature normalization, and feature extraction with high separability.

Since there is almost no human involvement in all processing, full automation can be realized and it does not take time.
Moreover, since it does not depend on individual judgment, the uniformity of the classification result can be maintained.

In addition to the above, the needs text has the effect of supporting communication between the sales unit and the research unit by cooperating with the authors of the technical literature and the know-how information of the inventor. As a program that can be executed, the program is stored in a storage medium such as a magnetic disk (floppy (registered trademark) disk, hard disk, etc.), an optical disk (CD-ROM, DVD, etc.), a magneto-optical disk (MO), or a semiconductor memory. It can also be distributed.

In addition, as long as the storage medium can store a program and can be read by a computer, the storage format may be any form.

In addition, an OS (operating system) running on a computer based on an instruction of a program installed in the computer from a storage medium, MW (middleware) such as database management software, network software, and the like realize the above-described embodiment. A part of each process may be executed.

Furthermore, the storage medium in the present invention is not limited to a medium independent of a computer, but also includes a storage medium in which a program transmitted via a LAN or the Internet is downloaded and stored or temporarily stored.

Further, the number of storage media is not limited to one, and the case where the processing in the above embodiment is executed from a plurality of media is also included in the storage media in the present invention, and the media configuration may be any configuration.

The computer according to the present invention executes each process in the above-described embodiment based on a program stored in a storage medium, and is a single device such as a personal computer or a system in which a plurality of devices are connected to a network. Any configuration may be used.

In addition, the computer in the present invention is not limited to a personal computer, but includes an arithmetic processing device, a microcomputer, and the like included in an information processing device, and is a generic term for devices and devices that can realize the functions of the present invention by a program. .

Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. Moreover, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

The figure which shows the whole system structure in a 1st Example. The figure which shows the whole system structure in a 2nd Example. The figure which shows the flow of the process in a 1st Example. The figure which shows the flow of a process in 2nd Example. The figure which shows learning text input. The figure which shows the text input for a classification | category. The figure which shows operation | movement of a data pre-processing apparatus. The figure which shows operation | movement of the feature preparation apparatus. The figure which shows operation | movement of a feature extraction apparatus. The figure which shows operation | movement of a classifier production | generation apparatus. The figure which shows operation | movement of a classification device. The figure which shows a classification result. The figure which shows a know-how search result.

Explanation of symbols

100, 200 ... technical document classification system, 101, 201 ... learning document input unit, 102
202, classification document input unit, 103, 203 ... data preprocessing unit, 104, 204 ... feature creation unit, 105, 205 ... feature extraction unit, 106, 206 ... classifier creation unit, 107, 207
... Classification part, 108, 208 ... Result output part, 209 ... Know-how information search part, 10 ... Needs D
B, 11, 21 ... technical document DB, 12, 23 ... classification result DB, 24 ... know-how DB.

Claims (5)

For technical documents that include technical elements for solving VOC documents containing needs information and the needs information,
A learning document input means for inputting a needs text as the needs information;
Document feature information generating means for generating document feature information based on the appearance frequency of the need text input by the learning document input means in the VOC document;
Classifier creating means for generating a feature pattern of a category of the VOC document;
Document input means for classification for inputting technical text as the technical element;
Classification means using an identification function executed on the technical text input by the classification document input means;
A result output means for outputting a classification result classified by the classification means;
The technical literature classification system characterized by comprising. 2. The technical document classification system according to claim 1, further comprising data preprocessing means for preprocessing the input various text data. The technical document classification system according to claim 1 or 2, further comprising engineer information search means for searching for engineer information as a writer of the input technical text. 4. The technical document classification system according to claim 2, wherein the data preprocessing means complements abbreviations for the various input text data and unifies synonyms and performs morphological analysis. . 5. The technical document classification system according to claim 4, wherein the feature extraction means extracts features using principal component analysis.

Images