KR102468930B1 - System for filtering documents of interest and method thereof - Google Patents

Abstract

The present invention extracts words included in documents such as patents or papers, uses them to select a reference seed document through information amount analysis, and clusters documents similar to the reference seed through semantic similarity analysis, in a plurality of documents. A system and method for filtering documents of interest capable of filtering out noise or core documents are provided, comprising the steps of selecting at least one specific document of interest among a plurality of documents as a reference seed; Calculating a probability distribution by topic by performing topic analysis of each document based on word statistical information about each document, and calculating a distance with a reference seed using the probability distribution by topic of each document, and the calculation A step of selecting and filtering documents of interest according to the obtained similarity may be included.

Description

System for filtering documents of interest and its method {SYSTEM FOR FILTERING DOCUMENTS OF INTEREST AND METHOD THEREOF}

The present invention relates to a document-of-interest filtering system and method for automatically extracting noise or core data of interest from collected population data.

In general, technical documents (population) such as patents or papers collected through keyword search include noise documents to be removed in addition to valid documents in the target technology field to be analyzed. In addition, valid documents include core documents that are more closely related to the target technology field.

Effective patent documents necessary for efficient patent analysis must be collected, and for this purpose, a filtering process to remove noise documents or extract key documents is required.

For example, conventionally, in order to remove noise documents, a person directly reviews patent documents from a population and removes noise documents through a simple repetitive task of removing noise, which requires a lot of time and manpower. Of course, it takes a lot of time to extract key documents by manually reviewing and extracting document contents.

The following patent literature relates to a document search and classification method and system, but does not provide a solution to the above problems.

Korean Patent Publication No. 10-2006-0047306

The present invention is to solve the above-mentioned problems of the prior art, after extracting words included in documents such as patents or papers, using them to select a reference seed document through information amount analysis, and analyzing semantic similarities A document filtering system and method capable of filtering noise or core documents from a plurality of documents by clustering documents similar to seeds are provided.

A method for filtering documents of interest according to an embodiment of the present invention includes the steps of: (a) selecting at least one specific document of interest among a plurality of documents as a reference seed; (b) Based on the word statistical information of the plurality of documents, topic analysis of each document is performed to calculate a probability distribution by topic, and distance with the reference seed is calculated using the probability distribution by topic of each document Calculating; and (c) selecting and filtering the document of interest according to the calculated similarity.

The reference seed in the step (a) is selected based on the information entropy calculated through the probability of the occurrence frequency of the preset keyword in a plurality of documents, and the reference seed is the range or condition of the calculated information entropy and the information entropy for the preset seed. can be selected by mutual comparison.

The step (b) may include calculating a probability distribution in which each document belongs to each topic by applying a topic modeling algorithm to the word statistical information; calculating a degree of similarity between a reference seed and a document to be compared using a probability distribution by subject of each document and a subject probability distribution of a document selected as a reference seed; and grouping and providing documents similar to the reference seed based on the calculated similarity.

A target-of-interest document filtering system according to an embodiment of the present invention includes: a reference seed selector selecting at least one specific document of interest among a plurality of documents as a reference seed; a seed clustering unit that calculates a probability distribution for each subject through topic analysis of each document based on the word statistical information of the plurality of documents, and calculates a degree of similarity with a selected reference seed using the probability distribution for each subject of each document; and a document of interest filtering unit for selecting and filtering a document of interest according to the similarity of each document calculated above.

In addition, the filtering system may include a word extraction unit for extracting words for each document from a plurality of collected documents; a statistical information acquisition unit that measures the frequency of occurrence of the word extracted above in the document, and acquires word statistical information using the frequency of occurrence of the word and the inverse document frequency; and an information entropy calculation unit calculating information entropy of each document through a pre-set probability of appearance in the document of the word of interest relative to the number of occurrences of all words in the document measured as above.

According to an embodiment of the present invention, in the case of a filtering system for documents of interest such as noise or key documents, it was possible to discriminate the degree of containing information about the technology group of the analysis target only by qualitative analysis by the Shannon entropy method. The possibility of a document of interest can be expressed as a quantified number to help expert judgment.

It is possible to help give and maintain consistency for the problem of inconsistency in determining whether a document of interest, which is one of the problems of the existing expert's long-term simple repetitive work method, and systematically compensate for the process in which mistakes or errors may occur. By doing so, the quality of work can be improved.

In addition, topic analysis modeling, which is a topic analysis, plays a role of selecting and clustering documents that are semantically similar to the reference seed. to be extracted in batches. Through this, it is possible to innovatively reduce the number of man-hours for filtering documents of interest to be performed by experts.

In addition, in the present invention, since the process of filtering the documents of interest is performed multiple times starting from the process of selecting the reference seed, the range of document sets requiring selection can be sufficiently reduced. This has the effect of increasing the speed and efficiency of R&D by innovatively reducing the time investment of experts compared to the time cost required by the existing method.

1 is a conceptual diagram illustrating a system for filtering a document of interest according to the present invention.
2 is a configuration diagram illustrating a system for filtering a document of interest according to an embodiment of the present invention.
FIG. 3 is a diagram illustrating the seed clustering unit shown in FIG. 2 .
4 is a flowchart illustrating a method for filtering a document of interest according to an embodiment of the present invention.
5 is a UI screen showing a document filtering system according to an embodiment of the present invention.
FIG. 6 is a flowchart for explaining the word extraction and appearance frequency calculation process of FIG. 4 .
7 is a diagram illustrating the process of FIG. 6 .
8 is a flowchart illustrating the seed clustering process of FIG. 4 .
9A to 9C are diagrams for explaining LDA applied to the calculation of the subject probability distribution of FIG. 8 .
10 is a screen showing a document similar to a reference seed extracted according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

However, the embodiments of the present invention can be modified in many different forms, and the scope of the present invention is not limited to the embodiments described below. In addition, the embodiments of the present invention are provided to more completely explain the present invention to those skilled in the art.

In the drawings referred to in the present invention, the same reference numerals will be used for components having substantially the same configuration and function, and the shapes and sizes of elements in the drawings may be exaggerated for clearer explanation.

The term '~unit' used in this embodiment means software or a hardware component such as a field-programmable gate array (FPGA) or ASIC, and '~unit' performs certain roles.

However, '~ part' is not limited to software or hardware. '~bu' may be configured to be in an addressable storage medium and may be configured to reproduce one or more processors.

Therefore, as an example, '~unit' refers to components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, and procedures. , subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables.

Functions provided within components and '~units' may be combined into smaller numbers of components and '~units' or further separated into additional components and '~units'.

In addition, components and '~units' may be implemented to play one or more CPUs in a device or a secure multimedia card.

1 is a conceptual diagram illustrating a system for filtering a document of interest according to the present invention.

In general, in a patent or paper search, all documents having various application technologies or heterogeneous characteristics are collected after an outline of a search formula is drawn up in order to obtain a population as wide as possible. As a result, noise is included in the population, so the process of removing noise or extracting key documents inevitably occurs.

Therefore, the present invention proposes an efficient noise or key document extraction method. The method applied in the present invention is conceived from the crystal growth method, which is a method used to obtain pure crystals, and undergoes repetitive purification processes. In the process of crystal growth, in order to obtain pure crystals of sufficient size, a relatively easy-to-obtain method of growing small-sized pure crystals is used. At this time, the small-sized crystal is a crystal seed, and the corresponding crystal seed induces crystallization of the same type of solute in a stable saturated solution, etc., and grows into a crystal of sufficient size.

Likewise, in the method proposed in the present invention, by using Shannon entropy and semantic similarity analysis, some of the documents of interest recommended by the system are selected as reference seeds, and based on this, documents with high semantic similarity are clustered to filter by batch extraction and repeated purification.

2 is a configuration diagram illustrating a system for filtering a document of interest according to an embodiment of the present invention.

Referring to FIG. 2 , a system for filtering documents of interest according to an embodiment of the present invention 1 includes a document storage unit 10, a word of interest storage unit 50, and an information processing unit 100, and an information processing unit 100 ) may include a word extraction unit 110, a statistical information acquisition unit 120, an information entropy calculation unit 130, a reference seed selection unit 140, a seed clustering unit 150, and a document of interest filtering unit 160. can

The document storage unit 10 may receive and store a plurality of documents collected through a keyword search formula by the document search system. In an embodiment, the document may be a patent document or a paper, and the document search system may include Google Patent, Delphi-in, Kipris, WIPSON, and WINTELIPS. , NDSL, and the like.

The interest word storage unit 50 stores the interest word input from the user, that is, the interest keyword. The word of interest is composed of at least one keyword, and the word of interest is used to calculate information entropy of a document to be described later.

The word extraction unit 110 is configured to extract words for each document from a plurality of collected documents. The word extraction unit 110 extracts sentences included in the document through natural language processing, extracts words corresponding to adjectives and nouns through part-of-speech analysis of the extracted sentences, and among the extracted words is included in a preset stopword list. Words are removed to extract the necessary words.

The statistical information acquisition unit 120 is configured to measure the frequency of occurrence of each extracted word in the document, and acquire word statistical information using the frequency of occurrence of the word and the inverse document frequency.

Here, the statistical information acquisition unit 120 may include a frequency measuring unit 121 and a statistical information generating unit 125 . The frequency measurement unit 121 is configured to measure each occurrence frequency in a document of words extracted through the word extraction unit 110 . The statistical information generation unit 125 calculates an inverse document frequency by dividing the number of documents including a random word from the total number of documents with respect to the extracted word, and multiplies the frequency of occurrence of the random word by the inverse document frequency. It is configured to obtain word statistical information.

Meanwhile, the information entropy calculation unit 130 is configured to calculate the information entropy of each document by summing the occurrence frequency probabilities of the preset word of interest in the document with the number of occurrences of all words in the measured document. The keyword of interest may be a core keyword or a noise keyword, and a different weight may be set for each keyword to be reflected in information entropy calculation.

Information entropy is an important concept in information theory and measures uncertainty in a situation. That is, it has a high information entropy value in a situation with high uncertainty, and a low information entropy value in a situation with low uncertainty. For example, a coin toss event has a lower uncertainty than a dice toss event, that is, a lower uncertainty and information entropy value than a case where there are 2 possible events and 6 events. In addition, the amount of information in one system changes according to the probability of occurrence of each event in the same situation. In other words, assuming that the probability of occurrence of each event is different under the same number of events and other conditions, prediction of the event becomes more difficult when the probability of occurrence of each event is the same, so in this case, it has the highest information entropy value.

Such information entropy can be measured using the Shannon entropy algorithm of Equation 1 below for discrete probability distribution. When the corresponding equation is used, the amount of information about one system can be expressed as a numerical value, which means the degree of diversity of information in the system. That is, if a document contains the information desired by the user, that is, the word of interest evenly, it is highly likely to be a document of interest. Therefore, the information entropy of the document is calculated by measuring the amount of information about the word of interest.

[Equation 1]

However, i is the word of interest, and pi is the probability value of the word of interest i.

The reference seed selection unit 140 selects at least one specific document of interest among a plurality of documents as a reference seed based on the information entropy of each document calculated by the information entropy calculation unit 130 . At this time, at least one or more reference seeds may be selected, and a document outside the reference value may be recommended as the reference seed by mutually comparing information entropy with a set reference value.

The seed clustering unit 150 calculates a probability distribution by subject through subject analysis of each document based on the word statistical information on a plurality of documents generated by the statistical information generating unit 125, and uses the probability distribution by subject of each document. It is configured to calculate the degree of similarity with the selected reference seed.

Here, the seed clustering unit 150 may include a subject probability distribution calculating unit 151, a similarity calculating unit 153, and a similar document clustering unit 155, as shown in FIG. The topic probability distribution calculation unit 151 is configured to calculate a probability distribution that each document belongs to each topic through a topic modeling algorithm for word statistical information generated by the statistical information generation unit 125, and a similarity calculation unit ( 153) compares the topic probability distribution of the document selected as the reference seed in the interested seed selection unit 140 with the topic probability distribution of other documents to be compared, and calculates the degree of similarity between the document selected as the reference seed and other documents to be compared. and the similar document clustering unit 155 is configured to group and provide documents similar to the reference seed based on the similarity calculated by the similarity calculating unit 153.

In the above, as the topic modeling algorithm, any one of LDA (Latent Dirichlet Allocation), LSA (Latent Semantic Analysis), and PLSA (Probabilistic Latent Semantic Analysis) may be used, and as the similarity measurement algorithm, Hellinger Distance distance), any one of cosine similarity and Jaccard similarity coefficient may be used. Since the above algorithms are known technologies, a detailed description thereof will be omitted.

The document of interest filtering unit 160 selects and filters the documents of interest according to similarity information for each document calculated by the seed clustering unit 150 . Here, the document of interest filtering unit 160 may recommend a document of interest similar to the reference seed by comparing the calculated similarity information for each document with a preset threshold, and store the document of interest including the reference seed in the document storage unit 10. It can be displayed separately in the stored population document data. Thereafter, if the document of interest is noise, it will be removed from the population, and if the document of interest is a core patent, it will be extracted from the population.

As described above, in the present invention, information entropy for each document is obtained using the word of interest, and then a reference seed of interest is selected based on the information entropy for each document. Afterwards, documents similar to the reference seed are found, clustered, and provided.

Here, the word of interest is used to calculate the information entropy of each document, the reference seed is selected based on the information entropy of each document, and the TF-IDF value for each document is used to calculate the topic probability distribution for measuring similarity. It is calculated by measuring the vector between the topic probability distribution and the topic probability distribution of other documents.

In addition, the Shannon entropy algorithm may be used to calculate the information entropy of each document, and the Latent Dirichlet Allocation, Latent Semantic Analysis or Probabilistic Latent Semantic Analysis algorithm may be used to obtain the subject probability distribution of each document. Hellinger distance, Cosine similarity, or Jaccard similarity coefficient algorithm may be used to measure the similarity between the two.

The overall operation process of the document-of-interest filtering system configured as described above will be described in more detail with reference to FIGS. 4 to 9 . Hereinafter, a case in which the document of interest is a patent document will be described as an example.

4 is a flowchart illustrating a method of filtering a document of interest according to an embodiment of the present invention.

First, a keyword search formula or patent number (application number, publication number, or registration number) input from the user is provided to search for related patent documents from the patent database, and the searched patent documents are loaded into the document filtering system 1 and document It may be stored in the storage unit 10. Subsequently, the document filtering system 1 may receive a keyword for the word of interest and store it in the word of interest storage unit 50 (S100). Here, the patent document is patent data in Excel format that includes noise and key patents when searched by keyword, and the document contains bibliographic information (application number, application date, publication number, publication date, applicant, title, etc.) and summary, claims and citation information, etc. may be included.

For example, the user interface of the document filtering system 1 may be configured as shown in FIG. It can be stored in the storage unit 10, and a keyword (②) for the word of interest can be input and stored in the word of interest storage unit 50. In this case, the word of interest is a keyword of interest and may be a core keyword or a noise keyword.

Next, the word extraction unit 110 of the filtering system 1 extracts words included in a plurality of documents stored in the document storage unit 10, and the statistical information acquisition unit 120 extracts words included in each of the plurality of documents. It is possible to generate word statistical information for a plurality of documents by calculating the appearance frequency of (S200).

In an embodiment, the word extraction unit 110 may extract sentences included in the document through natural language processing as shown in FIG. 6 (S210), and correspond to adjectives and nouns through part-of-speech analysis of the extracted sentences. A word can be extracted (S220). Here, the word extraction unit 110 may remove words included in a preset stopword list among the extracted words (S230).

Next, the statistical information acquisition unit 120 may calculate the frequency of occurrence of the extracted word for each document through the extracted word (S240). In addition, the statistical information acquisition unit 120 may calculate an inverse document frequency (idf), which means statistics on the number of documents including the corresponding word with respect to the total number of documents (S250).

Here, the frequency of occurrence of a word (tf) is a value representing how often a specific word appears in a document, and can be calculated by adjusting the frequency value of the word according to the length of the document. For example, the frequency measurement unit 121 may use the number of appearances as the word appearance frequency (tf), but may also calculate it by Equation 2 below. At this time, the word with the highest occurrence frequency in the document will have a value of '1', and the other words will have a value less than 1.

[Equation 2]

where t: any word, d: any document, w: any word in document d, f(t,d): frequency of word t in document d.

In addition, the statistical information generating unit 125 calculates an inverse document frequency (idf) indicating how commonly a given word is included in a plurality of documents, and the inverse document frequency is the number of documents in the entire document. It can be calculated by taking a logarithmic scale after dividing by the number of documents containing .

For example, the inverse document frequency (idf) can be calculated by Equation 3 below.

[Equation 3]

Here, t; any word, d; any document, D; total number of documents, |d∈D:t∈d|; The number of documents containing the word t.

Next, the statistical information generation unit 125 may calculate a term frequency-inverse document frequency (TF-IDF) value by multiplying the word appearance frequency (tf) and the inverse document frequency (idf), and the TF-IDF value may be obtained as word statistical information (S260).

For example, the TF-IDF value may be calculated by Equation 4 below. Here, the reason why ' is added to the inverse document frequency is to prevent the inverse document frequency from being negative depending on the base of the logarithmic scale. may be

[Equation 4]

where TF(t,d) is the number of occurrences of word t in document d, |D| is the number of all documents, and |d∈D:t∈d| is the number of documents containing word t.

The TF-IDF value increases as the occurrence frequency of a word in a specific document increases and as the number of documents including the word among all documents decreases.

The statistical information generating unit 125 may standardize a plurality of documents stored in the document storage unit 10 using the TF-IDF values of words extracted for each document.

For example, when the number of documents stored in the document storage unit 10 is 8 and the total number of words extracted from the documents is 7, the statistical information generator 125 generates TF-IDF for each word extracted from the document. Values can be calculated and standardized as shown in Table 1 below.

[Table 1]

In this way, the operation process of the word extraction unit 110 and the statistical information acquisition unit 120 is illustrated in FIG. 7 .

Meanwhile, the information entropy calculator 130 may calculate information entropy (Shannon Entorpy) based on the word of interest stored in the word of interest storage unit 50 (S300). The word of interest is a keyword related to noise or core patents, and may be set by the user, and a weight may be set differently for each word of interest. Here, the weight means that a weight can be applied to the frequency of occurrence of the word of interest.

Specifically, the information entropy calculation unit 130 may calculate the information entropy of the target word using frequencies of words included in the target word and words not included in the target word among words included in the document. Frequency information measured by the frequency measurement unit 121 may be used as the frequency information of each word. In the user interface of FIG. 5, the 'seed patent calculation (③)' button may be an information entropy calculation command.

Here, information entropy can be calculated by Equation 5 below.

[Equation 5]

Here, n: means the value of the required amount of information for each document, k: number of categories of interest words, h _i : probability of occurrence of each word of interest, a word corresponding to word of interest i relative to the number of occurrences of all words in one document The frequency of occurrence of is the probability.

For example, when the keywords of the keyword of interest are stereo, lithography, and 3D, the frequencies of words included in the keyword of interest and words not included in the keyword of interest of a specific document may be as shown in Table 2 below.

[Table 2]

Here, Hi of Stereo in a specific document may be a value obtained by dividing 4, which is the frequency of occurrence of Stereo, by 50, which is the frequency of occurrence of all words. Based on this, when the words of interest (k1 to k3) in Table 2 are applied to Equation 5, the information entropy of the document may be '1.08'.

The matrix of Table 2 is generated for each document. In the present invention, words other than the word of interest list are treated as one word to form a group of words (words of interest), thereby maximizing the composition of the amount of information on key words (words of interest).

Next, the interest seed selection unit 140 selects a reference seed document by comparing the calculated information entropy of each document with a preset information entropy (S400). Here, when extracting noise, if the information entropy of each document is smaller than the preset reference value, the corresponding document will be selected as the reference seed, and when extracting core patents, the corresponding document only when the information entropy of each document is greater than the preset reference value. will be selected as the reference seed.

Then, the seed clustering unit 150 may extract and cluster at least one or more documents similar to the reference seed selected by the interested seed selection unit 140 (S500). The seed clustering unit 150 extracts at least one similar document similar to the reference seed. In order to extract similar documents, the vector similarity between documents may be analyzed using a topic modeling algorithm to consider potential associations between documents. have. If a plurality of reference seeds are selected by the interested seed selection unit 140, the seed clustering unit 150 may extract at least one similar document for each reference seed document.

For example, as shown in FIG. 8 , the seed clustering unit 150 applies a topic modeling algorithm to weights (TF-IDF matrix) for each word of each document generated through the statistical information generation unit 125 to obtain each A probability distribution in which a document belongs to each subject may be calculated (S510). For example, the subject modeling algorithm may be a Latent Dirichlet Allocation (LDA) algorithm.

The LDA algorithm is a known technology and is a tool commonly used in classifying documents by topic, and will be briefly described with reference to the Matlab code as shown in FIG. 9 . Basically, the LDA algorithm starts with the premise that a document is a group of words, that a document has a specific subject, and that the subject is shared across documents. For example, assuming that there are 8 documents as shown in FIG. 9A and that each document consists of a total of 16 words, it is possible to display in color according to the frequency of occurrence of words. The darker the green, the higher the frequency of occurrence of the word, and the darker the blue, the lower the frequency of occurrence. In the case of document No. 7 in FIG. 9A, it can be seen that only the words in the matrix (3, 4) have a very high frequency.

9B shows the distribution of topics, and there are 8 topics (Topic1 to Topic8), and it shows what words each topic has. That is, the subject means the distribution of words. For example, in the case of topic 1, the frequency of occurrence of words from the first to the fourth ((1,1) to (1,4)) is high. Therefore, when LDA is applied to the weight of each word of each document, a similar pattern is shown as shown in FIG. 9b, and through this, each subject is found.

Figure 9c shows the distribution of topics for each document. Red is used when creating data, and blue is found through LDA. In other words, when the order of the topics corresponding to the x-axis is ignored, it can be seen that the topics of each document can be similarly found through LDA.

In addition to Latent Dirichlet Allocation (LDA), Latent Semantic Analysis (LSA) or Probabilistic Latent Semantic Analysis (PLSA) may also be used as a topic modeling algorithm.

The number of topics (corresponding to 'technical fields') can be set in advance in the filtering system 1, and it has been confirmed that the most appropriate number of topics is classified into 8 to 10 through several tests. Therefore, as shown in Table 3 below, the topics were first classified into 9 categories, and then LDA was applied to a plurality of documents to classify each topic.

As shown in Table 3 below, the number of patent documents belonging to each subject derived as a result of LDA and the main keyword information constituting each subject can be indicated. Using the main keyword information corresponding to each subject, the characteristics of the subject it is possible to judge For example, in the case of Topic 1, it can be inferred that it is a technology cluster of adhesive particulate bonding.

[Table 3]

In this way, the subject probability distribution calculation unit 151 may extract keywords for each subject and calculate a probability distribution belonging to each subject for each patent document as shown in Table 4 below.

[Table 4]

Subsequently, the similarity calculation unit 153 may calculate the similarity between documents by performing a similarity analysis between documents using the probability distribution belonging to each subject, and the similarity may be determined by Hellinger distance and cosine similarity. and Jaccard similarity coefficient (S520).

For example, the similarity calculation unit 153 determines the relationship between the reference seed selected in the interest seed selection unit 140 and other documents to be compared by the Hellinger distance (H(P,Q)) of Equation 6 below. similarity can be calculated.

[Equation 6]

Here, i is the topic, k is the number of topics, pi is the topic probability distribution of the reference seed document, and qi is the topic probability distribution of the comparison target document.

The result value (H(P,Q)) calculated by the Hellinger distance has a value between 0 and 1. The smaller the result value, the greater the degree of similarity between the two documents, and the greater the result value, the greater the degree of similarity between the two documents. degree is small Therefore, the final similarity value (S(P,Q)) is obtained by subtracting the resultant value (H(P,Q)) of the Hellinker distance by 1 as shown in Equation 7 below for easy intuitive understanding, and then the subtracted value is It can also be used as a similarity value.

[Equation 7]

Table 5 below shows an example of a similarity matrix for each document calculated by Equations 6 and 7 above.

[Table 5]

Next, the similar document clustering unit 155 clusters similar documents similar to the reference seed based on the similarity information calculated by the similarity calculating unit 153 and provides them (S530). In this case, if there are a plurality of reference seeds, it is also possible to cluster similar documents for each reference seed.

Thereafter, the document of interest filtering unit 160 may remove or extract the reference seeds and similar documents clustered through the seed clustering unit 150 by filtering based on a preset threshold (S600). That is, the interest document filtering unit 160 may select and recommend documents of interest similar to the reference seed by comparing similarity information between the reference seed and the comparison target document with a preset threshold value. Also, the document of interest filtering unit 160 may separately display the document of interest including the reference seed in the population document data stored in the document storage unit 10 . Thereafter, if the document of interest is noise, it will be removed from the population, and if the document of interest is a core patent, it will be extracted from the population.

That is, as shown in FIG. 10, the filtering system 1 clusters and provides documents similar to the reference seed. At least one reference seed ④ and similar documents similar to the reference seed ⑤ and corresponding Information on the information entropy value (⑦) of the document and the degree of similarity (distance; ⑧) with the reference seed is provided respectively. In FIG. 10, patent number 2002-018463, which is one of the reference seeds (④), and similar patents (⑤) similar to the reference patent are shown. Here, similar patents (⑤) are extracted from only patents having similarities (⑧) lower than a preset threshold (⑥). Of course, if another reference seed is selected, patents similar to the reference seed will be re-extracted and displayed.

The step of selecting the reference seed (S400) to the step of filtering similar documents (S600) may be repeated as necessary to filter core patents or noise step by step.

The present invention described above is not limited by the above-described embodiments and the accompanying drawings, but is limited by the claims to be described later, and the configuration of the present invention can be varied within a range that does not deviate from the technical spirit of the present invention. Those skilled in the art can easily know that the present invention can be changed and modified accordingly.

1: document filtering system 10: document storage unit
50: word of interest storage unit 100: information processing unit
110: word extraction unit 120: statistical information acquisition unit
121: frequency measuring unit 125: statistical information generating unit
130: information entropy calculation unit 140: interest seed selection unit
150: seed clustering unit 151: subject probability distribution calculation unit
153: similarity calculation unit 155: similar document clustering unit
160: interest document filtering unit

Claims (19)

(a) selecting at least one specific document of interest among a plurality of documents as a reference seed;
(b) Based on the word statistical information of the plurality of documents, topic analysis of each document is performed to calculate a probability distribution by topic, and distance with the reference seed is calculated using the probability distribution by topic of each document Calculating; and
(c) selecting and filtering documents of interest according to the calculated similarity;
The reference seed in step (a) is,
A method of filtering a document of interest in which a preset word of interest is selected based on information entropy calculated through probability of occurrence in a plurality of documents.
delete The method of claim 1,
The reference seed is selected by comparing the calculated information entropy with a range or condition of information entropy for a preset seed.
The method according to claim 3, wherein the information entropy (n) is,
A method for filtering documents of interest calculated by Shannon entropy, which is the following formula.
[Equation]

Here, n: means the value of the required amount of information for each document, k: number of classified interest words, h _i : occurrence probability of each interest word, corresponding to the interest word i for the total number of occurrences of words in one document The probability of occurrence of a word.
The method according to claim 1, wherein the word statistical information of step (b) is,
A method of filtering a document of interest generated by extracting words from a plurality of documents and measuring the frequency of occurrence of the extracted words in the document.
The method of claim 1,
The reference seed is a document of interest to be extracted, and is one of noise and core data.
The method according to claim 1, wherein the word statistical information of step (b) is,
extracting words included in a plurality of documents;
measuring a frequency of appearance in the document of the word extracted above with respect to each of the plurality of documents;
Calculating an inverse document frequency by dividing the number of documents including the word from the total number of documents with respect to the word extracted above; and
Multiplying the occurrence frequency of the word by the inverse document frequency; a document of interest filtering method obtained through the step.
The method according to claim 7, wherein the step of extracting the word,
extracting sentences included in the document through natural language processing on the document;
extracting words corresponding to adjectives and nouns through part-of-speech analysis of the extracted sentence; and
and removing words included in a preset stopword list among the extracted words.
The method according to claim 1, wherein the (b) step,
calculating a probability distribution in which each document belongs to each topic by applying a topic modeling algorithm to the word statistical information;
Calculating a degree of similarity between a reference seed and a document to be compared using a probability distribution by subject of each document and a subject probability distribution of a document selected as a reference seed; and
and grouping and providing documents similar to the reference seed based on the calculated similarity.
The method of claim 9,
The subject modeling algorithm is any one of LDA (Latent Dirichlet Allocation), LSA (Latent Semantic Analysis), and PLSA (Probabilistic latent semantic analysis), and the similarity is Hellinger distance, Cosine Similarity and a Jaccard similarity coefficient.
The method according to claim 1, wherein the similarity of step (b) is,
A method of filtering documents of interest calculated by applying a probability distribution in which each document belongs to a corresponding subject to a Hellinger distance (H(P,Q)) of the following equation.
[Equation]

Here, k is the topic, t is the number of topics, p _k is the topic probability distribution of the reference seed, and q _k is the topic probability distribution of the comparison target document.
delete The method of claim 1,
The target document filtering method of claim 1 , wherein the target word includes at least one of a core keyword and a noise keyword.
a reference seed selection unit that selects at least one specific document of interest among a plurality of documents as a reference seed;
a seed clustering unit that calculates a probability distribution for each subject through topic analysis of each document based on the word statistical information of the plurality of documents, and calculates a degree of similarity with a selected reference seed using the probability distribution for each subject of each document; and
A document of interest filtering unit for selecting and filtering documents of interest according to the similarity of each document calculated above;
The reference seed selection unit,
When the reference seed is selected, the target-of-interest document filtering system selects based on information entropy calculated through an occurrence frequency probability of a preset term of interest in a plurality of documents.
delete The method of claim 14,
A word extraction unit for extracting words for each document from a plurality of collected documents;
a statistical information acquisition unit that measures the frequency of occurrence of the word extracted above in the document, and acquires word statistical information using the frequency of occurrence of the word and the inverse document frequency; and
and an information entropy calculation unit configured to calculate information entropy of each document through a pre-set probability of occurrence of the word of interest in the document with respect to the number of occurrences of all words in the document measured as described above.
The method according to claim 16, wherein the statistical information acquisition unit,
a frequency measurement unit for measuring the occurrence frequency of the words extracted through the word extraction unit in the document; and
With respect to the extracted word, an inverse document frequency is calculated by dividing the number of documents including a random word from the total number of documents, and word statistical information is obtained by multiplying the frequency of occurrence of the random word by the inverse document frequency. A document filtering system of interest comprising a;
delete delete

Images