JP2012003550A - Document retrieval method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To construct a retrieval formula which is appropriate for retrieving a document matched with a purpose of retrieval even when there exists any peculiar negative example.SOLUTION: This document retrieval method includes: acquiring categories into which positive examples are classified in a document database classified into a plurality of hierarchized categories (step S2); acquiring feature words/phrases included in the positive examples classified into the categories, that is, feature words/phrases not included in negative examples classified into categories as category feature words/phrases for every acquired category (step S4); constructing a sub-retrieval formula based on the categories and the acquired category feature words/phrases for every acquired category (step S5); acquiring the feature words/phrases included in the positive examples (step S3); and constructing a main retrieval formula based on the sum set of the acquired feature words/phrases and the constructed sub-retrieval formula (step S6).

Description

  The present invention relates to a document search method and program for searching for a document that matches a search purpose. In particular, the present invention relates to a document search method and program for a document database classified into a plurality of hierarchical categories.

  Conventionally, there is a case in which a document that matches a search purpose is searched from a document database classified into a plurality of hierarchized categories as in a patent document search, for example. In this case, for example, a search formula is constructed and a search is performed using this search formula.

  The construction of a search expression is a condition for extracting documents that match the search objective (hereinafter referred to as “positive examples”), and excludes documents that do not match the search objective (hereinafter referred to as “negative examples”). It is considered that logical concept learning that seeks conditions is effective. Logical concept learning includes deductive learning techniques and inductive learning techniques. As a deductive learning method, there is learning based on explanation (EBL), and as an inductive learning method, there is learning based on similarity (SBL: Similarity Based Learning) (for example, see Non-Patent Document 1). ).

  The plurality of hierarchized categories can subdivide the conceptual structure into a tree structure and can be regarded as domain knowledge in EBL. For this reason, according to EBL, it is possible to obtain a condition for appropriately classifying positive examples based on domain knowledge, and to construct a search expression based on categories.

  On the other hand, according to SBL, in a state where words / phrases appropriate for a search purpose, that is, significant words / phrases are selected in advance, a combination of words / phrases including positive examples and excluding negative examples is obtained, and searching by words / phrases is performed. An expression can be constructed.

Nagao Makoto "Iwanami Course Software 14; Knowledge and Reasoning", Iwanami Shoten

  However, in EBL, if there is a specific negative example, there is a problem that a category in which the negative example can be excluded cannot be obtained by subdivision within the domain knowledge, and a search expression based on the category may not be appropriately constructed. .

  On the other hand, in SBL, it is difficult to set in advance how much significance a phrase should be selected, and there is a problem that a search expression based on the phrase may not be appropriately constructed.

  The present inventor has intensively studied to solve the above-described problems. As a result, in a document database classified into a plurality of hierarchical categories, a category in which positive examples are classified is obtained, and a combination of words and phrases that excludes negative examples from documents belonging to the obtained categories is obtained. The inventors have found that an appropriate search expression can be constructed, and have completed the present invention.

  (1) The present invention matches a search purpose from a document database (e.g., corresponding to a patent database described later) classified into a plurality of hierarchical categories (e.g., equivalent to a patent classification hierarchy in FIG. 6). A document search method for searching for documents, in which positive examples (e.g., corresponding to Patent Documents 1, 2, 4 to 6, and 9 to 19 in FIG. 7) are classified among a plurality of documents included in the document database. A first step of acquiring a category (for example, corresponding to step S2 in FIG. 2) and a feature word / phrase included in the positive example (for example, lighting fixture of formula (4), environmental control, electric blind, low cost, solar radiation) A second step (equivalent to step S3 in FIG. 2) for acquiring a state acquisition unit (equivalent to interlocking control) and a positive example classified into the category for each category acquired in the first step Included in Among the plurality of documents, feature words / phrases not included in the negative examples classified into the category are classified into category feature words / phrases (for example, lighting equipment, air conditioner, and lighting equipment of formula (1)). For each category acquired in the third step (e.g., corresponding to step S4 in FIG. 2) and the category acquired in the third step. Acquired in the fourth step (for example, corresponding to S5 in FIG. 2) and the second step for constructing a sub-search expression (for example, corresponding to the expressions (1) to (3)) based on the feature words and phrases A fifth step (for example, FIG. 2) for constructing the search formula based on the union of the feature words (for example, corresponding to the formula (4)) and the sub-search formula constructed in the fourth step. Equivalent to step S6) It proposes a document retrieval method with features.

  According to the present invention, in the document database classified into a plurality of hierarchized categories, the category into which the positive example is classified is acquired. Then, for each acquired category, a feature word / phrase included in the positive example classified into the category and not included in the negative example classified into the category is acquired as a category feature word / phrase. For each acquired category, a sub-search expression is constructed based on the category and the acquired category feature phrase. Further, feature words / phrases included in the positive examples are acquired. Then, the search formula is constructed based on the acquired union of feature words and the constructed sub search formula. For this reason, in a document database classified into a plurality of hierarchized categories, a category in which positive examples are classified can be obtained, and a feature word or phrase that excludes negative examples can be obtained from documents belonging to the obtained categories. . Therefore, even if there are unique negative examples, this search expression that can select significant feature words and exclude negative examples can be constructed, and an appropriate search expression to search for documents that match the search purpose Can be built.

  (2) In the document search method according to (1), in the first step, after obtaining a category in which the positive example is classified, all documents belonging to the positive example are selected from the category. The minimum number of categories to be included is extracted, and in the second step, after obtaining the feature words included in the positive example, the minimum number of all the documents belonging to the positive example are included from the feature words. Feature words / phrases are extracted, and in the third step, the category feature words / phrases are obtained for all documents belonging to positive examples classified into the category for each of the minimum number of categories extracted in the first step. Later, a minimum number of category feature phrases that include all documents belonging to the positive example are extracted from the category feature phrases, and in the fourth step, the most extracted category feature phrases are extracted in the first step. For each number of categories, a sub-search expression is constructed based on the category and the minimum number of category feature phrases extracted in the third step. In the fifth step, in the second step A document search method is proposed in which the search expression is constructed based on the union of the extracted minimum number of feature words and the sub-search expression constructed in the fourth step.

  According to the present invention, the minimum number of categories including all documents belonging to the positive examples, the minimum number of feature words / phrases including all documents belonging to the positive examples, and the minimum number including all documents belonging to the positive examples. A number of category feature phrases are extracted. Then, the search formula is constructed using the extracted minimum number of categories, the extracted minimum number of feature words and the extracted minimum number of category feature words. For this reason, it is possible to prevent the same category, the same feature word, and the same category feature word from being included in the union and sub-search expression of the feature words constituting the search expression. Therefore, the structure of this search formula can be simplified, and a search formula for searching for a document that matches the search purpose can be constructed more appropriately.

  (3) In the document search method according to (2), the first step is a first setting step (for example, obtaining a category in which the positive example is classified and setting the category as a category candidate (for example, , Corresponding to step S21 in FIG. 3), and register the positive example documents belonging to the category candidate as elements of the category candidate, and arrange the category candidates in descending order based on the number of registered positive example documents. The first sorting step to be replaced (e.g., corresponding to steps S22 and S23 in FIG. 3), the highest category candidate is adopted, and a positive example document belonging to the highest category candidate is extracted from other category candidate elements. A first adopting step (for example) of deleting a category candidate having no elements from the category candidates set in the first setting step 3), and the first alignment step and the first adoption step are repeated until the number of other category candidates becomes 0, and the first adoption step A first extraction step (for example, corresponding to steps S27 and S28 in FIG. 3) that extracts the adopted category candidates as the minimum number of categories, and the second step is included in the positive example A second setting step (for example, corresponding to step S31 in FIG. 4), and a positive example including the feature word candidate as an element of the feature word candidate A second alignment step (for example, steps S32 and S in FIG. 4), in which the feature word candidates are rearranged in descending order based on the number of registered positive examples. 33) and the top-level feature phrase candidate are adopted, the original document including the top-level feature phrase candidate is deleted from the other feature phrase candidate elements, and the feature phrase candidates with no elements are deleted. A second adoption step (for example, corresponding to steps S34 to S36 in FIG. 4) that is excluded from the feature word candidates set in the second setting step, and the number of other feature word candidates is zero. Until the second alignment step and the second adoption step are repeated until the feature phrase candidates adopted in the second adoption step are extracted as the minimum number of feature phrases (for example, Corresponding to steps S37 and S38 in FIG. 4), and the third step acquires the category feature phrase for each of the minimum number of categories extracted in the first step. A third setting step (for example, corresponding to step S41 in FIG. 5) for setting a category feature word / phrase as a category feature word / phrase candidate and registering a positive example document including the category feature word / phrase candidate as an element of the category feature word / phrase candidate And a third alignment step (for example, corresponding to steps S42 and S43 in FIG. 5) for rearranging the category feature words in descending order based on the number of registered positive examples. The candidate is adopted, and the positive example document including the highest category feature phrase candidate is deleted from the elements of the other category feature phrase candidates, and the category feature phrase candidate having no elements is set in the third setting step. The third adoption step (for example, steps S44 to S46 in FIG. ), And repeating the third alignment step and the third adoption step until the number of other category feature phrase candidates becomes zero, and the category feature phrase candidates adopted in the third adoption step are The document search method is characterized by comprising a third extraction step (for example, corresponding to steps S47 and S48 in FIG. 5) that extracts as the minimum number of category feature words.

  According to the present invention, for each category, feature phrase, and category feature phrase candidate, corresponding positive example documents are registered and sorted in descending order based on the number of corresponding positive example documents. . Then, the highest candidate is adopted, and the positive example document corresponding to the highest candidate is deleted from the other candidate elements, and the candidate having no elements is excluded from the rearranged candidates. Then, the above-described rearrangement and the above-described deletion and exclusion are repeated until the number of other candidates becomes 0 (zero). For this reason, the minimum number of categories encompassing all documents belonging to the positive example, the minimum number of feature words including all documents belonging to the positive example, and the minimum number of categories including all documents belonging to the positive example Feature phrases can be extracted.

  (4) In the document search method according to any one of (1) to (3), the present invention constructs the main search formula among documents searched from the document database based on the main search formula. A document search method comprising: a sixth step (for example, corresponding to step S102 in FIG. 12) that rearranges documents excluding documents used at the time of sorting in descending order based on the degree of conformity with the search formula. Has proposed.

  According to the present invention, out of the documents searched from the document database based on the search formula, those excluding the document used when constructing the search formula are determined based on the degree of conformity with the search formula. Sort in descending order. For this reason, documents that have not been confirmed whether or not they match the search purpose can be checked in descending order of suitability with this search formula, and the search for documents that match the search purpose can be performed efficiently. Can be done.

  (5) In the document search method according to any one of (1) to (4), the present invention is characterized in that the category feature word / phrase is not included in the feature word / phrase acquired in the second step. A document retrieval method is proposed.

  According to the present invention, the feature phrase included in the positive example and the category feature phrase are different words. For this reason, it is possible to prevent the feature word / phrase included in the union of feature words / phrases and the category feature word / phrase included in the sub-search expression from including the same word / phrase. Therefore, it is possible to prevent the same search condition from being set by the union of feature words and the sub search expression, and to construct a search expression for searching for a document that matches the search purpose more appropriately. Can do.

  (6) In the document search method according to any one of (1) to (5), in the third step, each category acquired in the first step is classified into the category. A feature word / phrase included in an example, which is a negative example classified into the category among the plurality of documents and not included in the negative example including the feature word / phrase acquired in the second step. The document search method is characterized by being acquired as a category feature word.

  According to the present invention, the category feature phrase acquired for each category into which the positive example is classified is a feature phrase included in the positive example classified into the category, and is a negative example classified into this category. The feature words / phrases are not included in the negative example including the feature words / phrases included in the positive example. Therefore, it is possible to more efficiently acquire category feature words used to construct this search expression.

  (7) In the document search method according to any one of (1) to (6), the document is a patent document, and the category is any one of FI, IPC, ECLA, USC, or a theme code. We propose a document retrieval method characterized by

  According to the present invention, when a document is a patent document, FI, IPC, ECLA, USC, and a theme code can be used as a category used to construct this search expression.

  (8) The present invention relates to the document search method according to any one of (1) to (7), wherein the document is a Japanese patent document, and an F-term is used instead of the feature word / phrase A search method is proposed.

  According to this invention, when a document is a Japanese patent document, F terms can be used instead of feature words.

  (9) The present invention matches the search purpose from a document database (e.g., corresponding to a patent database described later) classified into a plurality of hierarchized categories (e.g., equivalent to the patent classification hierarchy of FIG. 6) A program for causing a computer to execute a document search method for searching for a document, and a positive example (for example, Patent Documents 1, 2, 4 to 6, 9 to 9 in FIG. 7) of a plurality of documents included in the document database The first step (for example, corresponding to step S2 in FIG. 2) for obtaining a category into which the category is classified, and the feature word / phrase (for example, the lighting fixture of the formula (4), environment control) A second step (e.g., corresponding to step S3 in FIG. 2) for acquiring an electric blind, a low cost, a solar radiation state acquisition unit, and interlocking control), and a category acquired in the first step. For each feature word / phrase included in the positive example classified into the category and not included in the negative example classified into the category among the plurality of documents, a category feature word / phrase (for example, an expression) For each category acquired in the first step (equivalent to step S4 in FIG. 2) and the third step acquired as (1) lighting equipment, air conditioner, and lighting fixture) A fourth step (for example, S5 in FIG. 2) for constructing a sub-search expression (for example, equivalent to the expressions (1) to (3)) based on the category and the category feature phrase acquired in the third step. ), The union of the feature words acquired in the second step (e.g., equivalent to the equation (4)), and the sub-search expression constructed in the fourth step A fifth step in constructing an expression (eg, figure 2) (corresponding to step S6 in FIG. 2).

  According to the present invention, by causing a computer to execute a program, a category into which positive examples are classified is acquired in a document database classified into a plurality of hierarchical categories. Then, for each acquired category, a feature word / phrase included in the positive example classified into the category and not included in the negative example classified into the category is acquired as a category feature word / phrase. For each acquired category, a sub-search expression is constructed based on the category and the acquired category feature phrase. Further, feature words / phrases included in the positive examples are acquired. Then, the search formula is constructed based on the acquired union of feature words and the constructed sub search formula. For this reason, in a document database classified into a plurality of hierarchized categories, a category in which positive examples are classified can be obtained, and a feature word or phrase that excludes negative examples can be obtained from documents belonging to the obtained categories. . Therefore, even if there are unique negative examples, this search expression that can select significant feature words and exclude negative examples can be constructed, and an appropriate search expression to search for documents that match the search purpose Can be built.

  According to the present invention, even when a specific negative example exists, it is possible to construct this search expression that can select a significant feature word and exclude a negative example, and to search for a document that matches the search purpose. Appropriate search expressions can be constructed.

It is a block diagram which shows the structure of the document search apparatus which concerns on one Embodiment of this invention. It is a flowchart of the document search process which concerns on one Embodiment of this invention. It is a flowchart of the patent classification extraction process which concerns on one Embodiment of this invention. It is a flowchart of the characteristic phrase extraction process which concerns on one Embodiment of this invention. It is a flowchart of the category characteristic word phrase extraction process which concerns on one Embodiment of this invention. It is a figure which shows the hierarchical structure of a patent classification. It is a figure for demonstrating the specific example in the case of extracting the minimum number of patent classifications by a patent classification extraction process. It is a figure for demonstrating the case where the concept acquisition of "blind" is performed by SBL. It is a figure for demonstrating this search expression construction example. It is a figure for demonstrating this search expression construction example. It is a figure for demonstrating this search expression construction example. 3 is a flowchart of a dynamic screening method according to an embodiment of the present invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Note that the constituent elements in the following embodiments can be appropriately replaced with existing constituent elements, and various variations including combinations with other existing constituent elements are possible. Accordingly, the description of the following embodiments does not limit the contents of the invention described in the claims.

<Configuration of Document Retrieval Device 1>
FIG. 1 is a block diagram showing a configuration of a document search apparatus 1 according to an embodiment of the present invention.

  The document search device 1 is a device for searching for patent documents such as published patent gazettes and patent gazettes, and includes a control unit 11, a communication unit 12, a storage unit 13, and an input / output interface (input / output I / F) 14.

  The control unit 11, the communication unit 12, the storage unit 13, and the input / output I / F 14 are connected to each other via the bus 15. An input device 16 and an output device 17 are connected to the input / output I / F 14.

  The control unit 11 includes a CPU (Central Processing Unit) and a RAM (Random Access Memory), and operates according to a program stored in the storage unit 13.

  The storage unit 13 includes a hard disk, a semiconductor memory, and the like, and stores a patent document search program executed by the control unit 11 and a patent database in which patent document data is accumulated. The patent document search program is a program for searching for a patent document matching the search purpose from the patent database. When executed by the control unit 11, a document search process, which will be described later with reference to FIG. 2, is performed. It will be.

  The communication unit 12 is configured by a network interface card (NIC), and transmits information received via the network to the control unit 11 or sends information to the outside via the network according to a command input from the control unit 11. Or send.

  The input device 16 includes a keyboard and a mouse, and receives input from the user.

  The output device 17 is configured by a display or a printer, and outputs information corresponding to the processing performed by the control unit 11.

<Document search processing>
FIG. 2 is a flowchart of the document search process described above. When an input for executing a patent document search program is received from the user from the outside via the input device 16 or the communication unit 12, a document search process is started and a patent document matching the search purpose is searched from the patent database. Therefore, this retrieval formula is constructed.

  Specifically, first, in step S <b> 1, the control unit 11 reads a positive example and a negative example from the storage unit 13. A positive example is a patent document that matches the search purpose among a plurality of patent documents included in the patent database, and a negative example is a patent that does not match the search purpose among the patent documents included in the patent database. It is the literature. These positive examples and negative examples are selected in advance from a plurality of patent documents included in the patent database in the preliminary search.

  The preliminary search is performed in advance in order to perform the main search using the main search formula. In this preliminary search, some of the plurality of patent documents included in the patent database are extracted, and the user confirms whether or not each of the extracted patent documents matches the search purpose based on the description. According to this, the positive example and the negative example used in step S1 will be selected.

  Next, in step S2, the control unit 11 performs a patent classification extraction process which will be described later with reference to FIG. In this patent classification extraction process, a patent classification in which a positive example is classified is acquired, and a minimum number of patent classifications including all patent documents belonging to the positive example are extracted from the acquired patent classification.

  Next, in step S <b> 3, the control unit 11 performs a feature phrase extraction process to be described later with reference to FIG. 4. In this feature word / phrase extraction process, feature words / phrases included in the positive example are acquired, and a minimum number of characteristic words / phrases that include all patent documents belonging to the positive example are extracted from the acquired characteristic words / phrases. Here, the characteristic phrase is a phrase that characterizes the content of the patent document among the words included in the patent document.

  Next, in step S4, the control unit 11 performs a category feature phrase extraction process to be described later with reference to FIG. In this category feature phrase extraction process, a category feature phrase is acquired for each of the minimum number of patent classifications extracted in step S2, and the minimum number of all patent documents belonging to positive examples from the acquired category feature phrases is included. Extract category feature words. Here, the category feature word / phrase is a feature word / phrase included in a positive example classified into the patent classification for each minimum number of patent classifications acquired in step S2, and is a negative example classified into the patent classification. Is a feature word not included.

  The category feature word / phrase is a feature word / phrase included in the positive example classified in the patent classification for each minimum number of patent classifications acquired in step S2, and is included in the negative example classified in the patent classification. It may be a feature phrase that is not included in the feature phrase that is not included in the feature phrase acquired in step S3. In this case, in the category feature phrase extraction process, the minimum number of category feature phrases is extracted by excluding the feature phrase acquired in step S3 from the extracted minimum number of category feature phrases.

  Further, in the category feature phrase extraction process, each of the minimum number of patent classifications extracted in step S2 is a feature phrase included in the positive patent document classified in the patent classification, and is classified into the patent classification. A feature word / phrase that is a negative example and is not included in the negative example including the feature word / phrase acquired in step S3 is acquired as a category feature word / phrase, and all patent documents belonging to the positive example are included from the acquired category feature word / phrase. The minimum number of category feature words to be extracted may be extracted.

  Next, in step S5, the control unit 11 determines a sub-search formula (described later) based on the patent classification and the minimum number of category feature words extracted in step S4 for each minimum number of patent classifications extracted in step S2. (See formulas (1) to (3)).

  Next, in step S6, the control unit 11 obtains the union of the minimum number of feature words and phrases extracted in step S3 (see formula (4) described later), the union of the obtained minimum number of feature words and phrases, Based on the sub-search formula constructed in S5, the main search formula (see formula (5) described later) is constructed.

<Patent classification extraction process>
FIG. 3 is a flowchart of the above-described patent classification extraction process.

  First, in step S <b> 21, the control unit 11 acquires a patent classification in which a positive example is classified by EBL, and sets the acquired patent classification as a patent classification candidate. Patent classification is assigned to each patent document. Examples of this patent classification include IPC and FI. In the present embodiment, FI is used as the patent classification. As shown in FIG. 6, the FI is composed of a plurality of hierarchized categories, and the conceptual structure can be subdivided into a tree structure, so that it can be regarded as domain knowledge in EBL. For this reason, by setting an appropriate operability standard and inputting a small number of positive examples, it is possible to acquire an FI that is a target concept.

  For example, as an appropriate operability standard, subdivision to the first hierarchy of the subgroup is set, and as a small number of positive examples, Japanese Patent Laid-Open No. 07-119372 that discloses blinds is input, E06B / 24 in FIG. 6 can be acquired as the FI that is the target concept. Here, E06B / 24 refers to “a screen or other structure that can be protected against light, in particular sunlight; a similar screen for privacy or appearance”.

  Next, in step S22, the control unit 11 registers, for each patent classification candidate set in step S21, a positive example patent document belonging to the patent classification candidate as an element of the patent classification candidate. According to this, for each patent classification candidate set in step S21, a patent classification candidate and a positive patent document belonging to the patent classification candidate are linked and stored in the RAM of the control unit 11. It becomes.

  Next, in step S23, the control unit 11 sorts the patent classification candidates in descending order based on the number of positive patent documents registered in step S22. According to this, the patent classification candidate to which the patent document of the positive example belongs most often becomes the highest patent classification candidate.

  Next, in step S24, the control unit 11 employs the highest patent classification candidate among the patent classification candidates rearranged in step S23. According to this, the highest patent classification candidate is stored in the RAM of the control unit 11.

  Next, in step S25, the control unit 11 deletes the patent document of the positive example belonging to the highest patent classification candidate adopted in step S24 from the other patent classification candidates. According to this, for the positive patent document belonging to the highest patent classification candidate, the relationship linked to the other patent classification candidates is deleted from the RAM of the control unit 11.

  Next, in step S <b> 26, the control unit 11 excludes, from the patent classification candidates, patent classification candidates for which there is no positive example patent document to which they belong. According to this, the patent classification candidate that is not associated with the patent document of the positive example is not a patent classification candidate.

  Next, in step S27, the control unit 11 determines whether or not the number of other patent classification candidates is zero, in other words, whether or not the number of remaining patent classification candidates is “1”. . If the number of other patent classification candidates is zero, the process proceeds to step S28. If the number of other patent classification candidates is not zero, the process returns to step S23.

  Next, in step S28, the control unit 11 extracts all the patent classification candidates employed in step S24 as the minimum number of patent classifications.

  A specific example of extracting the minimum number of patent classifications by patent classification extraction processing will be described with reference to FIG. In FIG. 7, the patent classification candidates set in step S21 are represented as patent classification candidates A, B, C, D, and E. In addition, the positive patent documents read in step S1 are represented as patent documents 1, 2, 4, 5, 6, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19. The negative patent documents read in step S1 are represented as patent documents 3, 7, 8, and 20.

  FIG. 7A shows a stage where the process of step S22 is performed. At this stage, the patent documents 1, 2, 6, 11, 12, 16, 17 of the positive examples are registered as the elements of the patent classification candidate A, and the patent documents 1, 2 of the positive examples are registered as the elements of the patent classification candidate B. , 6 are registered, positive patent documents 12, 13, 14, 17, 18, 19 are registered as elements of patent classification candidate C, and positive patent documents 4, 5 are registered as elements of patent classification candidate D. , 9, 10, 14, and 15 are registered, and as patent classification candidate E elements, positive patent documents 5, 10, and 15 are registered.

  FIG. 7B shows a stage where the processes of steps S23 to S26 are performed from the stage of FIG. Specifically, first, in step S23, the patent classification candidates A to E are rearranged, and as a result, the patent classification candidates A, C, D, B, and E are arranged in this order. Next, in step S24, the patent classification candidate A is adopted as the highest patent classification candidate.

  Next, in step S25, the positive patent documents 1, 2, 6 belonging to the patent classification candidate A are deleted from the patent classification candidate B, and the positive patent documents 12, 17 belonging to the patent classification candidate A are patented. It will be deleted from the classification candidate C. According to this, there is no positive example patent document belonging to the patent classification candidate B.

  Next, in step S26, the patent classification candidate B is excluded from the patent classification candidates A to E. According to this, since the patent classification candidates C to E exist as other patent classification candidates described in step S27, the processes of steps S23 to S26 are performed again.

  (C) of FIG. 7 shows a stage where the processes of steps S23 to S26 are performed from the stage of (b) of FIG. Specifically, first, in step S23, the patent classification candidates C to E are rearranged, and as a result, the patent classification candidates D, C, and E are arranged in this order. Next, in step S24, the patent classification candidate D is adopted as the highest patent classification candidate.

  Next, in step S25, the positive patent document 14 belonging to the patent classification candidate D is deleted from the patent classification candidate C, and the positive patent documents 5, 10, and 15 belonging to the patent classification candidate D are patent classification candidates. E will be deleted from E. According to this, there is no positive example patent document belonging to the patent classification candidate E.

  Next, in step S26, the patent classification candidate E is excluded from the patent classification candidates C to E. According to this, since the number of other patent classification candidates described in step S27 is zero, in other words, the remaining patent classification candidate is only the patent classification candidate C, so the process of step S28 is performed.

  Next, in step S28, patent category candidates A, C, and D are extracted as the minimum number of patent categories. That is, the patent classification candidates A, C, and D include all of the positive patent documents among the patent documents 1 to 20.

<Feature word extraction processing>
FIG. 4 is a flowchart of the feature phrase extraction process described above.

  First, in step S31, the control unit 11 analyzes the name, abstract, and claim of the invention included in the patent document of the positive example by existing natural language processing or text mining, and characterizes the characteristic phrase of the patent document. To get. Then, similarly to step S21, the acquired feature word / phrase is set as a feature word / phrase candidate.

  Next, in step S32, as in step S22, the control unit 11 registers, for each feature word candidate set in step S31, a positive example patent document including the feature word candidate as an element of the feature word candidate. . According to this, for each feature word candidate set in step S31, the feature word candidate and the positive patent document including the feature word candidate are linked and stored in the RAM of the control unit 11. It becomes.

  Next, in step S33, similarly to step S23, the control unit 11 rearranges the feature word candidates in descending order based on the number of positive patent documents registered in step S32. According to this, among all the feature word candidates set in step S31, the feature word candidate included in the largest number of patent documents is the highest-order feature word candidate.

  Next, in step S34, as in step S24, the control unit 11 employs the highest-level feature word candidate among the feature word candidates rearranged in step S33. According to this, the highest-level feature word candidate is stored in the RAM of the control unit 11.

  Next, in step S35, similarly to step S25, the control unit 11 deletes the positive patent document including the highest-level feature word candidate adopted in step S34 from other feature word candidates. According to this, for the positive example patent document including the highest-level feature phrase candidate, the relationship associated with the other feature phrase candidates is deleted from the RAM of the control unit 11.

  Next, in step S 36, as in step S 26, the control unit 11 excludes the feature word / phrase candidate from which the included patent document of the example is missing from the feature word / phrase candidate. According to this, the feature word / phrase candidate not associated with the patent document of the positive example is not the feature word / phrase candidate.

  Next, in step S37, as in step S27, the control unit 11 determines whether or not the number of other feature word candidates is zero, in other words, the number of remaining feature word candidates is “1”. It is determined whether or not. If the number of other feature word candidates is zero, the process proceeds to step S38. If the number of other feature word candidates is not zero, the process returns to step S33.

  Next, in step S38, as in step S28, the control unit 11 extracts all the feature word candidates adopted in step S34 as the minimum number of feature words.

<Category feature word extraction processing>
FIG. 5 is a flowchart of the above-described category feature word extraction process.

  First, in step S41, as in step S31, the control unit 11 analyzes the name, summary, and claim of the invention included in the patent document of the positive example by using existing natural language processing and text mining. Get feature words that characterize the contents of. Then, using the acquired feature word / phrase, for each minimum number of patent classifications extracted in step S28, the feature word / phrase included in the positive example classified in the patent classification is classified into the patent classification. Feature phrases that are not included in the negative example are acquired as category feature phrases. Then, similarly to steps S21 and S31, the acquired category feature phrase is set as a category feature phrase candidate.

  In SBL, a concept space (version space) that solves a problem of constructing a conditional expression expressed by AND and OR of characteristic words and phrases included in a positive example so as to exclude a negative example and develops a conditional expression as a candidate. ) To set an appropriate degree of freedom (bias) and determine the conditional expression. This SBL will be described in detail with reference to FIG.

  FIG. 8 shows a case where the bias is a product set of two conditions when the concept of “blind” is acquired by SBL. In this case, the lowest layer of a single condition is excluded. Further, the most general concepts {color, awning} and {*, curtain} including negative examples are excluded. Then, after each positive example of red and white is processed, the final concept {color, blind} is acquired as a concept.

  Next, in step S42, as in steps S22 and S32, for each category feature phrase candidate set in step S41, the control unit 11 includes the category feature phrase candidate as an element of the category feature phrase candidate. Register patent documents. According to this, for each category feature phrase candidate set in step S41, the category feature phrase candidate and the positive patent document including the category feature phrase candidate are linked and stored in the RAM of the control unit 11. Will be.

  Next, in step S43, as in steps S23 and S33, the control unit 11 rearranges the category feature word / phrase candidates in descending order based on the number of positive patent documents registered in step S42. According to this, among all the category feature word candidates set in step S41, the category feature word candidate included in the most number of positive patent documents becomes the highest category feature word candidate.

  Next, in step S44, as in steps S24 and S34, the control unit 11 employs the highest-level feature word candidate among the feature word candidates rearranged in step S43. According to this, the highest category feature word candidate is stored in the RAM of the control unit 11.

  Next, in step S45, as in steps S25 and S35, the control unit 11 deletes the positive patent document including the highest-level feature word candidate adopted in step S44 from the other feature word candidates. According to this, for the patent document of the positive example including the highest category feature word / phrase candidate, the relationship associated with other category feature word / phrase candidates is deleted from the RAM of the control unit 11.

  Next, in step S46, as in steps S26 and S36, the control unit 11 excludes the category feature phrase candidates for which the included patent document of the positive example is lost from the category feature phrase candidates. According to this, the category feature word / phrase candidate that is not associated with the patent document of the positive example is not a category feature word / phrase candidate.

  Next, in step S47, as in steps S27 and S37, the control unit 11 determines whether the number of other category feature phrase candidates is zero, in other words, the number of remaining category feature phrase candidates is “ It is determined whether or not “1”. If the number of other category feature phrase candidates is zero, the process proceeds to step S48, and if the number of other category feature phrase candidates is not zero, the process returns to step S43.

  Next, in step S48, as in steps S28 and S38, the control unit 11 extracts all the category feature phrase candidates adopted in step S44 as the minimum number of category feature phrases.

<Example of building this search expression>
An example when the search formula is constructed by the above document search processing will be described below with reference to FIGS. Specifically, as a search target, a patent for controlling three elements of lighting, illumination, and air conditioning for countermeasures against room temperature rise by blind lighting shown in FIG. 9 is set, and a book for searching this patent is set. An example in which a search expression is constructed by the above-described document search process will be described.

  First, the preliminary search described in step S1 of FIG. 2 was performed. Specifically, when searching for patent documents related to the above-mentioned patents from the patent database using the search conditions and search formulas shown in FIG. 10, it was found that 30 patent documents corresponded. And it has confirmed that 12 patent documents among these 30 patent documents are the patent documents regarding the above-mentioned patent, ie, the patent document of a positive example. In other words, 12 positive examples and 18 negative examples are extracted from the patent database by the preliminary search.

  Next, the search formula was constructed by the document search process described above. Specifically, each process shown in FIGS. 2 to 5 was performed using 12 positive examples extracted by the above-described preliminary search and 18 negative examples. As a result, the following expressions (1), (2), and (3) were constructed as sub-search expressions, and the following expression (4) was constructed as the union of the minimum number of feature words. And based on Formula (1)-(4), the following formula | equation (5) was constructed | assembled as this search formula.

  According to the equations (1) to (3), as the minimum number of patent classifications, “F24F11 / 02” related to air conditioning control, “H04Q9 / 00” related to remote control, “E06B9 / 264” related to blinds, It can be seen that the following three were extracted. Also, as the minimum number of category feature words corresponding to “F24F11 / 02”, three “lighting equipment”, “air conditioner”, and “lighting fixture” are extracted, and the minimum number corresponding to “H04Q9 / 00” is extracted. One of “wireless communication” is extracted as a category feature phrase, and “carbon dioxide”, “electric blind”, “solar radiation state acquisition unit”, “linked” are selected as the minimum number of category feature phrases corresponding to “E06B9 / 264”. It can be seen that four of “control” have been extracted.

  In addition, according to the equation (4), the minimum number of feature words are “lighting fixture”, “environmental control”, “electric blind”, “low cost”, “solar radiation state acquisition unit”, “interlocking control” 6 It can be seen that one has been extracted.

  FIG. 11 is a schematic diagram showing the relationship between the sub-search expression, the union of feature words, and the positive and negative examples. In FIG. 11, the above-mentioned 12 positive patent documents are represented as Patent Documents 3, 6, 9, 10, 13, 15, 17, 18, 22, 25, 26, and 27, and the 18 negative patent documents described above. The patent document of an example shall be represented as patent document 1, 2, 4, 5, 7, 8, 11, 12, 14, 16, 19, 20, 21, 23, 24, 28, 29, 30.

  Satisfying the union expression of the feature words represented by the expression (4) is that 12 positive patent documents (Patent Documents 3, 6, 9, 25, 18, 13, 26, 17, 22, 27) , 15, 10) and four negative patent documents (Patent Documents 29, 23, 2, 30). For this reason, according to the union of the feature terms expressed by the expression (4), negative patent documents can be extracted, but all of the above 12 positive patent documents can be extracted. .

  “F24F11 / 02” in Formula (1) belongs to six positive patent documents (Patent Documents 3, 6, 25, 18, 17, and 15) and one negative patent document (Patent Documents) Reference 23). Here, “lighting equipment”, “air conditioner”, and “lighting fixture” in Expression (1) are category feature phrases in “F24F11 / 02”, and any of these three category feature phrases is described above. These are feature words that are included in the six positive patent documents but not included in the one negative patent document described above. For this reason, regarding the sub-search formula of Formula (1) constructed with “F24F11 / 02” and the above-described three category feature words, the above-mentioned six positive patent documents satisfy this sub-search formula. The above-mentioned negative patent document does not satisfy this sub-search formula. As described above, according to the sub-search expression represented by Expression (1), the above-mentioned six positive patent documents can be extracted from the patent documents belonging to “F24F11 / 02”.

  As with the sub-search expression of the above formula (1), the two sub-search formulas of the formula (2) are selected from two patent documents belonging to “H04Q9 / 00” (Patent Document 15). 10) can be extracted.

  Also, the sub-search formula of the formula (3) is similar to the sub-search formulas of the formulas (1) and (2) described above, from the patent documents belonging to “E06B9 / 264”. Patent documents (Patent documents 9, 25, 13, 26, 22, and 27) can be extracted.

  From the above, according to the present search formula expressed by the formula (5), which is constructed by the formulas (1) to (4), the above-mentioned 12 cases are selected from the 30 patent documents found by the preliminary search. All of the positive patent documents can be extracted. For this reason, this search formula represented by Formula (5) is a search formula suitable for the purpose of extracting all the positive patent documents from the 30 patent documents found by the preliminary search. You can think of it.

  Here, when the patent documents related to the above-mentioned patents were searched from the patent database using the search formula represented by the above-described formula (5), it was found that 212 patent documents corresponded.

<Dynamic screening method>
Then, next, the above-mentioned 212 patent documents were searched by the dynamic screening method shown in FIG. Normally, in the screening method of patent documents (population) extracted by a search formula or the like, confirmation of a subset to which an appropriate condition for the population is applied, such as a narrowing search by words, is repeated.

  However, here, in order to clarify the effectiveness of the above-described document search processing, from the state in which unconfirmed patent documents are arranged in order of filing date, following the above-mentioned 12 positive patent documents that have been confirmed. Sequentially, 212 patent documents were confirmed. Specifically, in the middle of repeating confirmation of unconfirmed patent documents (step S103), the main search formula is sequentially reconstructed by document search processing (step S104), and in descending order of suitability with the main search formula. Unidentified patent documents are rearranged (steps S101 and S102).

  More specifically, first, in step S101, the document search device 1 determines the degree of conformity with the search formula for each of the 212 patent documents described above. However, for a confirmed patent document, that is, a patent document used as a positive example or a negative example, the degree of conformity with the search formula is not determined. For this degree of matching, for each patent document, the evaluation value of each sub-search expression is obtained by the following three methods, and the largest of the evaluation values of each sub-search expression is set as the degree of matching with this search expression. The evaluation value of each sub-search expression may be obtained by any one of the three methods or may be obtained in combination. When the evaluation value could not be obtained by any of the following three methods, the evaluation value was set to zero.

  In the first method, when the patent classification of the patent document matches the patent classification of the sub-search formula, the evaluation value of the sub-search formula can be obtained. Specifically, a value obtained by dividing the number of confirmed positive examples belonging to the sub-search expression by the number of all confirmed positive examples is set as the evaluation value of the sub-search expression. Further, if the patent document includes the category feature word / phrase included in the sub-search expression when they match, the set evaluation value is multiplied by the first value. Here, as the first value, for example, a fixed value such as twice may be set, or it may be set according to the importance of the category feature phrase, or a category feature phrase not included in the patent document You may set according to the number of.

  In the second method, when the patent classification of the patent document does not match the patent classification of the sub-search expression, and the patent document includes the category feature word / phrase included in the sub-search expression, the sub-search expression Can be obtained. Specifically, the second value is set as the evaluation value of the sub search expression. Here, as the second value, for example, a value obtained by multiplying the value obtained by dividing the number of confirmed positive examples belonging to the sub-search expression by the number of all confirmed positive examples is employed. be able to.

  In the third method, when the patent document includes a feature word included in the union of the minimum number of feature words, the evaluation value of the sub search expression can be obtained. Specifically, the third value is set as the evaluation value of the sub search expression. Here, as the third value, for example, a fixed value such as “0.5” may be set, or may be set according to the number of feature words included in the patent document.

  Next, in step S102, the document search apparatus 1 rearranges unconfirmed patent documents based on the degree of matching obtained in step S101. Then, the confirmed patent documents are arranged at the top, and the rearranged unconfirmed patent documents are arranged below the confirmed patent documents.

  Next, in step S103, the user confirmed whether the patent documents rearranged in step S102 are positive examples or negative examples in order from the top. The number of cases to be confirmed is arbitrary, and it is not necessary to confirm all unconfirmed patent documents.

  Next, in step S104, the document retrieval apparatus 1 performs the above document retrieval based on the confirmation result in step S103, that is, the patent document newly classified as a positive example or a negative example and the confirmed patent document. Processed and rebuilt this query.

  Here, before the dynamic screening method shown in FIG. 12 is performed, all of the above-mentioned 212 patent documents are confirmed in advance, and in addition to the above-mentioned 12, three new patent documents are positive examples. It has been found. Therefore, hereinafter, the description will be given focusing on these three positive patent documents.

  When the above-mentioned 212 patent documents are rearranged in order of the filing date, the appearance ranks of the above-mentioned three positive patent documents are 139, 140, and 160, respectively.

  Therefore, the above-described dynamic screening technique was performed using the retrieval formula constructed by the above-described formulas (1) to (4). That is, the first dynamic screening technique was performed, and 200 unidentified patent documents out of the 212 cases described above were rearranged. As a result, the appearance ranks of the above three positive patent documents were 55th, 57th and 71st, respectively.

  In the first dynamic screening method, in step S103, eight patent documents out of 200 unconfirmed were confirmed. All of these eight patent documents were negative patent documents. In step S104, the search formula is reconstructed with 12 positive examples and 26 negative cases (18 + 8 cases).

  Therefore, the above-described dynamic screening technique was performed again using the retrieval formula reconstructed in the first dynamic screening technique. That is, the second dynamic screening technique was performed, and 192 patent documents that were not confirmed among the 200 cases described above were rearranged. As a result, the appearance ranks of the above-mentioned three positive patent documents were 27th, 30th and 71st, respectively.

  In the second dynamic screening method, 20 patent documents out of 192 unconfirmed were confirmed in step S103. Of these 20 patent documents, 2 were positive patent documents and the remaining 18 were negative patent documents. In step S104, the search formula is reconstructed with 14 positive cases (12 cases + 2 cases) and 44 negative cases (26 cases + 18 cases).

  Therefore, the above-described dynamic screening method was performed again using the retrieval formula reconstructed in the second dynamic screening method. That is, the third dynamic screening method was performed, and 172 patent documents that were not confirmed among the above 192 cases were rearranged. As a result, among the above-mentioned three positive patent documents, the appearance order of the patent document whose previous appearance order was 71 was 43rd.

  According to the above, the search formula is constructed, the patent documents are rearranged according to the degree of conformity with the search formula, the patent documents are confirmed in the sorted order, and the feedback of the confirmation results to the construction of the search formula By repeating the above, it can be seen that the appearance rank of the patent document of the positive example increases.

  According to the above document search device 1, the patent database in which the patent documents classified by the patent classification are accumulated, the patent classification in which the positive examples are classified is obtained, and the patent documents belonging to the obtained patent classifications are selected. Find feature words to exclude negative examples. Therefore, even if there is a specific negative example, this search formula that can select significant feature words and exclude negative examples can be constructed, and an appropriate search for searching patent documents that match the search purpose An expression can be constructed.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes a design that does not depart from the gist of the present invention.

  For example, F terms may be used instead of feature words. When the F term is used, the analysis range often extends not only to the claims but also to the whole sentence, and the accuracy of the search expression may be improved as compared with the use of feature words.

  In the above-described embodiment, the document search apparatus 1 searches for patent documents, but the present invention is not limited to this. This can be used when documents included in the document database are classified into a plurality of hierarchized categories and a document matching the search purpose is searched from the document database.

  In the above-described embodiment, the patent database is stored in the storage unit 13 of the document search apparatus 1. However, the invention is not limited to this. For example, the internal database of a server provided to be communicable with the document search apparatus 1. It is good also as what is built in.

  In the above-described embodiment, the FI is used as the patent classification. However, the invention is not limited to this, and it may be a plurality of hierarchized categories that can classify documents, such as IPC, ECLA, USC, and theme code.

DESCRIPTION OF SYMBOLS 1 Document retrieval apparatus 11 Control part 12 Communication part 13 Storage part 14 Input I / F
15 bus 16 input device 17 output device

Claims (9)

A document search method for searching for a document matching a search purpose from a document database classified into a plurality of hierarchical categories,
A first step of obtaining a category into which positive examples are classified among a plurality of documents included in the document database;
A second step of acquiring a feature phrase included in the positive example;
A feature word / phrase included in the positive example classified into the category for each category acquired in the first step, and not included in the negative example classified into the category among the plurality of documents. A third step of obtaining a phrase as a category feature phrase;
For each category acquired in the first step, a fourth step of constructing a sub-search expression based on the category and the category feature phrase acquired in the third step;
A fifth step of constructing this search expression based on the union of the feature words acquired in the second step and the sub-search expression constructed in the fourth step;
A document retrieval method comprising: In the first step, after obtaining a category into which the positive example is classified, a minimum number of categories including all documents belonging to the positive example are extracted from the category;
In the second step, after obtaining the feature words / phrases included in the positive example, a minimum number of feature words / phrases that include all documents belonging to the positive example are extracted from the feature words / phrases,
In the third step, for each minimum number of categories extracted in the first step, the category feature words / phrases are obtained for all documents belonging to positive examples classified in the category, and then the category feature words / phrases are obtained. Extract a minimum number of category feature words that include all documents belonging to the positive example from
In the fourth step, for each minimum number of categories extracted in the first step, a sub-search expression based on the category and the minimum number of category feature phrases extracted in the third step Build
In the fifth step, the main search formula is constructed based on the union of the minimum number of feature words extracted in the second step and the sub-search formula constructed in the fourth step. The document search method according to claim 1. The first step includes
A first setting step of acquiring a category into which the positive example is classified and setting the category as a category candidate;
A first alignment step of registering a positive example document belonging to the category candidate as an element of the category candidate and rearranging the category candidate in descending order based on the number of registered positive example documents;
The highest category candidate is adopted, the positive example document belonging to the highest category candidate is deleted from the elements of the other category candidates, and the category candidate having no elements is set in the first setting step. A first hiring step to exclude from category candidates;
The first alignment step and the first adoption step are repeated until the number of other category candidates becomes 0, and the category candidates adopted in the first adoption step are extracted as the minimum number of categories. A first extraction step to:
With
The second step includes
A second setting step of acquiring a feature phrase included in the positive example and setting the feature phrase as a feature phrase candidate;
A second alignment step of registering a positive example document including the characteristic word candidate as an element of the characteristic word candidate and rearranging the characteristic word candidate in descending order based on the number of registered positive example documents;
Adopting the highest-level feature phrase candidate, deleting the original example document including the highest-level feature phrase candidate from the other feature-phrase candidate elements, and setting the feature-phrase candidate with no elements in the second setting step A second adopting step of excluding from the feature word / phrase candidates set in
The second alignment step and the second adoption step are repeated until the number of other feature word candidates becomes zero, and the feature word candidates adopted in the second adoption step are determined as the minimum number of features. A second extraction step for extracting as a phrase;
With
The third step includes
A third setting step of acquiring the category feature phrase for each minimum number of categories extracted in the first step and setting the category feature phrase as a category feature phrase candidate;
A third alignment step of registering a positive example document including the category characteristic word candidate as an element of the category characteristic word candidate and rearranging the category characteristic word in descending order based on the number of registered positive example documents When,
Adopting the highest category feature phrase candidate, deleting the positive example document including the highest category feature phrase candidate from the elements of other category feature phrase candidates, and removing the category feature phrase candidate with no elements. A third adopting step of excluding from the category feature word / phrase candidates set in the setting step of 3;
The third aligning step and the third adopting step are repeated until the number of the other category feature phrase candidates becomes 0, and the category feature phrase candidates adopted in the third adopting step are determined as the minimum number. A third extraction step for extracting as a category feature word of
The document search method according to claim 2, further comprising:   Of the documents searched from the document database based on the main search formula, those excluding the document used when constructing the main search formula are sorted in descending order based on the degree of matching with the main search formula. 4. The document search method according to claim 1, further comprising a sixth step of rearranging.   5. The document search method according to claim 1, wherein the category feature word / phrase is not included in the feature word / phrase acquired in the second step.   In the third step, each of the categories acquired in the first step is a feature word / phrase included in a positive example classified into the category, and the negative phrase classified into the category among the plurality of documents. The feature word / phrase that is an example and is not included in the negative example including the feature word / phrase acquired in the second step is acquired as a category feature word / phrase. Document search method. The document is a patent document;
The document search method according to claim 1, wherein the category is any one of FI, IPC, ECLA, USC, or a theme code. The document is Japanese patent literature,
8. The document search method according to claim 1, wherein an F-term is used instead of the feature word / phrase. A program for causing a computer to execute a document search method for searching for a document that matches a search purpose from a document database classified into a plurality of hierarchical categories,
A first step of obtaining a category into which positive examples are classified among a plurality of documents included in the document database;
A second step of acquiring a feature phrase included in the positive example;
A feature word / phrase included in the positive example classified into the category for each category acquired in the first step, and not included in the negative example classified into the category among the plurality of documents. A third step of obtaining a phrase as a category feature phrase;
For each category acquired in the first step, a fourth step of constructing a sub-search expression based on the category and the category feature phrase acquired in the third step;
A fifth step of constructing this search expression based on the union of the feature words acquired in the second step and the sub-search expression constructed in the fourth step;
A program that causes a computer to execute.