JP2010118050A - System and method for automatically searching patent literature - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a system and method for automatically searching patent literatures by automatically searching relevant patents based on a small quantity of searched data and extracting a desired patent. <P>SOLUTION: The system and method for automatically searching patent literatures include: a relevant patent database which stores searched patents belonging to a technical field related to a search output as samples; a non-relevant patent database which stores searched patents belonging to technical fields other than that related to the search output; a morphological analysis means which performs morphological analysis to the patent literature data stored in the databases; a statistic information calculation means which determines a weight for the morphologically analyzed data from word appearance frequency and entropy to select a characteristic word; a discriminator which mechanically learns the characteristic word to discriminate patent literatures; and an extraction means which extracts all patent literatures belonging to the technical field related to the search output from the discrimination results of the discriminator. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

The present invention relates to a patent document automatic search system and a patent document automatic search method.

  There is an enormous amount of patent information, and there are multiple databases that provide patent information. For this reason, conventionally, various techniques for searching for patent information and technical information have been developed.

  For example, Patent Document 1 is a technique for performing a search by selecting the most appropriate document content using international patent classification.

  In addition, Patent Document 2 uses past research knowledge to create a logical operation search expression that appropriately combines search terms of technical terms, IPC, FI, F-term, applicant, etc. when conducting a search. As a result, high-accuracy surveys can be conducted efficiently.

  However, Patent Document 1 and Patent Document 2 cannot finally obtain a patent group desired by a user unless all patents obtained as search results are confirmed.

In particular, when searching for patents in a wide range of technical elements such as the environmental field and extracting desired patents, the user must read a large number of searched patents and determine whether they are related. Effort is required.
JP 2000-322447 A JP 2007-24204 A

  In view of the above prior art, the problem to be solved by the present invention is an automatic patent document search system and a patent document automatic search method for automatically searching related patents based on a small amount of searched data and extracting desired patents. Is to provide.

  An automatic search system for patent document data, in which related patent databases that store searched patents belonging to the technical field related to search output are sampled, and searched patents that belong outside the technical field related to search output are sampled A morpheme analysis unit that performs morphological analysis on patent document data stored in the related patent database and the unrelated patent database, and an appearance frequency and entropy for the morpheme analyzed data Patent discriminating unit for discriminating patent documents belonging to the technical field related to search output by machine learning using feature words selected using calculated weights, and technical field related to search output from the discrimination result of the patent discriminating unit Patent document characterized by comprising: an extraction unit that extracts all patent document data belonging to Consisting of dynamic search system.

  In particular, in the patent identification unit, for all words appearing in the positive example and the negative example, the weight is obtained from the appearance frequency and entropy, and if the weight in the positive example of the word is greater than twice the weight in the negative example of the word, It is characterized by using words as feature words as features of machine learning.

  An automatic search method for patent document data, a means for storing, as a sample, a searched patent belonging to a technical field related to search output in a related patent database, and a searched patent belonging to a field other than the technical field related to search output Means for storing in a non-related patent database as a sample, means for performing morphological analysis on patent document data stored in the related patent database and the non-related patent database, and an appearance frequency for the data subjected to morphological analysis A means for discriminating patent documents belonging to the technical field related to the search output by machine learning using the feature word selected using the weight calculated from the entropy as a feature, and a technical field related to the search output from the result of the identification means Means for extracting all patent document data, and an automatic patent document search method comprising: That.

In particular, the means for discriminating patent documents obtains weights from the appearance frequency and entropy for all words appearing in the positive and negative examples, and the weights in the positive examples of the words are greater than twice the weights in the negative examples of the words. For example, the word is used as a feature word for machine learning as a feature word.

According to the above-described system and method of the present invention, it is possible to automatically search and extract a patent document desired by a user based on a small amount of sample data that has been investigated in advance, thereby greatly reducing the work time required for patent search. .

  The patent document automatic search system according to the present invention includes a related patent database and an unrelated patent database that store sample patents that have already been investigated, a morpheme analysis unit that performs morpheme analysis, and a bias of words that appear in the accumulated patent document data. A statistic calculation unit for calculating weights from entropy, a patent identification unit for machine learning based on the words, and identifying and extracting related patents by a generated classifier, and a patent document based on the results of the patent identification unit It consists of the extraction part which extracts.

  FIG. 1 shows a procedure, that is, a flowchart of the patent document automatic search system according to the present invention. The operation will be described below with reference to FIG.

  Step 1: The same number of sample data as teacher data is selected from the related patent database section and the unrelated patent database section. Hereinafter, a sample belonging to the related patent database part is defined as a positive example, and a sample belonging to the unrelated patent database part is defined as a negative example.

  Step 2: The morpheme analysis unit performs morpheme analysis for each selected positive example and negative example data. Morphological analysis means dividing a sentence into words. For example, when the expression “the present invention is related to recycling of a refrigerator” appears, this expression is expressed as “the present invention //// refrigerator / of / recycled / related / related / thing. It is divided into word units. Here, / represents a delimiter.

  Step 3: Machine learning is performed by the patent discriminating unit using the appearance probability in the document as a learning feature, using words that are included only in positive examples extracted and calculated by the statistic calculation unit, that is, feature words. As a method of machine learning, there is a Support Vector Machine (hereinafter abbreviated as SVM), but any method capable of binary classification may be used. In the present invention, only the statistic of a word is used as a feature, so that it does not depend on a specific language. Therefore, patent documents written in all languages as well as Japanese can be targeted.

  The present invention is also characterized in that only word information called feature words is used. In search systems in other fields such as Web search, in addition to words, information unique to the document, for example, in Web search, tag structure, link structure, etc. are used at the time of search. There was no equivalent to the link structure, and it was necessary to search and extract only word information.

  Therefore, in a search system for patent documents, search systems in other fields such as the Web cannot be used as they are, and it is required to create an evaluation formula based on word information. Realize evaluation based on word information.

  As feature words, words that do not appear frequently in the patent group not related to the patent technology related to the search output but appear frequently in the patent group related to the technology are extracted. For example, if the patent technology desired by the user belongs to a group of patents related to recycling, “recycle”, “reuse”, etc. are selected.

  The feature word is extracted as follows. First, weighting is performed on the content words (hereinafter referred to as words) included in the positive example, and words whose weights are in the upper half are extracted. The weighting factor of the positive example is expressed by Equation 1.

Here, P (t _i , S _p ) and H (t _i , S _p ) on the right side are expressed by _Equation 2 and Equation 3, respectively.

Here _{P (t} i, _{S p)} refers to the probability of occurrence of the word _{t i} in the document set _{S p} of the positive cases.

H (t i, _{S p)} represents the entropy based on occurrence probabilities of words t _i of each document included in the document set S _p of the positive sample, the more entropy is high word, uniformly distributed document set positive cases It is a word.

  The reason for introducing this index is that words that appear dispersed in many documents in a positive document set are more likely to appear in documents compared to words that appear concentrated in a small number of documents. This is because it represents the characteristics of the set and is also effective as a feature.

  This method using entropy is optimal for the patent document automatic extraction task. This is because patent documents often use fixed technical terms and fixed expressions, such as “the present invention is”, and there are few individual differences in vocabulary compared to Web information using colloquial expressions. It is. If there are individual differences in vocabulary, even words with the same meaning are morphologically analyzed with different words, and as a result, there is a high possibility that correct entropy will not be obtained, but such problems do not exist in the patent literature. .

  Next, as in the case of the positive example, the words included in the negative example are weighted by using the following formula 5, and the word whose weight is the upper half is extracted.

However, _Sn is a document set belonging to a negative example in the training data. Then, the weight _{W p (t i, S p} ) in the positive example of a word _{t i} weight _{W n (t i, S p} ) is in the negative examples is greater than 2 times of selecting the word _{t i} as features . That is, a word t _i that satisfies the following formula 6 is selected as a feature.

  The reason for using the condition of Equation 6 will be described. With the weight expressed by Equation 1, a high weight is given even to a word that is not related to a related patent if it is a general word. However, such words can be given high weight even in negative examples. For example, the words “present invention” and “patent” are present in almost all patent documents regardless of positive examples and negative examples, and are inappropriate as feature words representing positive examples.

Here, the reason why W _p (t _i , S _p ) is twice as large as W _n (t _i , S _p ) as a criterion for feature word selection will be described with reference to the graph of FIG. In FIG. 2, the horizontal axis represents the weight and F-SCORE is plotted on the vertical axis. F-SCORE is an index indicating overall accuracy calculated from accuracy and recall. From this, it can be seen that F-SCORE peaks when the weight is about twice, and is an optimal numerical value in terms of accuracy. The reason why such an optimal value exists is that there is a trade-off relationship between the accuracy that means the accuracy of the extraction example and the recall that means the completeness of the extraction example.

Therefore, the weight W _p (t _i , S _p ) in the positive example for a certain word t _i is compared with the weight W _n (t _i , S _n ) in the negative example, and W _p (t _i , S _p ) is more _{W n (t i, S n} ) by selecting the larger word than twice the prevents common words are selected as features. That is, words that appear in general patent documents such as “present invention” and “patent” are not characteristic words.

By simply using the statistic Tf * Idf, which is often used in search systems, it is not possible to prevent a general word from being selected as a feature. Not selecting is one of the features of the present invention. In Tf * Idf, it is easy to extract an expression unique to a sentence as a feature. Therefore, feature words that are biased toward positive examples in general are evaluated low and may not be extracted. In addition, there is a problem that it is easy to extract special words that are frequently repeated in a small set of sentences. On the other hand, the weights of the present invention can extract words that are biased toward only positive examples as feature words, and are optimal for patent extraction.

  EXAMPLES Although this invention is demonstrated still in detail based on an Example, this invention is not limited only to these Examples.

  As shown in FIG. 3, the embodiment of the present invention includes a search input unit for inputting a search question, a patent database in which patents to be searched are stored, and a related patent database in which sample patents already searched are stored. , A non-related patent database that does not belong to the technical field desired by the user and stores already searched sample patents, and words that appear in the patent data stored in each of the related patent database and the non-related patent database Patent identification unit for identifying patent document data belonging to the technical field from bias, search result database for storing patent document data determined by the patent identification unit as a patent in the technical field desired by the user, and saving in the search result database It comprises a display unit for displaying the patent document data. FIG. 4 is a diagram illustrating a processing procedure and a concept according to the embodiment.

  The search input unit shown in FIG. 3 is a part in which a user inputs a search question and searches patent documents in order to collect samples of related patents and unrelated patents.

  Further, in the related patent database shown in FIG. 3, a small number of sample data of the patent group desired by the user is accumulated. This process corresponds to STEP1 in FIG. The format of the accumulated data may be any data including claims and specification parts, and is in a format such as a text format, a PDF (Portable Document Format) format, an XML (Extensible Markup Language) format based on a predetermined specification, etc. It doesn't matter. An example is shown in FIG.

  In the unrelated patent database shown in FIG. 3, a small number of sample data of the technical field desired by the user and the unrelated patent group are accumulated. The format of the accumulated data is not limited as long as it includes data including claims and specification parts, and may be in any format such as a text format, a PDF format, or an XML format based on a predetermined specification. An example is shown in FIG.

  First, the patent identification unit shown in FIG. 3 applies the unrelated patent database storing the already-searched patents, the related patent database, and the unrelated patent database to the patent data stored in FIG. Perform STEP2 and parse into words. Next, calculate the bias of the words that appear in each of STEP3, STEP4, and STEP5 in FIG. 1 (corresponding to the processing result of STEP3 in FIG. 4), and select the feature word. select. Then, machine learning is performed using the feature word as a feature as in STEP 6 in FIG. As a result, related patents are extracted based on the generated discriminator. For example, as shown in FIG. 4, machine learning is performed by SVM, and related patents are identified and extracted by a classifier generated thereby.

  The flow of the present invention will be described with reference to FIG. 1 and FIG. 4 using results obtained when the present invention is applied to “extracting patents related to recycling among technologies in the manufacturing process of refrigerators”.

  First, from the search input unit, a user uses a computer to input a query for a search keyword or IPC classification, and searches and collects sample patents from the patent database in FIG. This process corresponds to STEP1 in FIG. As a result, the patent document determined as a positive example is stored in the related patent database in FIG. 3, and the patent document determined as a negative example is stored in the unrelated patent database in FIG. The related patent database in FIG. 3 and the unrelated patent database in FIG. 3 are composed of an external storage device (storage or the like) that is directly connected to the terminal computer via a network. The format of the accumulated data is not limited to a format such as PDF, and any format can be used as long as it can describe a patent document. In this example, “Patents related to recycling among technologies in the manufacturing process of refrigerators” and “Patents related to recycling equipment” as shown in FIG. "Apparatus and method patents related to door manufacturing" apply. Negative examples include patents not related to recycling, such as “apparatus patent relating to maintaining the freshness of refrigerators” and “patent relating to deodorizing equipment” as shown in FIG. Here, sample patents, that is, teacher data, are collected for 76 positive and negative cases for 2003 and 2006 published patents.

  Next, a patent document judged as a positive example is extracted from the related patent database and the unrelated patent database, and morphological analysis is performed. This process corresponds to STEP 2 in FIG. Thereby, the patent document is divided into words. Specifically, as illustrated in FIG. 4, when an expression “the present invention relates to recycling of a refrigerator” appears in a patent document, the expression “the present invention / , / Refrigerator / of / recycle / to / relevant / related / thing / is / is ”. Here, / represents a break. The morpheme analysis is performed by the central processing unit in FIG. 3, and the analysis result is accumulated in the patent identification unit in the main storage device in FIG.

  A morphological analysis selects words that are characteristic of positive examples from the divided words. This processing corresponds to the processing of STEP3, STEP4, and STEP5 in FIG. The feature word is selected by using the above equation (6). In this embodiment, as shown in FIG. 4, {“recycle”, “discard”, “reuse”... The feature word selection is performed by the central processing unit in FIG. 3, and the result is accumulated in the patent identification unit in the main storage device in FIG.

  Using feature words as feature vectors, that is, bases for learning vectors, feature vectors are generated based on their appearance frequencies. For example, in a certain patent document, if “recycle” occurs 3 times, “dispose” 1 time, “reuse” 1 time, and other feature words appear 0 times, the feature vector is (3, 1 , 1, 0, ..., 0). The feature vector generation is performed by the central processing unit in FIG. 3, and the result is accumulated in the patent identifying unit in the main storage device in FIG.

  Note that the entropy calculation time in the statistic calculation depends on the hardware configuration of the computer, but when using public patent information for 10 years, it is only a few minutes and is sufficient for searching a large amount of patent document data. Can be used.

  Thereafter, as in STEP 6 in FIG. 1, machine learning is performed using the feature vector. As a method of machine learning, SVM is used this time as illustrated in FIG. SVM is an optimal technique for binary classification because it prevents overlearning by a technique called margin maximization. Machine learning by SVM is performed by a central processing unit, and the results are accumulated in the main memory.

  After machine learning, as shown in FIG. 4, patents determined to be positive examples by learning are accumulated in the search result database. The accumulated results are displayed on the search result display section. The display is performed in a list format such as an application number. The search result database is composed of an external storage device (storage or the like) that is directly connected to the terminal computer via a network. The format of the stored data is not limited to a format such as PDF, and any format can be used as long as it can describe a patent document.

  For the examples, search and extraction performance were evaluated. The number of patents used in the evaluation is 75 (31 positive examples, 44 negative examples). Moreover, SVM was used as a method of machine learning. In particular, the kernel uses a linear kernel.

  The accuracy rate in Table 1 is obtained by dividing the number of positive examples and examples determined by the system of the present invention that were truly positive examples by the number of positive examples and the example determined by the system of the present invention. The accuracy of.

  The recall shown in Table 1 is the total number of positive examples divided by the total number of positive examples among the positive examples and examples determined by the system of the present invention, and indicates the completeness of the positive examples.

  From this result, it can be seen that there is only about 10% of waste and that extraction can be performed with high accuracy. This means that subsequent screening is rarely necessary. In other words, when it is used roughly as statistical data, the time required for screening can be reduced. Specifically, collecting a sample patent took a week to tune keywords and carefully read patent documents, but in the present invention, it takes only a few minutes depending on the type of computer used. Collection of patent documents is completed.

  In this way, when patent data from 1993 to 2003 was extracted, 160 cases were found. In addition, in the keyword search type tuning based on thought and error (which is the norridge of Patent Document 2), 231 cases were searched. When evaluated from the relevance rate and the recall rate, the number of useless patents estimated by keyword search formula tuning is estimated to be about 70 (30% of the total). On the other hand, in the present invention, there are 10 useless patent documents (search errors). Degree.

  From the results, it can be seen that even with a very small number of samples, related patents are automatically extracted with good accuracy and recall. That is, the patent documents desired by the user are mechanically extracted without much effort.

  The search result database shown in FIG. 3 stores patent data determined as patents in the technical field desired by the user. The accumulated data can be used again as learning data. An example is shown in FIG.

The search result display unit shown in FIG. 3 displays patent data stored in the search result database. For example, the results can be displayed in an easy-to-understand manner using a browser.

  Hereinafter, the second embodiment will be described. Specifically, in the embodiment as shown in FIG. 3, a case where the present invention is applied to “extracting patents related to environmentally conscious resins out of resin technologies” will be described.

The reason for selecting the case of applying to “extracting patents related to environmentally conscious resins out of resin technologies” as Example 2 is “Recycling of technologies in the manufacturing process of refrigerators” selected in Example 1 When it is applied to `` extracting patents related to ``, the technology is mainly biased to the mechanical, electrical, and electronic fields. This is to prove that it is shown. If the result is good in the chemical / material field as well as the mechanical / electrical / electronic field, it can be suggested that the present invention is extremely effective for automatic patent search in a wide range of technical fields.

  Hereinafter, the procedure of Example 2 is followed. First, from the search input unit, the user uses a computer to input a search keyword or IPC classification as a query, and searches and collects sample patents from the patent database in FIG. This process corresponds to STEP1 in FIG. As a result, the patent document determined as a positive example is stored in the related patent database in FIG. 3, and the patent document determined as a negative example is stored in the unrelated patent database in FIG. The related patent database and the non-related patent database are made up of external storage devices (storage, etc.) connected directly to the terminal computer via a network or directly. The format of the accumulated data is not limited to a format such as PDF, and any format can be used as long as it can describe a patent document. In this example, “patent related to environmentally friendly resin among resin technologies” is a positive example, and for example, “patent for biodegradable resin and its manufacturing method” is applicable. Negative examples include patents not related to environmentally friendly resins such as “patents that improve resin yield”. Here, sample patents, that is, teacher data, are collected for 70 patents each for 2003 and 2006 published patents.

  Next, a patent document judged as a positive example is extracted from the related patent database and the unrelated patent database, and morphological analysis is performed. This process corresponds to STEP 2 in FIG. Thereby, the patent document is divided into words. For example, when the expression “the present invention relates to a biodegradable resin” appears in a patent document, this expression is an example of STEP 2 in FIG. 4 “invention / has // refrigerator / Similarly to “/ recycle / to / related / related / things / is / is” ”,“ invention / has /, / biodegradation / resin / to / related / related / things / is / ”is divided into word units. Is done. Here, / represents a break. The morphological analysis is performed by the central processing unit in FIG. 3, and the analysis result is stored in the patent identification unit in the main storage device in FIG.

  By the morphological analysis, words characteristic of positive examples are selected from the divided words. The feature word is selected by using the above equation (6). This processing corresponds to the processing of STEP 3 and STEP 4 in FIG. 1 ({“Recycling”, “Discarding”, “Reusing”... In FIG. 4 where extraction of refrigerator recycling is the target, {"recycling", "disposal", "reuse" ...}, but in this example, {"recycling", "biodegradation", "environment" ...・} Etc. were selected as feature words. The feature word selection is performed by the central processing unit shown in FIG. 3, and the result is stored in the patent identification unit in the main storage unit shown in FIG.

  Next, using the feature word as a feature vector, that is, the basis of a learning vector, a feature vector is generated based on the appearance frequency. The feature vector generation is performed by the central processing unit in FIG. 3, and the result is accumulated in the patent identification unit in the main storage device in FIG.

  Note that the entropy calculation time in the statistic calculation depends on the hardware configuration of the computer, but when using public patent information for 10 years, it is only a few minutes and is sufficient for searching a large amount of patent document data. Can be used.

  After generating the feature vector, machine learning is performed using the feature vector as shown in STEP 6 of FIG. As a machine learning technique, SVM is used as illustrated in FIG. SVM is an optimal technique for binary classification because it prevents overlearning by a technique called margin maximization. The machine learning by the SVM is performed by the central processing unit in FIG. 3, and the result is accumulated in the patent identification unit in the main storage device in FIG.

  Patents determined to be positive examples by the learning are accumulated in the search result database in FIG. These accumulated results are displayed on the search result display section. The display is performed in a list format such as an application number. The search result database is composed of an external storage device (storage or the like) that is directly connected to the terminal computer via a network. The format of the stored data is not limited to a format such as PDF, and any format can be used as long as it can describe a patent document.

  About said Example 2, search and evaluation of extraction performance were performed. The number of patents used for the evaluation is 100 (50 positive examples, 50 negative examples). In addition, a linear kernel is used as the SVM method.

Comparison of Table 2 as the result and Table 1 as the result of Example 1 showed that patent search can be performed with almost the same accuracy as Example 1. Therefore, the present invention is not limited to the first and second embodiments, and can be applied to automatic patent searches in a wide range of technical fields.

The flowchart regarding the patent identification part which concerns on this invention is shown. It is a graph which shows the basis which uses the conditions of several 6 by this invention. It is a figure which shows the structure of the patent document automatic search system which concerns on the Example of this invention. It is a figure which shows the procedure and concept which concern on the Example of this invention. The example of the database in this invention is shown.

Claims (4)

An automatic search system for patent document data,
A related patent database that stores, as a sample, patents that have already been searched in the technical field related to search output
An unrelated patent database that stores, as a sample, patents that have already been investigated that belong outside the technical field related to search output;
A morpheme analyzer that performs morphological analysis on patent document data stored in the related patent database and the unrelated patent database;
A patent identifying unit that discriminates patent documents belonging to the technical field related to search output by machine learning using feature weights selected using appearance frequencies and weights calculated from entropy for the morpheme-analyzed data; and
An automatic patent document search system comprising: an extraction unit that extracts all patent document data belonging to a technical field related to a search output from a determination result of the patent identification unit. In the patent identification unit according to claim 1, for all words appearing in the positive example and the negative example, a weight is obtained from the appearance frequency and entropy,
If the weight in the positive example of a word is larger than twice the weight in the negative example of the said word, the said word is used as a feature word of machine learning as a feature word, The patent document automatic search system characterized by the above-mentioned. An automatic search method for patent document data,
Means for storing patents already in the technical field related to search output in the related patent database as samples;
Means for storing, as a sample, an unpatented patent database of searched patents belonging to a field other than the technical field related to search output;
Means for performing morphological analysis on patent document data stored in the related patent database and the unrelated patent database;
A means for discriminating patent documents belonging to a technical field related to search output by machine learning using feature weights selected using appearance frequencies and weights calculated from entropy for the morpheme-analyzed data;
And a means for extracting all patent document data belonging to a technical field related to search output from the result of the identifying means. In the means for discriminating the patent document according to claim 3, for all words appearing in the positive example and the negative example, a weight is obtained from the appearance frequency and entropy,
If the weight in the positive example of a word is larger than twice the weight in the negative example of the said word, the said word is used as a feature word of machine learning as a feature word, The patent document automatic search method characterized by the above-mentioned.

Images