JP5577546B2 - Computer system - Google Patents

Description

  The present invention relates to a computer system, and more particularly to a computer system that extracts information from a document.

  A technique for determining whether content in a document is important and extracting an important part is one of techniques for determining the contents of a document in accordance with a preset determination rule, and has been conventionally proposed ( For example, see Patent Literature 1 and Patent Literature 2).

  Patent Document 1 converts presentation conditions and collation conditions input by a user into knowledge rules expressed by a directed graph, and uses the knowledge rules as determination rules. A collation condition that is a determination condition is described by a relationship between a word corresponding to a topic and a word corresponding to an element. Further, in the collation between the document to be determined and the knowledge rule, the apparatus of Patent Document 1 is such that the appearance position of the corresponding word matches the direction and hierarchical structure of the directed graph, and the length of the link is greater. Find a short word combination from the target document.

  Further, Patent Document 2 discloses a classifier using a support vector machine, which is a statistical method, by extracting feature vectors from training data previously labeled with important sentences (positive examples) or non-important sentences (negative examples). Thus, a separation plane that classifies the feature vector into a positive example and a negative example is generated. In the technique of Patent Document 2, the separation plane corresponds to a determination rule. A feature vector is also extracted for a new document and compared with the generated separation plane to determine whether it is an important sentence or an unimportant sentence. As each element of the training data and the feature quantity extracted from the new document, that is, the feature vector, the length of the sentence, the appearance position of the sentence, or the presence or absence of a keyword appearing in the sentence is used.

JP 2007-193500 A JP 2003-036262 A

  In the technology for determining the content of a document as listed in Patent Document 1 and Patent Document 2, when creating a standard for determination, the training data that is the basis, or the presentation condition and the matching condition are sufficient in advance. If prepared, it is possible to extract important sentences with high accuracy. However, in Patent Document 1, the presentation condition and the collation condition are premised on manual input.

  In Patent Document 2, a method for creating training data is not particularly disclosed, but usually, training data is often created manually. For this reason, in Patent Document 2, it can be said that the training data is similarly premised on manual creation.

  A method in which a user explicitly registers a determination criterion in advance in the system as in the technique disclosed in Patent Document 1 can be expected to be determined with high accuracy when it can be handled by a small number of determination rules. However, when the scale of a document to be determined becomes large and a large number of complicated determination criteria are required, it is difficult to manually input presentation conditions and collation conditions. Further, when such a large number of complicated determination criteria are required, there arises a problem that it is difficult to create and maintain a consistent determination rule having no contradiction.

  On the other hand, the technique disclosed in Patent Document 2 obtains a separation plane corresponding to a determination rule by statistically processing training data labeled as an important sentence or a non-important sentence. . As a result, it is expected that a consistent determination rule can be easily constructed as compared with Patent Document 1.

  However, in order to build a highly accurate judgment rule by a statistical method, it is necessary to collect a large amount of data and attach a label indicating whether the collected data is an important sentence or a non-important sentence. . If the data is small, labels can be easily assigned manually. However, if the data is large, it takes time for manual work, and thus the use of the data becomes difficult.

  Further, when collating the determination target document and the determination rule, the system disclosed in Patent Document 1 performs the verification based on whether or not the word in the determination target document matches the word in the determination rule. Furthermore, since the system shown in Patent Document 1 uses a synonym dictionary, it can cope with fluctuations in the notation of words to be collated.

  On the other hand, in Patent Document 2, information on the presence / absence of a word in a document to be determined is used as an element of a feature vector. However, a document usually contains a lot of descriptions related to numerical values. Regarding numerical values, numerical values such as “to more than”, “to less than”, or “to to to” are not included in the numerical values. Therefore, it is necessary to collate based on the size or the range of numerical values, and it is necessary to construct a determination rule for numerical values corresponding to them.

  The technique of Patent Document 2 corresponds to the construction of a judgment rule based on word notation, that is, matching or mismatching of character strings indicating words, but for construction of a judgment rule in consideration of the size of numerical values or the range of numerical values. Not supported. For this reason, when it is necessary to consider the magnitude | size or the range of a numerical value, there exists a problem that determination accuracy falls.

  It is technically easy to make a judgment rule correspond to a numerical value and to manually construct a judgment rule corresponding to the numerical value. It becomes difficult to efficiently construct a certain highly accurate determination rule.

  It is an object of the present invention to efficiently generate a determination rule used in a system for extracting a portion determined to be important from a document from a large number of documents, and to consider a numerical value and a numerical range. The provision of a system for easily generating a highly accurate determination rule.

  A typical example of the present invention is as follows. That is, a computer system that includes a processor that performs arithmetic processing and a storage device that is connected to the processor, and in which the processor analyzes a document, the document includes a plurality of words, and forms a sentence. A plurality of elements, the plurality of elements including a sentence or a paragraph; and the processor determines a plurality of first documents, a reference document including a reference to the first document, and an important part determination. The second document to be performed is input, and the elements are extracted from each first document, and a reference location to the first document is extracted from the reference document as reference information, Based on the similarity calculated from the elements extracted from each first document and the reference information, the similar elements between the elements extracted from each first document and the reference information are determined as important points. Is Then, a portion that is not similar to the first extracted element from the first document and the reference information is regarded as a non-important portion from the first document to the second element. The extracted element is divided, the first feature amount of each document is acquired based on the plurality of words included in the divided first element and second element, and the acquired first A determination rule for determining whether or not the important part is included is generated based on the feature amount of 1, and the element is extracted from the second document and included in the element extracted from the second document A second feature amount is obtained based on the plurality of words extracted, and extracted from the second document by comparing the generated determination rule with the obtained second feature amount Classify elements into important and non-critical parts.

  According to an embodiment of the present invention, a determination rule for extracting a portion determined to be important from a document is efficiently generated.

It is a block diagram which shows the structure of the important location determination system by the 1st Embodiment of this invention. It is a block diagram which shows the detailed structure of the document classification | category part of the 1st Embodiment of this invention. It is explanatory drawing which shows the example of the document division | segmentation rule of the 1st Embodiment of this invention. It is a block diagram which shows the detailed structure of the teacher data generation part of the 1st Embodiment of this invention. It is explanatory drawing which shows the example of the attribute name and unit contained in the attribute extraction rule of the 1st Embodiment of this invention. It is explanatory drawing which shows the example of the pattern contained in the attribute extraction rule of the 1st Embodiment of this invention. It is a block diagram which shows another structure of the teacher data generation part of the 1st Embodiment of this invention. It is a block diagram which shows the detail of the determination data generation part of the 1st Embodiment of this invention. The block diagram at the time of mounting the important location determination system of the 1st Embodiment of this invention in the computer is shown. It is a block diagram which shows the structure of the important location determination system of the 2nd Embodiment of this invention. It is a block diagram which shows the detailed structure of the teacher data generation part of the 2nd Embodiment of this invention. It is a block diagram which shows the detail of the determination rule production | generation part of the 2nd Embodiment of this invention. It is a block diagram which shows the detail of the determination data production | generation part of the 2nd Embodiment of this invention. It is a block diagram which shows the detail of the determination process part of the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the important location determination system of the 3rd Embodiment of this invention. It is explanatory drawing which shows the example of a display of the determination result of the 3rd Embodiment of this invention. It is explanatory drawing which shows another example of a display of the determination result of the 3rd Embodiment of this invention.

(First embodiment)
A first embodiment of the present invention will be described with reference to FIGS.

  FIG. 1 is a block diagram showing the configuration of an important point determination system according to the first embodiment of the present invention.

  The important point determination system of the present invention includes a case document 101, a reference information description document 102, a document classification unit 103, a teacher data generation unit 104, a determination rule generation unit 105, a determination rule 106, a determination target document 107, a document division unit 108, A document classification rule 109, a determination data generation unit 110, and a determination processing unit 111 are provided. The case document 101, the reference information description document 102, the determination rule 106, the determination target document 107, and the document classification rule 109 are a set of data. The document classification unit 103, the teacher data generation unit 104, the determination rule generation unit 105, the document division unit 108, the determination data generation unit 110, and the determination processing unit 111 are programs.

  The case document 101 is a document similar to a document (determination target document 107) for which an important part is determined, a document in the same field as the determination target document 107, a document including contents related to the determination target document 107, and the like. , Keep past documents as examples. For example, among the documents similar to the determination target document 107, documents accumulated in the past are accumulated in the case document 101.

  The reference information description document 102 holds a document including a description for referring to a specific part in the case document 101. For example, a document instructed by the user that a specific attribute in the case document 101 is important is stored in the reference information description document 102.

  The description for referring to the case document 101 of the reference information description document 102 includes the document name, chapter number, heading, number of pages, paragraph number, column number, line number, and sentence of the corresponding part of the case document 101. The information which combined the whole or a part etc. suitably is contained.

  The document classification unit 103 divides the case document 101 into units (case document elements) for determining important parts. Further, the divided case document elements are classified into important elements that are important parts or other non-important elements that are referred to from the reference information description document 102.

  The teacher data generation unit 104 converts important elements and unimportant elements classified by the document classification unit 103 into teacher data. The teacher data is used to generate the determination rule 106.

  The determination rule generation unit 105 generates the determination rule 106 based on the teacher data generated by the teacher data generation unit 104 by using a generally used machine learning technique.

  The determination target document 107 holds a document that is a target for determining an important part using the determination rule 106. The document dividing unit 108 divides the determination target document 107 into determination units (determination document elements) using the document division rule 109. The document division rule 109 holds information related to a sentence structure such as a chapter or paragraph or a table format.

  The determination data generation unit 110 generates determination data by converting each determination document element into a data format that can be used by the determination processing unit 111. In general, since the machine learning technique used to generate the determination rule 106 is paired with the technique used in the determination process, the data format output in the determination data generation unit 110 is teacher data. The data format is the same as that generated by the generation unit 104.

  The determination processing unit 111 applies the determination rule 106 to the determination data, determines whether each determination data is an important element or a non-important element, and outputs the result as a determination result 112.

  FIG. 2 is a block diagram illustrating a detailed configuration of the document classification unit 103 according to the first embodiment of this invention.

  The document classification unit 103 includes a document division unit 201, a reference information extraction unit 203, a matching unit 205, a classification processing unit 206, a document classification rule 202, and a reference information extraction rule 204. The document division unit 201, the reference information extraction unit 203, the matching unit 205, and the classification processing unit 206 are programs, and the document classification rule 202 and the reference information extraction rule 204 are storage areas such as a database.

  The document dividing unit 201 illustrated in FIG. 2 divides the case document into units (case document elements) for performing the determination process using the document dividing rule 202. The document dividing unit 201 may be the same as the document dividing unit 108 shown in FIG. 1, and the document dividing rule 202 may be the same as the document dividing rule 109 shown in FIG. The document division rule 202 holds information on the structure of a sentence such as a chapter heading or paragraph, or a table format.

  Here, an example of the document division rule 202 is shown below.

  FIG. 3 is an explanatory diagram illustrating an example of the document division rule 202 according to the first embodiment of this invention.

  The document division rule 202 shown in FIG. 3 indicates information related to the structure of a sentence by a pattern including a chapter heading or a table title. The document division rule 202 includes a pattern 301, a hierarchy level 302, and contents 303. 3 includes an entry 304 to an entry 307 in each line.

  A pattern 301 indicates a chapter heading or a table title pattern described in a document. The hierarchical level 302 indicates the depth of the hierarchical structure when the description shown in the pattern 301 has a hierarchical structure such as Chapter 1, Section 2, etc., as in the chapter heading. The hierarchical level 302 indicates that the higher the numerical value, the deeper the hierarchy.

  A content 303 indicates the content of an item shown in the pattern 301. The contents 303 are information for facilitating the management of the pattern 301 and are not particularly used in the processing by the document dividing unit 201.

  In the column of the pattern 301 shown in FIG. 3, “N” indicates that an arbitrary number is described, “A” indicates that an arbitrary alphabetic character is described, and “*” indicates Each indicates that an arbitrary character string is described. The document classification unit 103 extracts a portion that matches the pattern 301 by comparing the pattern 301 with the contents of each case document 101.

  For example, the document classification unit 103 extracts a character string such as “Chapter 1 general remarks” or “Chapter 2 configuration” from the case document 101 using the pattern 301 of the entry 304. Character strings such as “1 Overview” and “2.1 Specifications” are extracted. In addition, a character string such as “(a) memory capacity is 5 MB” or “(b) hard disk is 1 TB” is extracted by the pattern 301 of the entry 306, and “• two memory slots” are extracted by the pattern 301 of the entry 307. Extract a string.

  After extracting the part that matches the document division rule 202 in the case document 101, when the text exists immediately after the extracted part, the document dividing unit 201 collects the extracted part and the immediately following text as one. The case document 101 is divided as a case document element. When there is no text next to the extracted part, the document dividing unit 201 divides the case document 101 by using the extracted part as a case document element.

  When the case document 101 is divided by using the extracted part and the text immediately after as a single case document element, the divided case document element has a lower hierarchy in the hierarchical structure indicated by a chapter heading or the like. If included, the document dividing unit 201 acquires, as context information, the heading of the chapter in the higher hierarchy that matches the pattern 301 in which the hierarchical level 302 is stored, and also uses the case document element as the heading of the chapter in the lower hierarchy. Get as. Then, the acquired context information is added to the acquired case document element.

  For example, a case where the case document 101 includes the following hierarchical structure is shown.

“Chapter 2 Structure…
2.2 Specifications (a) Memory capacity is 5MB
(B) Hard disk is 1TB
... "... Case document 101
When the document division rule 202 shown in FIG. 3 is applied to the example of the case document 101 described above, it is determined that all the line items correspond to the pattern 301, and the determined items are all the line items. Since it is determined that the text does not exist immediately after, the following two case document elements are generated.

"Chapter 2 Configuration ... Context Information 2.2 Specification ... Context Information (a) Memory capacity is 5 MB" ... Example Document Element "Chapter 2 Configuration ... Context Information 2.2 Specification ... Context information (b) Hard disk is 1 TB ”... Case document element The above two documents are extracted as case document elements. Here, “Chapter 2 Configuration” and “2.2 Specification” are items added as context information because values are stored in the hierarchical level 302 of the corresponding pattern 301. In addition to chapter headings, the context information may include document names, page numbers, paragraph numbers, or line numbers.

  In addition, when the text immediately after the extracted case document element is divided into a plurality of paragraphs, even if the case document element is generated for each paragraph, the context information, the location extracted immediately before, and the paragraph are combined. Good. Furthermore, a case document element may be generated with each sentence in the text as a unit.

  In addition to the character string, the pattern 301 of the document division rule 202 is a rule regarding the format described in the case document 101, such as the type of font used, the size of the character, the presence or absence of an underline, or the presence or absence of indentation. Any rule may be stored as long as it is.

  The reference information extraction unit 203 illustrated in FIG. 2 extracts a description for referring to a specific location in the case document 101 from the reference information description document 102 based on the content of the reference information extraction rule 204. Similar to the document division rule 202, the reference information extraction rule 204 stores a unique pattern applicable to the description of the reference location, as with the pattern 301 shown in FIG.

  The reference information extraction unit 203 extracts the reference information from the reference information description document 102 by searching the reference information description document 102 for a place where these patterns match. That is, the reference information extracted by the reference information extraction unit 203 is information obtained by changing the document held in the reference information description document 102 into a format that can be compared with the case document element generated by the document dividing unit 201. .

  The reference information extracted by the reference information extraction unit 203 is the document name, chapter number, chapter heading, page number, paragraph number, column number, line number of the corresponding case document 101, Including some combinations. Which item is included in the reference information depends on the content of the reference information description document 102.

  The matching unit 205 matches each case document element extracted by the document dividing unit 201 with each reference information extracted by the reference information extracting unit 203. The matching unit 205 performs matching based on whether the corresponding item exists and whether the content of the corresponding item matches.

  In addition, when only a part of the sentence of the corresponding part is included in the reference information, the matching unit 205 performs reference by using a code matching technique based on dynamic programming (see, for example, JP-A-2002-221984). It is possible to flexibly match the content in the information with the corresponding text in the case document element.

  Next, the matching unit 205 calculates the matching degree between the reference information and the case document element based on the matching result. The degree of coincidence is calculated based on the number of matches between the matched reference information and the case document elements.

  Further, when a technique based on dynamic programming is used, the matching unit 205 can calculate the distance between the reference information and the case document element. For this reason, by calculating the reciprocal of the calculated distance, a function that obtains a larger value as the distance is smaller may be obtained, and the value calculated by this function may be used as the matching degree.

  As a result of matching, the classification processing unit 206 classifies the case document element having the highest matching degree with respect to each reference information as a case document element (important element) that is an important part, does not correspond to any reference information, and matches Class document elements with low degrees are classified as case document elements (non-important elements) that are not important parts. Then, the classified important elements and non-important elements (classification results) are sent to the teacher data generation unit 104.

  FIG. 4 is a block diagram illustrating a detailed configuration of the teacher data generation unit 104 according to the first embodiment of this invention.

  The teacher data generation unit 104 includes a word division unit 401, an attribute information extraction unit 402, an attribute extraction rule 403, a word totaling unit 404, a word list 405, and a data conversion unit 406. The word division unit 401, the attribute information extraction unit 402, the word totaling unit 404, and the data conversion unit 406 are programs. The attribute extraction rule 403 and the word list 405 are storage areas such as a database.

  In the present embodiment, a support vector machine (see, for example, Japanese Patent Application Laid-Open No. 2003-36262) is used for processing and determination processing in the determination rule generation unit 105. However, this is a machine learning technique that uses numerical data as teacher data. Any machine learning technique may be used as long as there is no particular limitation.

  A word dividing unit 401 in FIG. 4 divides sentences included in important and non-important elements input from the document classification unit 103 into words. As a technique for dividing a sentence into words, a morphological analysis technique generally used in the field of natural language processing or machine translation (for example, “Iwanami Course Software Science (15) Natural Language Processing”, Iwanami Shoten, 1996) is used. It may be used.

  In addition, when the target document is described in a language such as English, and words in the target document can be separated by a blank added in advance, the word dividing unit 401 extracts the blank by extracting the blank from the sentence. It may be divided into words.

  A document including important elements and non-important elements divided into words is sent to the word totaling unit 404 and also sent to the attribute information extracting unit 402. Based on the rules stored in the attribute extraction rule 403, the attribute information extraction unit 402 extracts attribute information including attribute names and attribute values from important elements and non-important elements.

  The contents of the attribute extraction rule 403 will be described with reference to FIGS. 5A and 5B. 5A and 5B show rules included in the attribute extraction rule 403. FIG.

  FIG. 5A is an explanatory diagram illustrating an example of attribute names 501 and units 502 included in the attribute extraction rule 403 according to the first embodiment of this invention.

  The attribute extraction rule 403 is a rule used by the attribute information extraction unit 402 to extract attribute information from important elements and non-important elements. The attribute extraction rule 403 includes an attribute name 501, a unit 502, and a pattern 504. The attribute name 501 indicates the notation of the attribute name in each document, and the unit 502 indicates the notation of the attribute value unit described together with the attribute value.

  Note that the attribute extraction rule 403 is generated in advance based on the attribute name, unit, and pattern included in the case document 101 and the like.

  In FIG. 5A, the attribute name 501 column stores notation of attribute names assumed to be included in important elements or non-important elements (hereinafter referred to as documents), and unit 502 corresponds to each attribute name 501. The unit description described together with the attribute value to be stored is stored. The entry 503-1 indicates that a notation in which the attribute name 501 is “memory size” and the unit 502 is “GB” or “GB” is described in the document. Similarly, the entry 503-2 and the entry 503-3 indicate what notation is described in the document.

  FIG. 5B is an explanatory diagram illustrating an example of the pattern 504 included in the attribute extraction rule 403 according to the first embodiment of this invention. A pattern 504 illustrated in FIG. 5B indicates a pattern in which attribute information is written in a document.

  According to the pattern 504 shown in FIG. 5B, when the document includes a pattern in which the attribute name 501 shown in FIG. 5A is described in <attribute name> and the unit 502 shown in FIG. 5A is described in <unit>. The information extraction unit 402 determines that the document including the pattern matches the attribute extraction rule 403. “N” shown in FIG. 5B indicates a pattern in which an arbitrary numerical value is described.

  For example, if there is a description “memory size is 1 GB” in the document, the description matches the pattern 504-1 of the attribute extraction rule 403, so the attribute information extraction unit 402 reads “memory size is 1” from the document. “GB” is extracted as attribute information.

  Also, for example, if there is a description “operating voltage 100 V” in the document, the description matches the pattern 504-2 of the attribute extraction rule 403, so the attribute information extracting unit 402 reads “operating voltage 100” from the document. “V” is extracted as attribute information.

  In the attribute extraction rule 403 shown in FIG. 5B, a rule indicating a numerical range indicated by a symbol such as “N” may be added. Further, the attribute information extraction unit 402 may extract not only the attribute name 501 and the attribute value indicated by the unit 502 but also the item name (item name) having the attribute indicated by the attribute name 501 as attribute information.

  When the thing name having the attribute indicated by the attribute name 501 is also extracted, “<thing name> <attribute name> is N <unit>” is added to the attribute extraction rule 403 in FIG. It is also possible to easily extract the name of an object by searching for a word corresponding to the location of, and include the extracted name of the object in the attribute information.

  Furthermore, the attribute extraction rule 403 may include a list of presumed object names such as an attribute name 501 and a unit 502 shown in FIG. 5A. In addition, any word or phrase (a set of consecutive words) that matches the attribute extraction rule 403 as described above may be extracted as an object name from an important element or a non-important element.

  The word totaling unit 404 shown in FIG. 4 totals all the words included in the important elements and non-important elements sent from the word dividing unit 401 and all the attribute names 501 extracted by the attribute information extracting unit 402. , Merge overlapping words and attribute names 501. The merged word and attribute name 501 are stored in the word list 405.

  Further, the data conversion unit 406 is necessary to generate the determination rule 106 based on the contents of the important element and the non-important element, the attribute information extracted by the attribute information extraction unit 402, and the contents of the word list 405. Generate teacher data. In this embodiment, since a support vector machine is used for determination rule generation processing and determination processing, the data conversion unit 406 converts the contents of important elements and non-important elements into multidimensional vectors. The teacher data is represented by a multidimensional vector.

  Specifically, first, the data conversion unit 406 assigns each word and attribute name 501 in the word list 405 to each element of the multidimensional vector. In addition, when the attribute name includes the thing name in the attribute information, the data conversion unit 406 assigns one element of the corresponding multidimensional vector to the set of the thing name and the attribute name 501.

  Next, the data conversion unit 406 stores “1” in the element corresponding to the word included in each important element and non-important element, and “0” in the other elements. However, if the corresponding word is extracted as attribute information, “0” is assigned to the element corresponding to the word, and the attribute value represented by the numeric character string is converted into numerical data, and the attribute name 501 is converted. Store as the value of the corresponding element.

  An example of converting each important element and non-important element into a multidimensional vector is shown below.

  It is assumed that the following words and attribute names 501 are stored in the word list 405.

Word: PC, CPU, memory, memory size, operating voltage Attribute name 501: Memory size, operating voltage When the above word and attribute name 501 are stored in the word list 405, the obtained multidimensional vector is the following 7 dimensions Vector.

Multi-dimensional vector: (<PC>, <CPU>, <memory>, <memory size>, <operating voltage>, [memory size], [operating voltage])
When a word corresponding to X exists, <X> indicates “1”, and when a word corresponding to X does not exist, <X> indicates “0”. When an attribute value corresponding to the attribute name 501 corresponding to Y is set, [Y] indicates “1”, and when an attribute value corresponding to the attribute name 501 corresponding to Y is not set, Y] indicates “0”.

  Here, it is assumed that the following character strings are included in the important elements or the unimportant elements sent from the word dividing unit 401 via the word totaling unit 404.

"PC has CPU and memory"
When the above-described character string is included in the important element or the non-important element, the obtained multidimensional vector is shown as follows by setting “1” to the element corresponding to the corresponding word.

Multidimensional vector: (1, 1, 1, 0, 0, 0, 0)
That is, in the multidimensional vector, the words “PC”, “CPU” and “memory” exist in the target important or non-important elements, and the words “memory size” and “operating voltage” do not exist. Indicates.

  Further, it is assumed that the following character strings are included in the important elements or the unimportant elements sent from the word dividing unit 401 via the word totaling unit 404.

“PC operating voltage is 100 V”
When the important element or the unimportant element is a character string as described above, the following character string is extracted by applying the attribute extraction rule 403 shown in FIGS. 5A and 5B.

Attribute name: Operating voltage Attribute value: 100V
For this reason, the obtained multidimensional vector is as follows.

(1, 0, 0, 0, 0, 0, 100)
That is, it is shown that the word “PC” exists in the target important element or non-important element, and no other word exists. Further, it is indicated that there is a description in which the operating voltage is “100 (V)” in the target important element or non-important element. Since the word “operating voltage” is extracted as attribute information, 0 is stored in the element indicating the presence or absence (the fifth element from the left), and only the element related to the attribute value (the rightmost element) has a value. Is stored.

  When the word “operating voltage” is described alone as an important element or an unimportant element, 1 is stored in the fifth element from the left of the vector, and 0 is stored in an element related to the attribute value. In the above example, the data conversion unit 406 ignores all words that are not included in the word list 405.

  Further, in the above example, the data conversion unit 406 stores the numerical data indicating the attribute value as it is in the corresponding multidimensional vector element, but the size is changed by multiplying a predetermined numerical value for each attribute name. The obtained value may be stored in an element of a multidimensional vector. Further, numerical data indicating the attribute value may be normalized between 0 and 1, and the obtained value may be stored in an element of a multidimensional vector.

  Furthermore, in the above-described example, all the words are assigned to vector elements, but the data conversion unit 406 may assign words having similar meanings to one vector element.

  FIG. 6 is a block diagram illustrating another example of the teacher data generation unit 104 according to the first embodiment of this invention.

  When words having similar meanings are assigned to one vector element by the data conversion unit 406, the data conversion unit 406 is connected to a synonym dictionary 407 storing a list of similar words. Then, by searching the contents of the synonym dictionary 407, the word determined to be a synonym in the word list 405 is assigned to the element of the same vector in the multidimensional vector.

  For example, in the above example, when “memory” and “memory size” are similar, the synonym dictionary 407 holds a set of “memory” and “memory size”. Then, the data conversion unit 406 refers to the set held in the synonym dictionary 407 and assigns the attribute names 502 of “memory” and “memory size” to the elements of one vector.

  The multidimensional vector generated as described above indicates teacher data. The data conversion unit 406 sends the generated teacher data to the determination rule generation unit 105.

  The determination rule generation unit 105 generates a determination rule 106 using the transmitted teacher data. As described above, the determination rule generation unit 105 in the present embodiment is assumed to use a support vector machine, and applies the support vector machine to multidimensional vectors classified for each important element and non-important element. The determination rule 106 is generated.

  With the support vector machine, the determination rule generation unit 105 generates a separation plane in the multidimensional vector. Accordingly, the determination rule 106 holds information indicating a range (or distribution) of values that can be taken by the attribute name 501 included in the important element and the non-important element and the attribute value corresponding to the attribute name 501.

  The determination processing unit 111 can acquire an important attribute name 501 and values that can be taken by the attribute value corresponding to the attribute name 501 according to the determination rule 106.

  On the other hand, processing for determining whether or not the content of the determination target document 107 shown in FIG. 1 is an important element will be described below. As described above, the determination target document 107 is a target document from which important elements are extracted.

  First, the document dividing unit 108 divides the determination target document 107 into determination document elements that are determination units. The processing performed by the document dividing unit 108 is the same as that of the document dividing unit 201 in the document classification unit 103, and the document dividing rule 109 used at this time is the same as the document dividing rule 202 used in the document dividing unit 201. The same can be used.

  That is, the determination target document 107 is divided into chapters or paragraphs by the document dividing unit 108 and sent to the determination data generating unit 110.

  Next, the determination data generation unit 110 converts each determination document element divided by the document division unit 108 into a data format that can be used by the determination processing unit 111. When the support vector machine is used for the determination processing unit 111 as in the case of the determination rule generation unit 105, the format of the teacher data and the determination data is the same. Therefore, the processing of the determination data generation unit 110 is the same as that of the teacher data generation unit 104. The same process.

  That is, the determination data generation unit 110 divides each determination document element sent by the document division unit 108 into words and further converts them into multidimensional vectors. Accordingly, the determination processing unit 111 can compare the determination rule 106 with the multidimensional vector converted from the determination document element.

  FIG. 7 is a block diagram illustrating details of the determination data generation unit 110 according to the first embodiment of this invention.

  The determination data generation unit 110 includes a word division unit 701, an attribute information extraction unit 702, an attribute extraction rule 703, a word list 704, and a data conversion unit 705. The word division unit 701, the attribute information extraction unit 702, and the data conversion unit 705 are programs. The attribute extraction rule 703 and the word list 704 are storage areas such as a database.

  The word division unit 701, the attribute information extraction unit 702, the attribute extraction rule 703, and the data conversion unit 705 are respectively the word division unit 401, the attribute information extraction unit 402, the attribute extraction rule 403, and the data conversion unit 406 in the teacher data generation unit 104. The same.

  The difference between the teacher data generation unit 104 and the determination data generation unit 110 is that the word totaling unit 404 in the teacher data generation unit 104 is not included in the determination data generation unit 110, and the word list 704 used in the data conversion unit 705 is different from the teacher data generation unit 104. This is to use the word list created by the word totaling unit 404 in the data generation unit 104. Thus, the determination document element is converted into a multidimensional vector based on the word and the attribute name 501 included in the case document.

  When the teacher data generation unit 104 includes the synonym dictionary 407, the determination data generation unit 110 also includes the synonym dictionary 407. The data conversion unit 705 is connected to the synonym dictionary 407 and assigns similar attribute names 501 among the attribute names 501 to elements of the same vector in the multidimensional vector.

  The determination data generation unit 110 sends the converted multidimensional vector (determination data) to the determination processing unit 111.

  The determination processing unit 111 compares the determination data sent from the determination data generation unit 110 with the determination rule 106 using a support vector machine, thereby determining whether each determination data is an important element or an unimportant element. And the result is output as the determination result 112.

  Specifically, the determination processing unit 111 searches the determination rule 106 based on the multidimensional vector included in the determination data, and determines whether the attribute corresponding to the attribute name 501 included in the determination data is an important element or a non-important element. Get what it is.

  Further, the determination processing unit 111 searches the determination rule 106 for the attribute name 501 and the attribute value indicated by the multidimensional vector of the determination data, thereby determining the attribute value included in the determination data and the attribute value included in the determination rule 106. And get away that the distance. That is, the determination processing unit 111 searches the determination data in the determination rule 106 to obtain how far the attribute value included in the case document 101 is different from the attribute value included in the determination target document 107. it can.

  FIG. 8 is a configuration diagram when the important point determination system according to the first embodiment of this invention is mounted on the computer 100.

  The computer 100 includes an information processing device 801, an input device 802, a display device 803, a storage device 804, a case document 814, a reference information description document 815, a determination target document 816, a document division rule 817, a reference information extraction rule 818, and an attribute extraction rule. 819, a word list 820, and a determination rule 821.

  The information processing device 801 is an arithmetic device such as a CPU for executing various programs necessary for important point determination processing. The input device 802 is a device for a system user to operate the system, and is a device such as a keyboard or a mouse that is generally used. The display device 803 is a device for outputting the determination result 112, and is a generally used device such as a monitor or a speaker.

  The storage device 804 stores various programs necessary for the important part determination process and information on the progress of the process. The storage device 804 includes a document division program 805, a reference information extraction program 806, a matching program 807, a classification processing program 808, a word division program 809, an attribute information extraction program 810, a word aggregation program 811, a data exchange program 812, and a determination A processing program 813 is stored.

  The document division program 805 performs processing corresponding to the document division unit 201 and the document division unit 108 included in the document classification unit 103. The reference information extraction 806 performs processing corresponding to the reference information extraction unit 203 included in the document classification unit 103. The matching program 807 performs processing corresponding to the matching unit 205 included in the document classification unit 103.

  The classification processing program 808 performs processing corresponding to the classification processing unit 206 included in the document classification unit 103. The word division program 809 performs processing corresponding to the word division unit 401 included in the teacher data generation unit 104 and the word division unit 701 included in the determination data generation unit 110.

  The attribute information extraction program 810 performs processing corresponding to the attribute extraction unit 402 included in the teacher data generation unit 104 and the attribute extraction unit 702 included in the determination data generation unit 110. The word totaling program 811 performs processing corresponding to the word totaling unit 404 included in the teacher data generation unit.

  The data exchange program 812 performs processing corresponding to the data conversion unit 406 included in the teacher data generation unit and the data conversion unit 705 included in the determination data generation unit 110. The determination processing program 813 performs processing corresponding to the determination processing unit 111.

  The case document 814 stores the case document 101 shown in FIG. 1, and the reference information description document 815 stores the reference information description document 102 shown in FIG. The determination target document 816 stores the determination target document 107 shown in FIG. 1, and the document division rule 817 stores the document division rule 202 shown in FIG. 2 and the document division rule 109 shown in FIG.

  The reference information extraction rule 818 stores the reference information extraction rule 204 shown in FIG. 2, and the attribute extraction rule 819 stores the attribute extraction rule 403 shown in FIG. 4 and the attribute extraction rule 703 shown in FIG. In the word list 820, the word list 405 shown in FIG. 4 and the word list 704 shown in FIG. 7 are stored, and in the determination rule 821, the determination rule 106 shown in FIG. 1 is stored.

  According to the first embodiment of the present invention, it is necessary to generate a determination rule 106 for determining an important element in a document by obtaining a correspondence relationship between the case document 101 and the reference information description document 102. Can be efficiently constructed, and attribute information is extracted from the case document 101, and attribute values converted into numerical data are embedded in the teacher data. It is possible to easily construct a determination rule based on the size and range of numerical values such as “to” and configure an accurate important point determination system.

  For example, when it is desired to extract particularly important system information (CPU performance, memory size, etc.) from an enormous requirement specification presented when designing a computer system, according to the first embodiment. The determination rule 106 is generated based on the case documents such as the design documents accumulated in the past and the reference information describing the request from the user. As a result, system information can be efficiently extracted from the requirement specifications. Further, even when the numerical value indicated in the requirement specification is separated from the numerical value indicated in the past design document by the attribute value included in the determination rule 106, it can be extracted.

(Second Embodiment)
A second embodiment of the present invention will be described with reference to FIGS.

  In the first embodiment, the numerical data obtained from the attribute information extracted from the case document is embedded in the teacher data, thereby generating one type of determination rule 106, and further generating the one type of determination rule 106. Therefore, it is possible to determine an important point with high accuracy. On the other hand, the second embodiment is characterized in that a rule for determining a case document element excluding attribute information is separated from a rule for performing determination regarding attribute information.

  FIG. 9 is a block diagram showing the configuration of the important point determination system according to the second embodiment of the present invention.

  The difference between the configuration shown in FIG. 1 and the configuration shown in FIG. 9 is that the determination rule 106, which is one type in FIG. 1, is a document element determination rule for determining a case document element that does not include attribute information in FIG. There are two types of rule 903 and an attribute value determination rule 904 that performs only determination regarding attribute information.

  The configuration and function of the teacher data generation unit 901, the determination rule generation unit 902, the determination data generation unit 905, and the determination processing unit 906 are the same as the teacher data generation unit 104, the determination rule generation unit 105, and the determination data generation unit illustrated in FIG. 110 and the determination processing unit 111, but the details of the processing are different.

  In addition, the case document 101, the reference information description document 102, the document classification unit 103, the determination target document 107, the document division unit 201, and the document division rule 202 of the first embodiment are determined as important point determinations of the second embodiment. It also has a system.

  FIG. 10 is a block diagram illustrating a detailed configuration of the teacher data generation unit 901 according to the second embodiment of this invention.

  The teacher data generation unit 901 illustrated in FIG. 10 and the teacher data generation unit 104 illustrated in FIG. 4 are different in processing in the data conversion unit and output teacher data. In the second embodiment, the teacher data output from the data converter 1001 is document element teacher data 1002 and attribute value teacher data 1003.

  The document element teacher data 1002 is teacher data for generating a determination rule for a case document element that does not include attribute information from the teacher data in the first embodiment. The attribute value teacher data 1003 is teacher data for generating a determination rule for a case document element including attribute information.

  Similar to the first embodiment, in the second embodiment, a support vector machine is used to generate a determination rule from a case document element and to determine a determination target document element. The data conversion unit 1001 of the second embodiment assigns a multidimensional vector element to each word in the word list 1004, and assigns 1 to each case document element as a multidimensional vector element corresponding to the word in the case document element. Is set to 0 in addition to that, teacher data is generated.

  At this time, the data conversion unit 1001 removes the word in the case document element corresponding to the location extracted as the attribute information by the attribute information extraction unit 1005 in advance. Alternatively, the word corresponding to the attribute value is replaced with a specific character string that does not depend on the content of the word and does not appear as a word having another meaning while excluding the attribute name word. For example, the word corresponding to the attribute value may be replaced with “NNN” or the like.

  The data conversion unit 1001 outputs the multidimensional vector generated in this way as document element teacher data 1002. As a technique for generating the document element determination rule 903 from the document element teacher data 1002, in addition to the support vector machine, a generally used machine learning technique can be used. In this case, the document element determination rule 903 may be generated using a format suitable for machine learning to be used.

  Further, the attribute information extraction unit 1005 outputs a list of attribute information acquired by extracting the attribute name 502 and the attribute value converted into numerical data as attribute value teacher data 1003. As in the first embodiment, the attribute information extraction unit 1005 may extract the thing name having the corresponding attribute together with the attribute information, and output it by including it in the attribute value teacher data 1003.

  Note that each teacher data in the document element teacher data 1002 and the attribute value teacher data 1003 is appended with an identifier regarding whether the teacher data is generated from an important element or teacher data generated from an unimportant element. Is done.

  FIG. 11 is a block diagram illustrating details of the determination rule generation unit 902 according to the second embodiment of this invention.

  The determination rule generation unit 105 of the first embodiment is based on the premise that the determination rule 106 is generated by one type of machine learning. However, the determination rule generation unit 902 of the second embodiment has a document element determination rule 903. In addition, in order to generate two types of determination rules, an attribute value determination rule 904, a document element determination rule generation unit 1101 and an attribute value determination rule generation unit 1102 are provided.

  The determination rule generation unit 902 may generate a determination rule using a technique suitable for each in accordance with the format of the teacher data. For example, the method for generating the document element determination rule 903 can use a support vector machine when the teacher data is given by a multidimensional vector as described above. As a result, as in the first embodiment, the determination rule can be easily generated.

  On the other hand, in the case of the attribute value determination rule 904, since the teacher data is a set of attribute values converted into attribute names and numerical data, the determination rule generation unit 902, for example, sets attribute values of important elements for each attribute name. And generating a rule for determining whether or not an attribute value exists in a range determined by the average and variance. Also, a determination rule for calculating the probability of being an important element is generated by adding the average and variance of the non-important elements. Needless to say, a commonly used machine learning technique may also be used in generating a determination rule related to attribute information.

  FIG. 12 is a block diagram illustrating details of the determination data generation unit 905 according to the second embodiment of this invention.

  The difference between the determination data generation unit 104 illustrated in FIG. 7 and the determination data generation unit 905 illustrated in FIG. 12 is the processing in the data conversion unit 1201 and the determination data to be output. The determination data generation unit 905 illustrated in FIG. 12 outputs two types of document element determination data 1202 excluding attribute information and attribute value determination data 1203 related to attribute information.

  The word division unit 701, the attribute extraction rule 703, and the word list 704 in the second embodiment are the same as those in the first embodiment.

  When a support vector machine is used for the determination rule generation unit 902 and the determination processing unit 906, the data conversion unit 1201 is based on the content of the determination document element excluding the portion corresponding to the attribute information, similar to the document element teacher data described above. A multidimensional vector is generated as document element determination data 1202. Alternatively, the word corresponding to the attribute value may be replaced with a specific character string that does not depend on the content and does not appear as a word, while excluding the attribute name word. For example, it may be replaced with “NNN”.

  As a format of the document element determination data generated by the data conversion unit 1201, a format suitable for the technique used in the document element determination rule generation and the document element determination processing unit 1301 described later may be selected. Further, the determination data generation unit 905 performs the same processing as the teacher data generation unit 901. That is, the attribute information extraction unit 1204 extracts attribute information converted into attribute names and numerical data from attribute values, and outputs the extracted attribute information as attribute value determination data 1203. The extracted attribute information may be added together with the name of an object having the corresponding attribute and output as attribute value teacher data 1203.

  FIG. 13 is a block diagram illustrating details of the determination processing unit 906 according to the second embodiment of this invention.

  The determination processing unit 906 includes a document element determination processing unit 1301 and an attribute value determination processing unit 1302. The document element determination processing unit 1301 performs a determination process based on the document element determination rule 903 and the document element determination data 1202. The attribute value determination processing unit 1302 performs a determination process based on the attribute value determination rule 904 and the attribute value determination data 1203. As a technique used for each determination process, a method suitable for the technique used for generating the determination rule can be selected.

  For example, when the document element determination rule 903 is generated using a support vector machine as in the above-described example, the support vector machine may be similarly used for the document element determination processing unit 1301. Further, when the attribute value determination rule 904 is given by, for example, the average value and the variance value of the attribute values obtained from the case document, it is determined probabilistically using a well-known normal distribution formula. A method or the like may be used.

  The result integration unit 1303 integrates the determination result in the document element determination processing unit 1301 and the determination result in the attribute value determination processing unit 1302 and outputs the final determination result. For example, when the attribute information is extracted from the determination document element that is the determination target, the integration method may be a method in which only the result of the attribute value determination processing unit 1302 is used as the determination result.

  In addition, the result integration unit 1303 determines that the determination result from the document element determination processing unit 1301 and the determination result from the attribute value determination processing unit 1302 are important elements or non-important elements such as probability values that are important elements. When output is a continuous value indicating the degree, the results of both may be integrated by using a function that calculates one continuous value from both values, such as a weighted average or an arithmetic average of both determination results. .

  When the result from each of the document element determination processing unit 1301 and the attribute value determination processing unit 1302 is a result of only the determination as to whether it is an important element or a non-important element, the result integration unit 1303 By defining 1 as the case of 1 and 0 as the case of the non-important element, the results of both may be integrated using a function such as a weighted average or an arithmetic mean.

  The determination result output by the result integrating unit 1303 outputs the values obtained by the above-described method, and determines whether the determination result is an important element or a non-important element based on those values and a predetermined threshold value. The result may be output.

  As described above, according to the second embodiment of the present invention, it is necessary to generate a determination rule for determining an important element in a document by obtaining a correspondence relationship between a case document and a reference information description document. It is possible to efficiently construct the teacher data to be used, and to separate the part where the character string is important in the document from the part where the numerical value such as the attribute value is important, and the decision rule suitable for each. It is possible to configure a flexible and highly accurate important point determination system to which the generation technique and the determination processing technique are applied.

(Third embodiment)
A third embodiment of the present invention will be described with reference to FIGS.

  FIG. 14 is a block diagram showing a configuration of an important point determination system according to the third embodiment of the present invention.

  In the components in FIG. 14, the determination result correction unit 1401 and the classified case document 1402 are different from the important part determination system of the first embodiment. Other components are the same as those of the first embodiment.

  In the third embodiment, the output from the document classification unit 103 is stored in the classified case document 1402. The teacher data generation unit 104 refers to the contents of the classified case document 1402 and generates teacher data. In the third embodiment, a list of determination results is presented to the system user in the determination target document, and the system user confirms the determination results and provides means for correcting the erroneous determination results. It is. Further, after the system user corrects, a means for using the determination target document and the determination result as a new case to generate a determination rule is provided.

  The determination result correction unit 1401 illustrated in FIG. 14 presents a list of determination results received from the determination processing unit 111 to the system user, and also provides the system user with a means for the system user to correct the determination result. Specifically, a determination document element that is a result of dividing a determination target sentence into units to be determined and a determination result corresponding to the determination document element are displayed on a monitor screen, and by operating a keyboard or a mouse, on the screen, Enable to correct the judgment result.

  FIG. 15 is an explanatory diagram illustrating a display example of the determination result according to the third embodiment of this invention.

  The determination result list illustrated in FIG. 15 includes a determination target document 1501, a determination result 1502, a correction button 1503, an “add to case” button 1506, and an end button 1507.

  The determination target document 1501 displays the contents of each determination document element. In the determination handling document 1501, only the contents of the corresponding determination document element may be displayed, or may be displayed together with the context information. In addition, a button or the like may be arranged on the screen as an operation unit for switching the displayed content, and the displayed content may be switched by operating the button.

  In the determination result 1502, the determination result for the determination document element indicated by each determination target document 1501 is displayed. When a value indicating a flag is stored in the determination result 1502 of FIG. 15, the corresponding row indicates that it is determined as an important element, and a blank indicates that it is determined as a non-important element. That is, the display example illustrated in FIG. 15 indicates that it is determined that only the determination document element described in the determination target document 1501 in the lines 1504 and 1505 is an important element.

  As a method for displaying whether the determination document element is determined to be an important element or a non-important element, any method may be used as long as the two are distinguished from each other. For example, you may display by a different color or you may display by a different character.

  In the correction button 1503, a portion described as “correction” is a button for correcting the determination result of the determination document element of each determination target document 1501. When the system user operates this button with a keyboard or a mouse, the contents displayed in the determination result are displayed by switching between the display of the important element and the display of the non-important element each time the button is operated.

  For example, the determination result 1502 in the row 1504 displays the determination result that it is an important element. In this state, when the system user operates the correction button 1503 once, the determination result 1502 is switched to display of an unimportant element, and the column of the determination result 1502 in the row 1504 is blank. When the system user operates the correction button 1503 again, the determination result 1502 in the row 1504 switches to the display of the important element again, and the value indicating the flag is displayed.

  In the correction button 1503, in order to correct the determination result 1502, in addition to the above-described method using the button, a display method in which the system user directly inputs content indicating whether it is an important element or an unimportant element in the determination result column May be used. As described above, the determination result correction unit 1401 illustrated in FIG. 14 records all the portions of the determination target document 1501 in which the system user corrects the determination result 1502 by the system.

  When the system user operates the “add to case” button 1506 with the keyboard or the mouse, the determination result correction unit 1401 classifies the determination target element that has been corrected into an important element and / or an unimportant element as a classified case. It is added to the document 1402. The contents sent from the determination result correction unit 1401 are the same as the contents sent from the document classification unit 103 (case document elements classified as important elements and case document elements classified as non-important elements). Therefore, the case data element sent from the determination result correction unit 1401 can be used in the teacher data generation unit 104 by simply being added to the classified case document 1402.

  It is also possible to acquire content in a format that can be processed by the determination rule generation unit 105 from the determination processing unit 111. In this case, a storage area for storing the content output from the teacher data generation unit 104 is provided. The contents output from the determination result correction unit 1401 may be added to the storage area. In this case, since the word list generation process in the teacher data generation unit 104 is omitted, it is desirable that all words included in the determination target document are included in the case document.

  Further, the determination target element sent from the determination result correction unit 1401 to the classified case document 1402 or the like may be information regarding all combinations of determination target elements and determination results, or information regarding determination target elements corrected by the system user. You may send only. Alternatively, a combination of the determination target element determined as an important element by the system and only the determination target element corrected by the system user may be sent to the classified case document 1402. Furthermore, the system user may be able to select them.

  When the system user operates the “end” button 1507 shown in FIG. 15, the display of the determination result is ended without adding the determination target element to the classified case document 1402. Further, when the process is finished, information regarding the determination result and correction made by the system user may be separately stored.

  Furthermore, as a method of displaying the determination result, as shown in FIG. 15, in addition to the method of displaying for each determination document element that is a unit to be determined, the content of the determination target document is displayed as it is and it is determined as an important element. A method may be employed in which the determined location is displayed in a form that can be identified on the determination target document. An example of the display method in this case is shown in FIG.

  FIG. 16 is an explanatory diagram illustrating another display example of the determination result according to the third embodiment of this invention.

  The determination results shown in FIG. 16 include a result display screen 1601, an “add to case” button 1602, and an “end” button 1603.

  The result display screen 1601 displays the contents of the determination target document. By using the processing result in the document dividing unit 108, the context information on the determination document corresponding to each determination document element can be acquired. Therefore, the determination result for each determination document element is reflected in the context information of the determination document. Then, the result display screen 1601 shown in FIG. 16 displays the positions of the lines 1602 and 1603 corresponding to the determination document element determined to be an important element while changing the background color.

  The important element display method in FIG. 16 may be any method other than the method of changing the background color as long as the difference from the non-important elements can be easily identified. For example, it may be displayed by changing the color of the character, underlining or adding a frame.

  Further, in the method of correcting the determination result in the display method shown in FIG. 16, each time the mouse is double-clicked on the line corresponding to the determination document element to be corrected, the corresponding part is an important element or an unimportant element. It is possible to use a method of switching the display indicating whether or not there is. Alternatively, a means such as a correction button 1503 shown in FIG. 15 may be prepared, and the correction may be performed by operating the button after selecting a portion to be corrected using a keyboard or a mouse. Further, an “add to case” button 1604 and an “end” button 1605 shown in FIG. 16 have the same functions as the “add to case” button 1506 and the “end” button 1507 shown in FIG.

  As described above, according to the third embodiment of the present invention, the system user confirms the determination result by the system and provides means for easily correcting the determination result, and the result is used as a new case document. By adding, it becomes possible to construct a determination rule with high determination accuracy while using the system.

101 case document 102 reference information description document 103 document classification unit 104 teacher data generation unit 105 determination rule generation unit 106 determination rule 107 determination target document 108 document division unit 109 document division rule 110 determination data generation unit 111 determination processing unit 112 determination result 203 Reference information extraction unit 205 Matching unit 206 Classification processing unit 401 Word division unit 402 Attribute information extraction unit 404 Word aggregation unit 406 Data conversion unit 1401 Determination result correction unit

Claims (8)

A computer system comprising a processor for performing arithmetic processing and a storage device connected to the processor, wherein the processor analyzes a document,
The document includes a plurality of elements each including a plurality of words and constituting a sentence ,
The plurality of elements includes a sentence or a paragraph,
The processor is
A plurality of first documents, a reference document including a reference to the first document, and a second document to be determined as an important part are input;
Extracting the element from each of the first documents, extracting a reference location to the first document from the reference document as reference information;
Based on the similarity calculated by the elements extracted from the first documents and the reference information, the elements similar between the elements extracted from the first documents and the reference information are important. A location that is not similar between the element extracted from each first document and the reference information is regarded as a non-important location in the second element, and the first element is assumed to be a location. Split the elements extracted from the document,
Acquiring a first feature amount of each document based on the plurality of words included in the divided first element and second element;
Based on the acquired first feature value, a determination rule for determining whether or not the important part is included is generated,
Extracting the element from the second document;
Obtaining a second feature amount based on the plurality of words included in the element extracted from the second document ;
A computer that classifies elements extracted from the second document into important and non-important parts by comparing the generated determination rule and the acquired second feature amount. system. Wherein the plurality of words comprises the attribute name representing an attribute of the object indicated by the words contained in the document, the attribute value is a value of the attribute corresponding to the attribute name, as a word,
The computer system holds extraction information for extracting the attribute name and the attribute value,
The processor is
Extracting all the words indicating the attribute name and attribute value from the first element and the second element , respectively, by extracting locations that match the extraction information ,
Corresponds to information indicating whether or not all words indicating the extracted attribute name are included in each of the first element and the second element, and all words indicating the extracted attribute value And obtaining the first feature amount including a value calculated from the attribute value included in each of the first element and the second element,
Corresponding to information indicating whether or not all the words indicating the extracted attribute name are included in the second document and all words indicating the extracted attribute value, The computer system according to claim 1, wherein the second feature amount including a value calculated from the included attribute value is acquired. The plurality of words include an attribute name representing an attribute of a target indicated by a word included in the document, an attribute value that is an attribute value corresponding to the attribute name, and an article name having an attribute corresponding to the attribute name; As a word ,
The computer system holds extraction information for extracting the attribute name, the attribute value, and the article name,
The processor is
Extracting all the words indicating the attribute name, the attribute value, and the article name from the first element and the second element by extracting a portion that matches the extraction information ,
Information indicating whether or not all the words indicating the extracted attribute name and article name are included in each of the first element or the second element, and all indicating the extracted attribute value A first feature amount corresponding to a word and including a value calculated from the attribute value included in each of the first element and the second element;
Corresponding to information indicating whether or not all words indicating the extracted attribute name and article name are included in the second document, and all words indicating the extracted attribute value, the second The computer system according to claim 1, wherein the second feature amount including a value calculated from the attribute value included in the document is acquired. The processor is
Obtaining the first feature quantity and the second feature quantity as teacher data expressed in a vector format;
3. The computer system according to claim 2, wherein the determination rule is generated from the first feature amount using a support vector machine. The processor is
Obtaining the words whose meaning is similar to the plurality of words,
The computer system according to claim 2, wherein the attribute name includes the similar word. The computer system holds information indicating a plurality of other elements whose meanings are related to the plurality of elements ,
The processor identifies a location in the document where the plurality of other elements are included based on information indicating the plurality of other elements;
The meaning of the document is extracted by extracting the plurality of other elements, or the plurality of other elements and the plurality of elements having meanings related to the other plurality of elements, from the specified portion. Split into multiple related elements,
The first element includes a group of the plurality of elements including the important part ,
The computer system according to claim 1, wherein the second element includes a group of the plurality of elements not including the important part . Wherein the plurality of words comprises the attribute name representing an attribute of the object indicated by the words contained in the document, the attribute value is a value of the attribute corresponding to the attribute name, as a word,
The computer system holds extraction information for extracting the attribute name and the attribute value,
The processor is
By extracting a portion that matches the extraction information, the first element is divided into the third element that does not include the attribute name and attribute value, and the fourth element that includes the attribute name and attribute value. Divided into
By extracting a portion that matches the extraction information, the second element is divided into the fifth element that does not include the attribute name and attribute value, and the sixth element that includes the attribute name and attribute value. Divided into
Based on the third element and the fifth element, a third feature quantity including the feature quantity of the third element and the feature quantity of the fifth element is acquired,
Generating a first determination rule for determining whether or not a document that does not include the attribute name and attribute value includes the important part based on the acquired third feature amount;
Based on the fourth element and the sixth element, a fourth feature quantity including the feature quantity of the fourth element and the feature quantity of the sixth element is acquired,
Generating a second determination rule for determining whether or not a document including the attribute name and attribute value includes the important part based on the acquired fourth feature amount;
Extracting the element from the second document;
Dividing the element extracted from the second sentence into a plurality of seventh elements not including the attribute name and attribute value and a plurality of eighth elements including the attribute name and attribute value;
Acquiring the fifth feature amount based on the seventh element;
Obtaining the sixth feature amount based on the eighth element;
Comparing the generated first determination rule with the acquired fifth feature amount;
Comparing the generated second determination rule with the acquired sixth feature amount;
A comparison result of the first determination rule and the fifth feature amount, by Rukoto using both the comparison result of the second determination rule and the feature of the sixth, extracted from the second document The computer system according to claim 1 , wherein the elements are classified into the important part and the non-important part . The processor is
Connected with the output device,
Outputting the result of classifying elements extracted from the second document into important and non-important parts to the output device;
Obtaining a correction of the classification result extracted from the second document;
According to a result of classification of the second document and the correction, an element extracted from the second document is added to the first element or the second element, and the determination rule is generated. The computer system according to claim 1.

Images