JP2003281186A - Example-based search method and search system for similarity determination - Google Patents

Abstract

The present invention relates to an example-based document search method and a search system for automatically searching for a document, and quantitatively calculates the degree of similarity between a document in an already constructed database and an example document. Provided to a user is an example-based search method and a search system for determining similarity that are useful for searching for related technologies that are the same as or similar to an example document in a short time. The present invention provides an index including a step of inputting a related art document, a step of expressing the related art document as a word vector by analyzing a specific structure of the document, and a step of storing the expressed word vector. And inputting an example document, expressing the example document as a word vector by analyzing the specific structure of the document, and similarity between the word vector for the related technical document stored in the indexing process and the vector for the example document. And a search system therefor.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an example-based search method and search system for automatically searching documents,
More specifically, the present invention relates to an example-based search method and a search system for determining similarity using the specific structure of a document to search for the same or similar related technologies.

[0002]

2. Description of the Related Art The development of new technology is based on a completely new idea, but to date, it is more often produced from improved technology utilizing the technology of the same or related technical fields, and its contents are more complex and diverse. And its value is increasing. Therefore, in the industrial field, it is often necessary to judge the identity and / or similarity between technologies already developed and announced.

For example, when a new technology is developed in a company or a research institute, or while the new technology is being developed or completed, it is searched whether or not there is a related technology which is the same as or similar to the technology, and the technology is searched. Work is being performed to determine whether or not there is an identity and similarity with the related technology retrieved.

As described above, when a conventional keyword search system is used to determine the identity and similarity between technologies, the user must understand the example technology (for example, new technology). After searching the technology related to the example technology, the example technology and the related technology found are identified as
Or, you have to directly judge the similarity (difference).

As described above, conventional keywords
When using the search system, the user must understand new knowledge and surely recognize the document content, and the sameness and / or similarity (difference) between the example technology and the related technology searched It takes a lot of time to judge such a problem, and there is a possibility that the related art may be overlooked because the search is performed by using some keywords, and the accuracy of the search is poor.

Thus, the conventional keywords as described above
(keyword) Techniques related to example-based retrieval have been proposed to solve the problems of the retrieval system.

For example, some functions related to software provided as a part of the search system by a regular search system manufacturer such as Verity's Search 97 and Oracle's Intermedia are realized. Furthermore, theoretically, Gerald Salton (Non-Patent Document 1), Ricardo and Berthier
Documents such as (Non-Patent Document 2) also describe basic methodologies for example-based retrieval.

In general, the example-based search is a query (quer) for finding a specific document in an information search system.
Instead of using some core keywords in combination as y), it refers to the method of directly searching the example document selected by the user as a query. That is, a keyword is extracted from a document and expressed by a word vector,
This is called an index, and such a word vector plays the same role as a combination of keywords.

However, in the regular search system and the methodologies presented in the literature, the handling of keywords and the handling of documents are regarded as the same, and only the appearance or non-appearance of specific words is emphasized in the indexing process. There was a problem that important information indicating the subject of the document was overlooked. In other words, since the feature of the document is not processed and the content of the important part and the content of the unimportant part in the document are not distinguished, there is a problem that the accuracy rate of the search decreases.

It is considered that such problems are caused by the fact that the document includes many fields due to the structural characteristics of the document. In consideration of such matters, some commercial systems provide a method for allowing a user to divide a document into several fields, and to search by using the simple similarity between the fields desired by the user. is doing. However, there is a problem that the search by comparing the document parts does not meet the precise processing request for the content of the entire document.

[0011]

[Non-Patent Document 1] Gerard Salton. (1989). Automatic t
ext processing: the transformation, analysis, and
retrieval of information by computer.Addison-Wes
eley, Reading, Massachusetts.

[Non-Patent Document 2] Ricardo Baeza-Yates & Berthier Rib
erio-Neto. (1999). Modern information retrieval. Add
ison-Weseley, Reading, Seoul

[0012]

SUMMARY OF THE INVENTION The present inventors have conducted research to solve the problems of the prior art and have proposed the present invention based on the results of the research. It is an object of the present invention to provide an example-based search method and a search system for similarity determination, which displays the same and / or similar related technologies together with their similarity so that the similarity can be determined more quickly and accurately.

[0013]

The present invention will be described below. The present invention provides an indexing process including inputting a related technical document, expressing the related technical document as a word vector by analyzing a specific structure of the document, and storing the expressed word vector; The steps include inputting, expressing the example document as a word vector by analyzing a specific structure of the document, and obtaining a similarity between the word vector for the related technical document stored in the indexing process and the word vector for the example document. The present invention relates to an example-based search method including a search process.

Further, the example-based search method for determining the preferred similarity according to the present invention includes an example-based index process and an example-based search process, wherein the example-based index process includes inputting a related technical document; In the related technical documents, a paragraph is divided according to the structural characteristics of the document, and a keyword is extracted for each divided paragraph; a weight value in each paragraph is calculated for each keyword extracted from each paragraph, and a keyword is obtained for each paragraph. Expressing the weights in a word vector; storing the keywords expressed in the word vector and the weights, and the example-based search process inputs an example document in which example techniques are described. ; A step of dividing paragraphs in the input example document according to the structural characteristics of the document and extracting a keyword for each divided paragraph Calculating a weight value in each paragraph for the keyword extracted from each paragraph and expressing the keyword and its weight value in each paragraph by a word vector; storing the word vector by paragraph in the expressed example document and the indexing process The degree of similarity between the corresponding paragraphs of the example document and the related technical document is obtained by using the word vector for each related technical document, and the similarity between the example document and the related technical document is obtained using the similarity between the paragraphs. And a step of arranging related technical documents and providing them to a user in descending order of the calculated similarity (claim 1).

Further, according to the present invention, a related technology input section for inputting a related technology document, a related technology document expression section for expressing the related technology document input in the related technology input section by a structure analysis of the document by a word vector, and An index unit including a related technical document storage unit that stores the word vector expressed in the related technical document expression unit; an example document input unit for inputting an example document in which example techniques are described, and an example document input unit for analyzing the structure of the document. In the example document expression unit that expresses the example document input in 1. with the word vector, the word vector for the related technical document stored in the related technical document storage unit and the word vector for the example document expressed in the example document expression unit are used. Align the related technical documents in the descending order of the similarity calculated by the similarity calculating unit and the similarity calculating unit for obtaining the similarity to the example document It relates example based retrieval system for similarity determination comprising a display unit for providing the user Te.

Further, the example-based search system for determining the preferred similarity according to the present invention includes an example-base index unit and an example-base search unit, and the example-base index unit inputs related technical documents. Document input unit; first keyword extraction unit that divides paragraphs in the related technical document input by the input unit according to the structural characteristics of the document and extracts keywords for each divided paragraph; each paragraph in the first keyword extraction unit A first word vector expression part for expressing a keyword and its weight value for each paragraph by a word vector, and calculating a weight value in each paragraph for the keyword extracted from; and a word vector expressed by the first word vector expression part The example base search unit includes a word vector storage unit that stores keywords and their weights, and the example base search unit is a description of example techniques. Examples document input unit for inputting the example document; the example by dividing the paragraph by structural characteristics of the document in the example document input by the document input unit, the second keyword extracting unit for extracting a keyword by segmented paragraph;
A second word vector expression unit that obtains a weight value in each paragraph for a keyword extracted from each paragraph in the second keyword extraction unit and expresses the keyword and its weight value by a word vector for each paragraph; the second word vector expression Using the paragraph-based word vector for the example document represented by the section and the paragraph-based word vector for the related technical document stored in the word vector storage unit, the similarity between corresponding paragraphs of the example document and the related technical document is obtained, A similarity calculation unit that obtains the similarity between the example document and the related technical document using the inter-paragraph similarity; and the related technical documents are arranged and provided to the user in descending order of the similarity calculated by the similarity calculation unit. And a display unit for displaying the information (claim 6).

The present invention will be described in detail below. The "example technology" used here means a technology for determining whether or not there is a similar technology, and the "related technology" means all the other technologies of the above example technology. Related art includes not only those publicly known prior to the example technology but also those publicly known thereafter.

As used herein, "similarity determination" means determining whether or not an example technique is the same and / or similar to a related technique. For example, when the example technology is an invention related to a patent, the “similarity judgment” means the completion point of the invention and / or
Alternatively, it may be determined whether or not the invention is the same as and / or similar to the invention (device, technology, etc.) described in the patent document filed before or after the filing date, or the publications distributed earlier or later. Mean, identical or similar
It is applied to the determination of patent requirements (newness, inventive step, prior application relationship) based on (inventive step), application judgment, patent infringement judgment, etc.

The term "example document" used herein is used to determine the identity and / or similarity between documents in which technologies (including inventions, inventions, etc.) are described and the similarities with other related technologies. Or, it means a document in which technology for which it is determined whether there is similarity is described, and "related technical document" is technology for which it is determined whether there is identity and / or similarity. It means a document that describes other related technologies of.

Typical examples of the above-mentioned example documents and related technical documents are general technical documents and technical materials, patent documents (patent specifications, etc.) in which an invention or the like is described according to the description requirements required by each national patent office, and An application document may be one in which the described content partially satisfies the above described requirements (invention declaration, proposal, etc.) or does not satisfy at all (invention declaration, proposal, etc.). Examples of the patent documents and the like include non-disclosed specifications, patents or utility model publications, patents or utility model publications, and patents or utility model registrations, in which an invention or invention is described. Further, the application document includes a document summarizing research subjects, a document summarizing research results, a document summarizing completed technical contents (employee invention declaration, proposal, etc.).

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT A preferred embodiment of the present invention will be described below with reference to the accompanying drawings. INDUSTRIAL APPLICABILITY The present invention quantitatively calculates the degree of similarity between a document in an already constructed database and an example document and provides it to the user, which helps to find a related technique that is the same as or similar to the example document in a short time. An example-based search method and a search system for determining similarity are provided.

FIG. 1 shows an overall configuration diagram of an example-based search system for similarity determination according to the present invention.
Referring to FIG. 1, the example-based search system 100 according to the present invention is similar to a general information search system in the index unit 11.
It is roughly divided into 0 and the search unit 120. The index section 11
Reference numeral 0 indicates a related technical document input section 11 into which a related technical document is input.
1, a related technical document expression unit 112 that expresses a related technical document by structural analysis of the document, and a related technical document storage unit 113 that stores the expressed document, and the search unit 120 inputs an example document to which an example document is input. Section 121, an example document expression section 12 for expressing an example document by analyzing a specific structure of the document
2 and the similarity calculation unit 123.

In the present invention, the "index" means a process of recording a related technical document in the example-based system of the present invention in a structure such that the document can be searched in advance in order to search the related technical document. '' Means a process of analyzing an example document (example technique) presented by the user for similarity determination related to the related technical document and searching for a technical similarity from the indexed related technical document. .

Generally, document retrieval and information retrieval are based on the theory of vector space model. Also in the present invention, the index and the search are constructed based on the vector space model.

In order to explain the present invention, first, an example-based retrieval system based on a vector space model will be described with reference to FIG. Referring to FIG. 2, in the general example-based search system 200, a document representation process, which is a common process for expressing the given example document and the related document given by both the index unit 210 and the search unit 220, is performed.

In the example-based retrieval system based on the vector space model, all documents are represented by vectors consisting of words. If the number of words appearing in the set of documents to be stored is n, the document D _i is an n-dimensional word vector (W
_{i, 1} , W _{i, 2} , ..., W _{i, n} ).

At this time, W _{i, j} means a weight value of the word T _j with respect to the document D _i . In general document retrieval, the weight value W _{i, j} of the word T _j with respect to the document D _i is the word frequency count (t
f: term frequency) and document inverse frequency (idf: inverse document)
frequency).

Word frequency number t of word Tj in document Di
f _{i, j} is the number of occurrences of the word T _j appears in document D _i, word T _j is a measure of whether the representative of how the contents of the document.

[0029] On the other hand, document frequency document reverse frequency of the word T _j is the ratio of the number of documents in which the word T _j appears in the document set
It means the reverse of (document frequency). The smaller the number of documents in which the word T _j appears, the more the word T _j becomes the word T _j.
It has a high ability to distinguish the document in which is shown from other documents. The document inverse frequency of the word T _j serves as a measure of document discrimination.

Various word weight calculation methods using the word frequency and the document inverse frequency have been studied. In the present invention, the word weight calculation method of the widely known INQUERY system is used. The weight value W _{i, j} of the word T _j with respect to the document D _i is obtained by the following equation (1).

[Equation 1] (Where tf_{i, j}: Document D_iFor the word T_jFrequency
Number, max_tf: Document D _iThe most frequently appearing in
Frequency of words, N: number of whole documents, n: word T_jAppears
Number of documents)

If the weighted value of each word appearing in the document is calculated by the equation (1), the document can be expressed by the word and the word weighted value.

Referring to FIG. 2, after the related technical document input to the related technical document input unit 211 forming the index unit 210 is expressed in the first document expression unit 212 by a word and a word weight value, the document expression is expressed. In order to record the data into the system for quick and easy retrieval, the document storage unit 213 undergoes an inverted index file storage step of storing it in an inverted indexing file storage structure. The reverse index file storage step has been traditionally used in information retrieval.

On the other hand, the search unit 2 of the example-based search system
In FIG. 20, according to FIG. 2, after the example document input to the example document input unit 221 by the equation (1) is expressed by the second document expression unit 222 by a vector composed of words and weights thereof, document-document After the document-document similarity calculation process of calculating the similarity by comparing the vector representation of the related document already stored in the document storage 213 in the similarity calculation unit 223, the similarity is greater than 0 in the display unit. The related documents are arranged in order of similarity and provided to the user.

An example document D _i given in the indexing process
And similarity of each related document D _x stored [sim (D _x ,
D _i )] is traditionally the cosine similarity (cosin
e similarity) and the following equation (3)
product similarity) formula.

[0035]

[Equation 2] (Where W _{x, j} : weight value of word T _j for document D _x , W _{i, j} : weight value of word T _j for document D _i )

[0036]

[Equation 3] (Where W _{x, j} : weight value of word T _j for document D _x , W _{i, j} : weight value of word T _j for document D _i )

As can be seen from FIGS. 1 and 2, the present invention improves the steps of index representation, the step of document expression in the retrieval step, and the step of obtaining the similarity in the retrieval step in the example-based retrieval method. That is, the improvement point of the present invention is
As shown in FIG. 1, in the example-based index process and the example-based search process, document representation is performed based on the characteristic structure grasp of the example document, and in the example-base search process, similarity is used to grasp the characteristic structure of the example document. It is to obtain using the document expression based on.

An example of a search system for example-based search according to the present invention is shown in FIG. Referring to FIG. 3, the search system 300 of the present invention includes an example base index unit 310 and an example base search unit 320, and the example base index unit 31.
Reference numeral 0 indicates a related technical document input unit 31 for inputting a related technical document.
1, a first keyword extraction unit 312, a first word vector expression unit 313, and a word vector storage unit 314.

The first keyword extracting unit 312 is configured to divide a paragraph in the related technical document input by the input unit according to the structural characteristics of the document, and extract a keyword for each divided paragraph. The word vector expression unit 313 is configured to obtain a weight value in each paragraph with respect to the keyword extracted from each paragraph in the first keyword extraction unit 312, and express the keyword and the weight value for each paragraph by a word vector. .

Further, the word vector storage unit 314 is configured to store the keyword expressed by the word vector in the first word vector expression unit 313 and its weight value.

On the other hand, the example base search unit 320 inputs an example document in which example techniques are described, as an example document input unit 3
21, second keyword extraction unit 322, second word vector expression unit 323, similarity calculation unit 324, and display unit 325.
including.

The second keyword extraction unit 322 is provided with an example document (eg, the example document input by the example document input unit 321).
(Eg, new technology), the paragraphs are divided according to the structural characteristics of the document, and the keywords are extracted for each of the divided paragraphs. The weight value in each paragraph for the keyword extracted from is calculated, and the keyword and its weight value are expressed by a word vector for each paragraph.

Further, the similarity calculation unit 324 and the word vector storage unit 3 and the word vectors for each paragraph with respect to the example document expressed by the second word vector expression unit 323.
The similarity between the corresponding paragraphs of the example document and the related technical document is calculated using the unit vector for each related technical document stored in 14, and the similarity between the example document and the related technical document is calculated using the paragraph similarity. The display unit 325 is further configured to arrange related technical documents in order of descending power of the similarity calculated by the similarity calculating unit 324 and provide the related technical documents to the user.

The search method according to the present invention based on the search system of FIG. 3 will be described below.

In order to perform the example-based search according to the present invention, first, the related technical document is input to the related technical document input unit 311 of the index unit 310. Then, the input related technical document is divided into paragraphs according to the structural characteristics of the document, and the keywords are classified into the divided first paragraphs by the first keyword extracting unit 31.
Extract with 2. Next, the weight value in each paragraph for the keyword extracted from each paragraph as described above is obtained, and the keyword and its weight value are expressed by a word vector in the first word vector expression unit 313 for each paragraph. Then
The keyword expressed by the word vector and its weight value are stored in the word vector storage unit 314.

Next, the example document in which the example technique is described is input to the example document input unit 321. Next, in the input example document, paragraphs are divided according to the structural characteristics of the document, and the keywords are extracted by the second keyword extraction unit 322 for each divided paragraph. Next, the weight value in each paragraph for the keyword extracted from each paragraph is obtained, and the keyword and its weight value are expressed by a word vector in the second word vector expression unit 323 for each paragraph. Then
Between the paragraphs corresponding to the example document and the related technical document using the paragraph-based word vector for the example document expressed as described above in the similarity calculator 324 and the paragraph-based unit vector stored for the related technical document in the indexing process. Is calculated, and the similarity between the example document and the related technical document is calculated using the similarity between paragraphs. Then, the related technical documents are arranged on the display unit 325 in the descending order of the calculated similarity and provided to the user.

In FIG. 4, a paragraph division unit 1221 for dividing paragraphs into the example document input to the example document input unit 121 by the structural analysis of the document and a document is expressed for each divided paragraph (a word weight value is set). ) Paragraph-specific document representation unit 1222
An example of the example document expression unit 122 of the present invention is shown.

According to FIG. 4, in the example-based retrieval method for similarity determination according to the present invention, the document representation method by grasping the characteristic structure of the document, which is commonly included in the index process and the retrieval process, is roughly divided into two. It consists of stages. That is,
Referring to FIG. 4, the document representation method according to the present invention, which is based on the patent structure analysis, includes a paragraph division unit 122 of the example document representation unit 122 for the example document input to the example document input unit 121.
In step 1, the paragraph is divided by the characteristic structure analysis of the document, and the paragraph-based document representation unit 1222 represents the document in the paragraph.
It consists of the steps of (setting the word weights).

For example, when the example technique is a patent-related invention (hereinafter, also referred to as "example invention"), it can be performed as follows. That is, in order to express the example document described in the example invention according to the present invention, first, each input example document is divided into paragraphs by grasping the patent structure. One example is to divide paragraphs into paragraphs. 1. Title of invention 2. Summary 3. Index word: When the document creator writes other important keywords of the title of the invention. Detailed Description of the Drawings 5. Detailed explanation of the present invention: Applicable when the category of "related art, purpose of invention, structure, action, effect, field of use" is described in a form not specifically described as follows6. Related Art and Technical Field of Invention 7. Technical problem to be achieved by the present invention (or object of invention): In some patent documents, the following structure may appear in a form such as "object and structure of invention" or "object, function and effect of invention". However, in such a case, it is classified into the "purpose of the invention" which is the earliest description. 8. Structure of the invention: As explained in connection with "Object of invention",
When it is created in a form such as “structure and action of invention”, it is classified as “structure of invention”. 9. Action of the invention: Similarly, when it is in the form of "action and effect of the invention", it is classified as "action of the invention". 10. Effects of the invention: Similarly, in the case of "effects and fields of use of the invention", etc., they are classified as "effects of the invention". 11. Fields of Use of Invention 12. Contents whose structure cannot be grasped: All contents whose specific patent structure is not clearly understood are classified into this structure. In the case of unstructured patent documents, the entire content is divided into this structure. 13. In the example document entered in each claim of the claims, such paragraphs are patent documents corresponding to the application format that satisfies the description requirements required by the JPO.
In the case of patent specifications, etc., they can be easily classified by the title of each paragraph.

In particular, SGML (Sta
Since it is easy to distinguish between paragraph titles and paragraphs in a patent document made up of ndard Generalized Markup Language) or a patent document created in XML (Extensible Markup Language), the paragraphs can be easily distinguished.

In the case of a patent document created in a markup language capable of easily distinguishing paragraphs as described above, the paragraph dividing unit 1221 is not performed in the present invention, and the paragraph document expression unit 1222 directly divides each paragraph. The whole document is represented by word vectors.

When the paragraphs cannot be easily divided as described above, a method of searching the subject of each paragraph is required to find a specific paragraph in the document. For example, in the case of example inventions, there is a need for a method that searches for the subject of each paragraph in order to find a specific paragraph in a document that satisfies the description requirements required by the JPO and, in particular, in a document that partially or completely does not satisfy the description requirements. Is.

An example of a preferred system and method for searching for paragraph titles in technical documents according to the present invention is shown in FIGS. 5 and 6, respectively.

FIG. 5 shows an example of a preferred paragraph title extraction system consistent with the present invention. Referring to FIG. 5, a preferred paragraph title extraction system 400 according to the present invention includes a sentence extraction unit 410, a phrase extraction unit 420, and a paragraph title display unit 43.
0, a selection unit 440, a paragraph title validity determination unit 450, and a paragraph title extraction unit 460.

The sentence extracting unit 410 is configured to extract a sentence from the related technical document or the example document respectively inputted, and the phrase extracting unit 420 makes the sentence extracting unit 410.
Configured to extract phrases from sentences extracted from
The paragraph title display unit 430 is configured to display the paragraph titles whose phrases extracted by the phrase extraction unit 420 correspond to the structure determination rule.

The selection unit 440 is configured to select a paragraph title that matches the largest number of phrases in the sentence in which the paragraph title display process is completed in the paragraph title display unit, and the paragraph title hit / miss determination unit 450 selects the selection. A constant value (0.8) by judging whether the ratio of the number of matching clauses / the total number of clauses with respect to the clause selected by the section 440 is a certain value or more, preferably 0.8 or more.
If it is more than the above, it is judged as a new paragraph title, and if the ratio is less than a constant value (0.8), it is included in the related paragraph.
It is configured to extract what is determined as a paragraph title in 50 into a paragraph title.

The paragraph title extraction system is preferably provided in each of the first keyword extraction unit 312 and the second keyword extraction unit 322.

A process of extracting the paragraph titles of the example document and the related technical document using the paragraph title extraction system shown in FIG. 5 will be described with reference to FIG.

Referring to FIG. 6, in order to search for a paragraph title of an example document or a related technical document according to the present invention, first, a sentence extracting unit 410 extracts a sentence from the input example document or a related technical document (step 510). A phrase is extracted from the sentence extracted by the phrase extracting unit 420 (step 52).
0). In the present invention, it is preferable to use a unit divided by a return character as a unit of the extracted sentence. Then, the paragraph title display unit 430 displays and accumulates the paragraph titles whose phrases extracted as described above correspond to the paragraph determination rule (step 530). The phrase extraction and paragraph title display process is repeated until the extracted sentence is completed (step 540).

Then, when the paragraph title display process is completed for the extracted sentence, the paragraph title corresponding to the most words in the sentence is selected in the selection unit 440 (step 55).
0). Next, when the ratio of the number of matching clauses / the total number of clauses with respect to the clause selected as described above in the paragraph title validity determination section 450 is a certain value or more (0.8 or more in FIG. 6). If it is 0.8 or more, it is determined as a new paragraph title, and if the ratio is less than 0.8, it is included in the related paragraph (step 560). Next, the paragraph title validity determining unit 450 determines (generates) a paragraph title as a paragraph title in the paragraph title extracting unit 460.
(Step 570). By repeating the above process until the final sentence of the document, the paragraph title of the example document can be checked (step 580).

Hereinafter, an example of a process of checking whether the extracted sentence is a paragraph title that notifies the start of a new paragraph, that is, a process of extracting a paragraph title from the extracted sentence will be described.

As a preferable example of the paragraph judgment rule for checking whether the item is the above-mentioned paragraph title, the following configuration can be given. [Paragraph name] [Cue word set (mutual OR relationship)] [Matching degree] [Necessity] $ ・ [Paragraph name]: Pointing to a specific paragraph when the following rules match ・ [Cue word set]: Starting point It means that any of them should match if they are listed together. That is, it is the relation of OR. -[Matching degree]: There are three kinds of matching degree. The "+" display must exactly match the input phrase with the word in the clue word set, the "-" display must partially match the clue word, and the "=" display will show the clue word. This means that the paragraphs will match exactly. That is, it means that it is clearly understood that it is the first sentence of the structure without the need to apply other rules.・ [Necessity]: There are two types. The "y" type means that the current rules must be satisfied in order to be recognized as its structure, and the "n" type is not necessarily required to be recognized as the current structure, and can be Means that. $: A recognizer that distinguishes the end of one rule.

In order to check whether the extracted sentence is a paragraph title, first, words are extracted from the extracted sentence,
The extracted words are compared with the rules to determine whether they match. For example, the rules for understanding the structure of the "detailed description of the invention" are as follows. 6 {Simple drawing of the drawing on the drawing name side} -y
$ 6 {Attachment Description Name Name Name Simplified Explanation Book Name
Composition} − y $ 6 {Simple detailed} + n $ 6 {Invented invention} + n $ 6 {Each book} − n $ 6 {Main} − n $ 6 {Compare} + n $ 6 {Sign} − n $ 6 {part} − n $

In the above example, "6" is "detailed description of the drawing".
It is a "paragraph name" field that indicates. And the first rule is "drawing, design, drawing name, face, simple drawing,
References to six clue words, such as "Figure," mean that they may "partially match" the input phrase.
Here, "partial match" means that when the given sentence is "detailed description of the drawing", the clue word "drawing" partially matches the phrase "of the drawing".

The third rule is an exact match of "+" and is applied to the word "detailed" in the preceding input sentence. In this case, if the given text is "the description is more detailed if the drawing is detailed ...", the first rule matches "the drawing is". However, even if the third rule partially matches "if it is detailed", it does not apply because it does not exactly match any word.

Among the rules for the above "Detailed Description of Drawings", the rules which must be applied are the first and second rules indicated by "y". That is, in order to indicate that the input sentence is the beginning of a paragraph for the "detailed description of the drawing", that is, the paragraph title, there must be a phrase satisfying these two rules in the input sentence.

The most important rule in the process of analyzing the characteristic structure of a document is that a new paragraph title is obtained when 80% or more of the entire phrases of the input sentence are correctly examined by the rule indicating one paragraph. Is. For example, when a sentence such as “as described in the description of drawings, drawing 1 is…” is an input sentence, “drawing” and “explanation” are applied to the first and second rules described above. There is a high possibility that it will be a paragraph title corresponding to "Detailed explanation of", but the rest of the words in the whole sentence are not applied by other rules, so that structure, that is, the paragraph title, is not analyzed.

Next, as described above, words are extracted for each paragraph divided by grasping the characteristic structure of the document, the weighted value is obtained, and the word and the weighted value are expressed by a word vector.

In the present invention, a document is represented by a vector of paragraphs, unlike displaying a document by one vector in the related example base search. For example, if the example technology is a patent-related invention, the example document (patent document) as described above
Can be defined in 13 paragraphs, of which 13 claims, each claim of which is dependent on the patent document. Therefore, in the present invention, there are at least 13 or more vectors for each patent document, and a document can be represented by 13 or more vectors.

Therefore, the patent document D _i is
When expressed by (F _i1 , F _i2 , ..., F _im ), m has a value of 13 or more.

Furthermore, each paragraph F _ij is represented by an n-dimensional word vector (W _{ij, 1} , W _{ij, 2} , ..., W _{ij, n} ). At this time, W _{ij, q} indicates the weighted value of the word T _q in the paragraph F _ij of the document D _i . Therefore, the equation (1) of the related weight calculation method must be changed to the following equation (4).

[0072]

[Equation 4] _{(Where, tf ij, q:} frequency number of documents _{D i} word _{T q} in paragraph _{F ij} of, _{max tf:} paragraph of the document _{D i} F
₍ The frequency of the most frequently occurring words in _ij , N: the number of documents in total, n: the number of documents in which the word T _q appears)

The document expression method based on the characteristic structure analysis (grasping) of the document is commonly used for the example base index unit and the search unit of the present invention. In the example-based index, it is preferable that such a representation method is stored in the related technical document storage unit of the index unit as an inverted index file storage structure as in the conventional method, so that the index can be quickly searched.

An example-based search process consistent with the present invention is
As shown in FIG. 1, the similarity calculation unit of the search unit includes a step of determining a technical similarity as described later. For example, in determining the patentability of a new invention, the technical similarity, that is, the patent similarity determination process is as follows.

In the example-based search process of the present invention,
In order to enable patent similarity calculation that is useful for determining patentability,
The patent specialist follows the logical act of making a patentability decision.

Generally, in order to determine the patentability of a new invention, the patent specialist determines that the patentability between the two patents is the earliest when the claims appear in the same form in the related patent. . Then, when the purpose and effect of both patents are similar to the subject, it is considered that the patentability is similar, and then it is examined whether or not each claim and the structure and operation of the invention are similar.
The rest of the content will be examined later.

Therefore, also in the present invention, if the patent similarity judgment of both patent documents is 100% even if one of the claims of both patents is the earliest, the patent similarity of both patents is 100% without letting go. However, it is preferable to consider that there is a high degree of patent similarity between the two patents if the important paragraphs are similar to each other.

For example, when the example document and the related technical document are patent documents (patent specifications), it is preferable to determine the similarity between the example document and the related technical document as follows.

If any one of the same claims exists in the claim paragraphs of both documents, the two patents are judged to be the same, and there is no same claim in the claim paragraphs of both documents. In this case, the similarity determination is performed as follows.

That is, after calculating the similarity between the "title of invention, object of invention and effect of invention" paragraphs of both documents, the highest similarity among the similarities thus obtained is selected as the representative similarity value of these paragraphs. Given the highest weighted value, the similarity between the "abstract, composition of invention, action of invention, and claim" paragraphs was calculated, and the highest similarity among the similarities thus obtained was selected as the representative similarity value of these paragraphs. After giving the post-weighting value, determining the similarity between the remaining paragraphs and selecting the average value thereof as the representative similarity value, the lowest weighting value is given. The similarity is determined by comparing the values obtained by multiplying the respective representative similarity values by the respective weighted values.

On the other hand, the inter-paragraph similarity judgment according to the present invention can be carried out by using the similarity equation in which the document vector is changed to the paragraph vector in the equations (2) and (3). In the present invention, it is preferable to obtain the inter-paragraph similarity using the cosine similarity expression of the above equation (2).

Similarity between the j-th paragraph F _{ij of the} example document D _i and the q-th paragraph F _{pq of the} related technical document D _p sim_
F (F _ij , F _pq ) can be defined by the following equation (5).

[Equation 5] (Where W _{ij, l} : the word T _l is the weight value in the j-th paragraph F _ij of the document D _i , W _{pq, l} : the word T _l is the weight value in the j-th paragraph F _pq of the document D _p )

Since the above equation (5) uses the cosine similarity equation, the inter-paragraph similarity sim_F (F _ij , F _pq ) is always 0.
The value is between 1 and 1. Similarity between both paragraphs sim_F (F
_{When ij} and F _pq ) are 1, it means that the vectors are 100% mutually coincident.

In the paragraph notation F _ij , j matches the paragraph order obtained from the paragraph grasp described above. Thus, for example, F _i1 means "entitled" paragraph patent document D _i, F _i2 means "abstract" paragraph patent documents D _i.

Next, after the inter-paragraph similarity is calculated as described above, the example document (example technology) D _i and the related technical document (related technology) between _{D p} technical similarity sim_P _(D i, it is preferable to determine the _{D p).}

[Equation 6]

For example, when the example document and the related technical document are patent documents, the first term of the formula (6) is the patent similarity between the two patents if the claims of both patents match. This is a mathematical expression that sim_P (D _i , D _p ) is 1 and is considered to be completely patentable. Then, the second term, if not, means to obtain the paragraph importance described above. At this time, α, β, and μ indicate the importance of each paragraph group. Therefore, the sum of α, β, and μ must always be 1. In the present invention, α value is set to 0.5, β value is set to 0.3, and μ value is set to 0.2 by experiments.

Therefore, the second term is, for example, a value obtained by multiplying the highest value among the similarities of “object of invention”, “effect of invention” and “title of invention” by a weight value of 0.5, and “summary”. Value, which is obtained by multiplying the highest value from the sum of the “book”, the “structure of the invention”, the “object of the invention”, and the “claims” by a weighted value of 0.3, and finally, the average similarity between the remaining paragraphs is 0.2. It means a mathematical expression for obtaining the patent similarity between both patents by a value obtained by adding all the values multiplied by.

The above formula (6) is merely an example, and it goes without saying that the importance of each paragraph group can be set to two or four or more instead of three, and the value of the importance can be changed.

When the technical similarities of all the related technical documents to the example document D _i are calculated by the equation (6), they are arranged in descending order of power and provided to the user. Then, the user can search for related technologies in the order of technical similarity to the example technology.

The preferred embodiments of the present invention are disclosed for the purpose of illustration, and those skilled in the art can make various modifications, changes and additions within the spirit and scope of the present invention. Modifications and changes should be regarded as belonging to the technical scope of the present invention.

[0091]

As described above, according to the present invention, technically similar related documents are displayed together with their degree of similarity, so that the similarity can be judged easily and quickly.
Furthermore, when the present invention is used at the time of filing or filing a new invention, the patentability judgment specialist or the inventor can compare similar related documents together with the degree of similarity, so that the patentability of the invention can be easily and quickly obtained. The effect that can be judged is produced. Further, the present invention can directly provide a document in which the technology is described to the system of the present invention, so that the user does not need to learn and understand the technology, and thus the search time can be significantly shortened.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of an example-based search system for similarity determination according to the present invention.

FIG. 2 is an overall configuration diagram of a normal example-based search system.

FIG. 3 is a configuration diagram showing an example of a preferred example-based search system according to the present invention.

FIG. 4 is a flowchart showing an example of a method of expressing a document by grasping a specific structure of the document according to the present invention.

FIG. 5 is a configuration diagram showing an example of a paragraph title extraction system for extracting a paragraph title by grasping a specific structure of a document according to the present invention.

FIG. 6 is a flowchart showing an example of a method of extracting a paragraph title by grasping a specific structure of a document according to the present invention.

[Explanation of symbols]

100, 300 ... Search system, 110, 310 ... Index section, 111, 311 ... Related technical document input section, 112 ... Related technical document expression section, 113 ... Related technical document storage section, 120, 320 ... Search unit, 121, 321 ... Example document input section, 122 ... Example document expression part, 123, 324 ... Similarity calculation unit, 312 ... the first keyword extraction unit, 313 ... First word vector expression part, 314 ... Word vector storage section, 322 ... the second keyword extraction unit, 323 ... Second word vector expression part, 325 ... Display, 400 ... Paragraph title extraction system, 410 ... a sentence extraction unit, 420 ... a word extraction unit, 430 ... Paragraph title display section, 440 ... Selector 450 ... Paragraph title judgment unit, 460 ... Paragraph title extraction unit.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Park Johnson             South Korea, Cholanand, Don Kwangi             Yang-si, Kang Ho-dong 700 Kwang             Gyan Iron Factory (72) Inventor Piyun Jin             South Korea, Cholanand, Don Kwangi             Yang-si, Kang Ho-dong 700 Kwang             Gyan Iron Factory (72) Inventor Kim Jin Sun             Republic of Korea, Kyongsan Book-do, Poha             Ngushi, Nanku, Donchon-Don             Within 5 POSCO (72) Inventor Song Nam Gong             Republic of Korea, Kyongsan Book-do, Poha             Ngushi, Nank, Cordon-Don 1               In POSCO (72) Inventor Lee Jong Hyuk             Republic of Korea, Kyongsan Book-do, Poha             Ngushi, Nanku, Hyoja Don, Sa             31 Pohang University             In Buu Science and Technology (72) Inventor Kwon Ou             Republic of Korea, Kyongsan Book-do, Poha             Ngushi, Nanku, Hyoja Don, Sa             31 Pohang University             In Buu Science and Technology F term (reference) 5B075 ND03 NK02 NK32 NR05 PP24                       PQ02 PQ46 PQ74 PR06 QM08

Claims (8)

[Claims] 1. The method includes an example-based index process and an example-based search process, wherein the example-based index process inputs a related technical document; the input related technical document is divided into paragraphs according to structural characteristics of the document, and is divided. Extracting a keyword for each selected paragraph; determining a weight value in each paragraph for the keyword extracted from each paragraph as described above, and expressing the keyword and its weight value in each paragraph by a word vector;
And storing the keyword expressed by the word vector and its weight as described above, and the example-based search process inputting an example document in which example techniques are described; The paragraph is divided according to the structural characteristics of the paragraph, and the keyword is extracted for each of the divided paragraphs; as described above, the weight value in each paragraph is obtained for the keyword extracted from each paragraph, and the keyword and its weight value are obtained for each paragraph. Expressing with a word vector;
And a similarity between corresponding paragraphs of the example document and the related technical document is obtained using the paragraph-based word vector for the example document expressed as described above and the paragraph-based word vector for the related technical document stored in the indexing process. A step of obtaining a similarity between the example document and the related technical document using the similarity between paragraphs; and a step of arranging the related technical documents and providing them to the user in descending order of the similarity calculated as described above. An example-based search method for similarity determination comprising. 2. In the indexing process and the searching process, the paragraph division between the related technical document and the example document is performed based on the paragraph title described in the patent document satisfying the description requirement required by each national patent office. An example-based search method for determining similarity according to claim 1. 3. A step of extracting a sentence from the input example document and extracting a phrase from the extracted sentence for a paragraph title of the document; a paragraph in which the extracted phrase corresponds to the structure judgment rule. Displaying the title; Repeating the phrase extraction and paragraph title display process until the extracted sentence is finished; When the paragraph title display process is completed for the extracted sentence, it matches with the largest number of phrases in the sentence Selecting a paragraph title to be selected; it is determined whether the ratio of the number of matching phrases / the total number of phrases to the selected phrase as described above is 0.8 or more, and if it is 0.8 or more. If the ratio is less than 0.8, it is included in the relevant paragraph; and the step is repeated until the final sentence of the document, and the paragraph title for the example document is extracted. The type according to claim 2, characterized in that Example based search method for the sex judgment. 4. The example-based search method for determining similarity according to claim 3, wherein the structure determination rule is configured as follows. [Paragraph name] [Cue word set (mutual OR relationship)] [Match degree] [Necessity] $ {where [Paragraph name]: [Cue word set, which indicates a specific paragraph when the following rules match] ]: A list of starting words, which means that if they are listed together, only one of them may match, that is, the relationship of OR. [Matching degree]: There are 3 kinds of matching degree However, "+" display must exactly match the input phrase with the word in the clue word set, "-" display must partially match the clue word, and "=" indicates the clue It means that the paragraphs will be matched exactly as long as the word appears, that is, it is possible to clearly judge that it is the first sentence of the structure without applying other rules. [Necessity]: 2 There is a type, but because the "y" type is recognized as its structure Not in the sense that should the current rule is satisfied, "n"
The type is not necessarily required to be recognized as the current structure, and means that it is possible. $: A recognizer that distinguishes the end of one rule. } 5. The example document and the related technical document consist of a patent specification, and the similarity between the example document and the related technical document is determined by the same claim if any of the same claims exists in the claim paragraphs of the two documents. Determining that both patents are the same; and if there are no identical claims in the claims paragraphs of both documents,
The highest weighted value after selecting the similarity between the "title of invention, object of invention and effect of invention" paragraphs of both documents and selecting the highest similarity among the thus obtained similarity as the representative similarity value of these paragraphs Then, the similarity between paragraphs of “abstract, structure of invention, action of invention, and claims” is obtained, and the highest similarity among the thus obtained similarity is selected as the representative similarity value of these paragraphs, and then the weighted value is calculated. And give the similarity between the remaining paragraphs,
After the average value is selected as the representative similarity value, the lowest weighted value is given, and the value obtained by multiplying the representative similarity value by each weighted value is compared to determine the similarity. The example-based search method for determining similarity according to claim 1. 6. An example base index unit and an example base search unit, wherein the example base index unit inputs a related technical document, a related technical document input unit, and a structural structure of a document in the related technical document input by the input unit. A first keyword extracting unit that divides paragraphs according to characteristics and extracts a keyword for each divided paragraph; a weighting value in each paragraph is calculated for each keyword extracted from each paragraph in the first keyword extracting unit, and a keyword for each paragraph and A first word vector expression unit that expresses the weighted value in a word vector; and a word vector storage unit that stores the keyword expressed in the word vector in the first word vector expression unit and the weighted value, and the example base The search unit inputs an example document in which example techniques are described; an example document input unit; input by the example document input unit A second keyword extracting unit that divides paragraphs according to the structural characteristics of the document in the generated example document and extracts a keyword for each divided paragraph; within each paragraph for the keyword extracted from each paragraph in the second keyword extracting unit. A second word vector expression unit for obtaining the weighted value of each keyword and expressing the weighted value as a word vector for each paragraph; the word vector for each paragraph and the word vector storage unit for the example document expressed in the second word vector expression unit. The similarity between the corresponding paragraphs of the example document and the related technical document is obtained by using the paragraph-based word vector for the stored related technical document, and the similarity between the example document and the related technical document is calculated by using the paragraph similarity. Similarity calculation unit to be sought; and display for arranging related technical documents in order of descending power of similarity calculated by similarity calculation unit and providing to user Example-based retrieval system for the comprising at similarity determination. 7. The first keyword extracting unit and the second keyword extracting unit respectively divide paragraph sections of a related technical document and an example document into a paragraph subject described in a patent document satisfying a description requirement required by each national patent office. The example-based search system for similarity determination according to claim 6, wherein the example-based search system is configured to be performed based on the above. 8. The first keyword extracting section and the second keyword extracting section extract a sentence from a related technical document or an example document respectively input; a sentence extracting unit; a phrase from a sentence extracted from the sentence extracting unit. A paragraph extraction section that displays the paragraph titles whose phrases extracted from the phrase extraction section correspond to the structure judgment rule; the paragraph title display process is completed in the paragraph title display section. A selection unit that selects a paragraph title that matches the largest number of phrases in the sentence; is the ratio of the number of matching phrases / the total number of phrases to the phrase selected in the selecting unit 0.8 or more? Judgment, 0.
If it is 8 or more, it is judged as a new paragraph title, and if the ratio is less than 0.8, it is included in the related paragraph. The example-based search system for similarity determination according to claim 7, further comprising a paragraph title extraction system including a paragraph title extraction unit for extracting a title.