JP2003330947A - Fact data extracting device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To construct database by automatically extracting multiple types of fact data from a lot of technical documents. <P>SOLUTION: This fact data extracting device is provided with a means 16 for inputting and storing a text 11 of the technical documents, a means 5 for inputting and storing technical terms and their category data 10, a means 1 for putting a tag on the technical document text using the technical terms, a means 2 for preparing an extract pattern of the fact data from the technical document text with the tag, and a means 3 for extracting the fact data from the technical document text to be the extract target of the fact data and the extract pattern by a matching operation. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for extracting fact data, and is particularly suitable for use in extracting fact data contained in technical documents such as papers, experimental reports, research reports and the like. is there.

[0002]

2. Description of the Related Art Fact data includes physical property values of materials such as
It means the characteristic value of a natural phenomenon obtained by experiment or observation. These fact data are converted into a database and used as common assets for humankind. Also, clinical data of patients in the medical field can be said to be fact data.

The creation of the database of the above fact data was basically done manually. Specifically, from technical documents such as papers, experimental reports, and research reports compiled by engineers and researchers,
Fact data has been manually extracted and stored in a database.

On the other hand, as a technique for extracting information from texts, there is a research for extracting merchandise information from newspaper articles, a service for retrieving travel information from the Internet, extracting places, costs, etc. and organizing and presenting them. Is being done.

[0005]

However, in the above-mentioned conventional technique, since the fact data is manually extracted, there is a problem that a great deal of labor is required for these operations.

The present invention has been made in view of the above problems, and an object of the present invention is to make it possible to automatically extract fact data from technical documents such as articles in specific fields, experimental reports, and research reports.

[0007]

A fact data extraction device of the present invention is a fact data extraction device for extracting fact data contained in a technical document from a text of a technical document, which corresponds to the technical document prepared in advance. Using the technical term dictionary file in the technical field, each sentence of the learning text file is tagged with respect to the extraction item, and a tagging means for creating a tagged file, and a tagged sentence selected from the tagged file, Extraction pattern match template creating means for creating a template for extracting the fact data to be extracted from the specified extraction information, and input the text to be the extraction target of the fact data, and perform matching using the created template. And the character string of the input sentence corresponding to the extraction item of the template is set as the solution candidate. Further subjected to restriction judgment candidate solutions, wherein the fact extracting means for extracting the fact data to include a database storage means for storing the fact data which the extracted database. Another feature of the present invention is that the tagging means classifies the terms in the text into categories of technical terms by means of morphological analysis incorporating technical terms to perform basic tagging, and based on the basic tags, It is characterized by combining words and tagging them. Another feature of the present invention is that the tagging means tags the structure of the text based on structure information of a sentence in the text. Another feature of the present invention is that the constraint determination of the solution candidate in the fact extraction means determines a constraint condition based on a morphological analysis result of the solution candidate. Another feature of the present invention is that the fact extraction means divides the extraction result into individual extraction items by dividing the solution candidates of the expression that spans a plurality of extraction items in determining the constraints of the obtained solution candidates. It is characterized by further having a function. According to the present invention configured as described above, a technical document file in which a technical document text is input,
You can use the technical term dictionary file that contains the technical terms and their categories to tag technical document texts with technical terms, and create a template pattern that extracts fact data from the tagged technical document texts. It becomes possible to do. Then, the fact data corresponding to the extraction item can be extracted by performing a character string matching operation between the technical document text that is the extraction target of the fact data and the extraction template pattern. The extracted fact data can be stored / stored as a fact database. Further, the tagging unit can classify the terms in the text into categories of parts of speech and technical terms and perform basic tagging by morphological analysis incorporating the technical terms. Furthermore, it becomes possible to synthesize and tag a compound word using the basic tag. Further, the tagging unit can tag the structure of the text by using the structure information of the sentence in the text. Further, the fact extracting means relates to part-of-speech information, character breaks, and technical term categories based on the result of morphological analysis from solution candidates obtained by performing a matching operation between a technical document text that is an object of fact data extraction and an extraction pattern. The data that satisfies the constraint is used as the extracted data. Further, the fact extracting unit can further divide the numerical value and the unit or the numerical value into the extracted data when performing the matching operation from the technical document text to be the extraction target of the fact data and the extraction pattern. According to the present invention, many fact data can be automatically extracted from technical documents such as papers, accumulated as a fact database, and used as basic data for diagnosis and control. Also, it becomes possible to store as a tagged file.

[0008]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of a device for extracting fact data according to the present invention will be described below with reference to the accompanying drawings. The target field of the fact data extraction device of the present invention is not particularly limited, but the medical field will be described below as an example. The fact data to be extracted are the data of the experiment subject, the experiment condition, and the result thereof.

FIG. 1 is a block diagram showing an example of the configuration of a medical fact data extraction device to which the fact data extraction device of this embodiment is applied. This device is realized by, for example, a personal computer.

In FIG. 1, reference numeral 1 denotes a tagging processing unit, which uses a technical term dictionary file 5 into which technical term data 10 has been input to perform tagging processing on a learning text file 6 into which learning text 11 has been input. Output attachment file 7.

The technical term data 10 is "(technical term ( for
Word categories ((entry word (the term name 1.0)) (read terms
Read )))) ”format. In the above, the technical term classification category is expressed as a term classification item. Technical terms are described in term names. 1.0 means weight.

More specifically, technical term data 10
Is, for example, "(Terminology (Disease ((
1.0)) (reading Nikgesh)))) ”format.

The file created by inputting the technical term data 10 in such a format is the technical term dictionary file 5. The technical term dictionary file 5 is created for each extraction target field. FIG. 2 shows an example of the term classification items in the medical field.

The tagging process in the tagging processing unit 1 is performed according to the process flow shown in FIG. In the tagging process 13 by morphological analysis using the technical term dictionary, the vocabulary registered in the technical term dictionary file 5 is searched for the character information of the learning text file 6, and when the registered vocabulary exists, A tag having a corresponding element name is added to each term classification item of the vocabulary.

Here, adding a tag means inserting a start tag and an end tag before and after a character string to be tagged. The start tag is the element name and attribute specification "<" and ">"
Enclosed in, and the end tag is the element name "</" and ">"
Expressed by enclosing in. For example, when expressed as "<element name attribute specification> tag target character string </ element name >", the start tag is <element name attribute specification> and the end tag is </ element name> Is.

As described above, the technical term dictionary file 5
The vocabulary registered in is searched, and morphological analysis is performed to determine whether the registered vocabulary exists. For the morphological analysis in this embodiment, "Japanese morphological analysis system JU
MAN (Kyoto University) "was used.

In the morphological analysis, a given sentence is decomposed into parts by the part of speech, and the part of speech name and the part of speech subclass classification name of each decomposed morpheme are obtained. The term classification item is obtained as a result. This information is used to do the basic tagging in the above format. For attribute specification, the appearance number of the element name is No = 1
It is expressed in the form.

Next, the tagging process 1 by analyzing the line structure 1
Do 4. As a text file for learning, the abstract paper collection is targeted. The abstract collection consists of multiple abstracts. One or more blank lines are inserted between each abstract.

Each abstract is composed of a plurality of sections such as "title", "author", and "text".
It is assumed that one or more blank lines are inserted between each section. Therefore, a structure tag is added to distinguish these sessions.

For example, in the abstracts of papers recorded in "Abstracts of the 58th Annual Meeting of the Japanese Society of Medical Radiology", structure tags as shown in FIG. 4 are attached to the entire abstracts, each abstract, and each section.

This structural tag is added by judging its type by the recognition and order of blank lines. Insert <doc>, <article>, <head>, <title> before the first non-blank line. Insert </ title> and <authors> in the next blank line. And on the next blank line </ authors>, </ head>, <body
Insert>. On the next blank line </ body>, </ article>, <art
Insert icle>. A structural tag can be added to the whole by repeating the following.

When a blank line is not inserted, the beginning can be detected by a term such as the article number and purpose, and a blank line can be inserted. Also, the body of the abstract can be divided into several description parts such as "purpose" and "result". For example, in the abstract of a paper recorded in "Abstracts of the 58th Annual Meeting of the Japanese Society of Medical Radiology", the text is four descriptions of "purpose", "object / method", "result", and "conclusion". It is divided into parts. Each content description section starts with a specific character string such as "[purpose]" according to the content.

A tag corresponding to the description content is added to each line of each description section. Each line ends with "." Or ".". A start character string for instructing the start of the content description part and a tag element name to be added to each line of the description content indicated by the character string are set by the "description content start instruction character" dictionary file (description content start dictionary file) 23 It shall be.

The description for each description content starts from the description content start instruction character string and continues until the next description content start instruction character string appears or until the abstract ends.

FIG. 5 shows an example of the "description content start character dictionary file". In the above example (FIG. 5), the tag element name added to each line of the “purpose” description part is “objective”,
The tag element name to be added to each line of the "target / method" description is "
"results" is set as the element name added to each line of the "methods" and "results" description sections, and "conclusions" is set as the tag element name added to each line of the "conclusion" description section.

In the tag combining process 15, a specific new tag is added to the set of the specified tag element and the appearance pattern of the specific character string. When adding a new tag, the conventional tag specified by the tag addition condition pattern shall be removed.

Specific examples are shown below. In the first example, the personal name (Yamada) and the personal name (Taro) are combined into the personal name (Taro Yamada). The same applies hereinafter.

Personal name Personal name + Personal name Personal name + "" + Personal name However, in the above, """" indicates a blank space (the same applies hereinafter).

Affiliation organization name + organization department + organization department + ... place name + "university", "large" place name + "hospital", "hospital" affiliation + ("" character string other than "(blank))

Disease name Disease name modifier + Disease name Disease name modifier + "Gender" + Disease name Disease name + Disease name ending Body organization + Disease name ending (Kanji character string (including alphabets)) + Disease name disease name + (Kanji character string (including alphabets) ) Is not a disease name

Diagnostic equipment Device modification + diagnostic equipment "ultra high speed" + diagnostic equipment Kanji (including alphabets) character string connected to "used equipment is"

Diagnostic method Diagnostic device + analysis method Diagnostic device + "by" + analysis method

Diagnostic Source Diagnostic Device + “Image” | “Image” Body Tissue + Diagnostic Device + “Image” | “Image” However, in the above, ““ | ”” means “or” (hereinafter the same. ).

Numerical value Numerical value + "±" + Numerical value Numerical value + "-" + Numerical value Numerical value + "-" + Numerical value "(" + Numerical value + ")" is not a numerical value

Report Item Statements ending with "Examine." In section describing objectives Statements ending with "Evaluate" in section describing objectives

In the extraction item tagging process 16, information such as the upper tag in which the tag appears, the preceding and following tags, and the character string is used to convert the tag for tagging process such as the term tag and the structure tag into the extraction item tag. By doing so, the extraction item tags are added to the following items among the items to be extracted from the fact data. 6 and 7 show a list of extraction items.

1. Paper number A numerical value that appears at the beginning of the title part of a paper abstract. 2. The title A string of characters that excludes the number that appears at the beginning of the title part of the abstract and the number that follows the number. 3. Presenter name "Person name" that appears in the "Author" section of the abstract.

4. “Affiliation” that appears in the “author” section of the abstract of the paper belonging to the presenter. 5. "Disease name" that appears in the "Title" and "Purpose" sections of the disease abstract. 6. "Body tissue" that appears in the "Target / method" section of the medical site abstract.

[0039] 7. "Disease name" that appears in the "Target / method" section of the medical treatment case abstract. 8. Number of cases treated A "numerical value" concatenated with the "case unit" that appears in the "Object / Method" section of the paper abstract. 9. The following character strings that appear in the "Target / Method" section of the age range paper abstract.
"Numeric value" + "Age unit" connected to "Age unit" +
“～” | “From” + “number” connected to “age unit” +
"Age unit".

10. The following character strings appearing in the “object / method” section of the medical practice date abstract. "Numerical value" + "Year / Month unit" +
"~" | "From" +. "Numeric value" connected to "year / month unit"
+ “Year / month unit”. 11. "Diagnostic device" that appears in the "Target / method" section of the medical device abstract. 12. "Diagnosis method" that appears in the "Title" section and the "Target / method" section of the medical treatment method paper abstract.

13. "Diagnosis Source" that appears in the "Target / Method" section of the medical data type paper abstract. 14. Data analysis method "Analysis method" that appears in the "object / method" section of the paper abstract. 15. Treatment Method The "treatment method" that appears in the "Title" and "Object / Method" sections of the abstract.

16. “Therapeutic drug” that appears in the “Object / Method” section of the therapeutic drug abstract. 17. A "numerical value" concatenated with the "dosage unit" that appears in the "Target / Method" section of the therapeutic dose abstract. 18. Therapeutic dosage unit The "dosage unit" that appears in the "Subject / Method" section of the paper abstract.

19. Radiation name "Radiation name" that appears in the "object / method" section of the abstract of the thesis. 20. A "numerical value" concatenated with the "dose unit" that appears in the "object / method" section of the radiation dose abstract. 21. Dose unit The "dose unit" that appears in the "object / method" section of the paper abstract.

22. “Frequency” connected to “frequency unit” appearing in the “object / method” section of the treatment frequency paper abstract. 23. A "numerical value" concatenated with the "period unit" that appears in the "object / method" section of the treatment period abstract. 24. Medical treatment classification The "medical treatment classification" that appears in the "Results" section of the abstract.

25. Medical treatment Gr classification "Medical treatment Gr classification" that appears in the "Results" section of the abstract. 26. Medical treatment Gr classification supplement "Medical Gr classification supplement" that appears in the "Results" section of the abstract. 27. Evaluation items A "diagnostic scale" that appears in the "Results" section of a paper abstract.

28. Evaluation value A "numerical value" concatenated with a "unit" that appears in the "results" section of the paper abstract. 29. Evaluation unit "Unit" that appears in the "Results" section of the abstract.

30. Total number of evaluation cases The following character strings appearing in the "Results" section of the abstract. M + “in the example” + N + “example”. M of N + “/” + M. However, in the above, M and N represent "numerical values."

31. Number of evaluation examples The following character strings appearing in the "Results" section of the abstract. M + “in the example” + N + “example”. N + “/” + M. N + N in "example". However, in the above, M and N represent “numerical values”.

32. “Supplemental Evaluation” in the “Result” section of the supplementary paper for evaluation. 33. "Evaluation" in the "Results" section of the evaluation paper abstract. 34. Conclusion Each line in the "Conclusion" section of the abstract. However, it does not include the start identifier of the "conclusion" part, and does not include the "." Or "."

35. “Disease name” that appears in the “Conclusion” section of the abstract of the subject paper. 36. "Diagnostic methods" and "treatment methods" appearing in the "Conclusion" section of method paper abstracts. 37. A specific character string that appears in the "Conclusion" section of the effect paper abstract. The character string is specified in the "result character string file". Examples of the character string specified in the “result character string file” include “useful”, “recognized”, “low”, “high” and the like.

In the tag shaping process 17, a tag unnecessary for the extraction pattern template creation process (process in the extraction pattern match template creation processing unit 2) is deleted. Information on tags to be formatted is "Tag formatting file"
Shall be specified. Also, the format of the "tag shaping processing file" is based on the XSL format. In addition, the tagged files are XML
It complies with. An example of the tag to be left finally is shown in FIGS.

Returning to FIG. 1, the extraction pattern matching template creation processing unit 2 automatically creates a template for extracting the fact data to be extracted. Describe the terms used to define the template.

The "extraction item" is a label showing the content of information to be extracted, and for example, "disease name", "medical treatment Gr classification", "evaluation item", "evaluation value", "evaluation unit", "evaluation". Examples include “number of cases” and “evaluation supplement”.

"Extraction information" is a character string (fact data) representing information corresponding to the extraction item. "pattern"
Is a character string of one or more characters that is the target of pattern matching. The “fixed pattern” is a characteristic character string that frequently appears in the extraction target, and examples thereof include “diagnose” and “image”.

The “wild card” is a symbol that can match any character string having a character string length of 0 or more in pattern matching.

The template is defined as follows. L
, Where P is a pattern and W is a wildcard, the template T is defined as in (Equation 1) below.

[0057]   T = C0L1C1L2 ... Cn-1LnCn   However, Ci = P0W1P1W2 ... Pm-1WmPm (i = 0 to n) ... (1 formula)

Note that C may be an empty character string. In addition, the template is created for each sentence, and C and L appear alternately. When expressing the template, the extraction items are expressed in the form of [item]. item is the extraction item name. Wild cards are represented by *.

The template is automatically created from the tagged template creating data. The procedure will be described using an example.

A sentence to be converted into a template is selected from the template creation data, and the extraction information is determined.

The case where the following sentence is selected will be described as an example. Example) “<SNB> malware tumor lesion </ SNB> is <SNS> 29 </ SNS>
<SNV> 14 </ SNV><SU> cases </ SU>(<SNA> 50% </ SNV)
SNA>). Here, as extraction information, SNB: medical treatment Gr classification, SNS: number of evaluation examples, SNV: evaluation value, SU: evaluation unit, SNA: evaluation supplement, etc. are set in advance.

The selected sentence is divided into words by using the morphological analysis system JUMAN. Example) “<SNB> Halmal tumor lesion </ SNB> is <SNS> 29 </
SNS> Out of cases <SNV> 14 </ SNV><SU> Cases </ SU>(<
SNA> 50% </ SNA>). "

The extraction information in the sentence is replaced with the extraction item. Example) "[Medical treatment Gr classification] was recognized in [Number of evaluated cases] [Evaluation value] [Evaluation unit] ([Supplemental evaluation])."

One word before and after the extraction information and a character string other than the fixed pattern are replaced with wildcards. as a result,
The template looks like this: Example) "[Medical treatment Gr classification] was found in [number of evaluated cases] [evaluation value] [evaluation unit] ([supplementary evaluation]) in the cases."

The template created by the above procedure is stored in the template file 8. When the input sentence is successfully matched with the template, the character string in the input sentence sandwiched between the patterns is extracted as the extraction information of the corresponding extraction item.

If there are a plurality of templates in the template set that match the sentence for which information is to be extracted, the information extracted will differ depending on the template set application method. Here, a method is adopted in which the templates are matched in a predetermined order, and the processing is ended when the matching is successful.

For that purpose, a template set arranged in order from a template having a high possibility of extracting a correct solution is created as follows.

Pattern matching is performed on each template using the extraction system, and information is extracted from all the sentences of the template creation data. Find the number of sentences that match the template, and sort the templates in order from the one with the smallest number.

Next, the fact extraction processing unit 3 using the template extracts the text of the fact extraction target (abstract paper).
Divide 12 by punctuation. The fact extraction processing unit 3 performs the fact extraction processing using the template shown in FIG. 10 on each of the divided sentences.

In FIG. 10, in the template changing process 18 by surface expression, the following process is performed. Change the notation of a specific item. Since the total number of evaluation examples and the number of evaluation examples can take only numerical values and “total”, they are replaced with the following notation pattern (equation (2)). Pattern of "total number of diagnosis cases" = ([0 1 2 3 4 5 6 7 8 9 all] +) (2 expressions) Express the relationship between items by notation. In the pattern of evaluation item-evaluation unit, medical treatment Gr classification-evaluation item-evaluation unit, a numerical notation is given to the evaluation item and a unit notation pattern is given to the evaluation unit.

Pattern matching process 1 by template
In 9, the matching processing of the input sentence and the template changed in the template changing processing 18 is performed. When the input sentence is successfully matched with the template, the character string in the input sentence sandwiched between the patterns is extracted as a solution candidate of the corresponding extraction item.

In the extraction result division processing 20, the extraction result is divided. The “medical care Gr classification”-“evaluation” pattern and the like also extract expressions that span multiple items. Therefore, the individual items are extracted by focusing on the “numerical value + unit” by the extraction result division processing 20.

For example, medical treatment Gr classification 0 is “bone lesion”, evaluation 0
Is extracted as “3 cases of sphenoid dysplasia and 1 case of occipital bone defect”, it is divided by “numerical value + unit”, and “sphenoid dysplasia” is evaluated item 0, “3” is extracted as an evaluation value 0, “case” is an evaluation unit 0, “occipital bone defect” is an evaluation item 1, “1” is an evaluation value 1, and “case” is an evaluation unit 1. The number after each extraction item name is a number that distinguishes the same extraction item.

In the constraint processing 22, constraint determination of solution candidates is performed based on morpheme information. When the sentence is entered, the above JUMA
Morphological analysis is performed using N, and delimiter information and part-of-speech information of each sentence are recorded. The following processing is performed for each extraction item for the solution candidates obtained by the template matching.

The word breaks are examined, and if the breaks before and after the solution candidate are incorrect, the solution candidate is excluded. Check the part-of-speech sequence of the solution candidate, and if the part-of-speech other than the specific part-of-speech is included, exclude the solution candidate. When the part-of-speech of a solution candidate is a technical term, the category is examined, and the non-target one excludes the solution candidate.

A specific example will be shown using the following input sentence examples and solution candidates extracted therefrom. As an example of the input sentence, "Examine the diagnostic ability of CT angiography in the evaluation of Takayasu arteritis."
I will give you.

It is assumed that, after morphological analysis is performed on the input sentence, four character strings shown in Table 1 below are extracted as disease name solution candidates by pattern matching processing.

[0078]

[Table 1]

When the sentence division is determined, the second solution candidate is excluded from the solution candidates because it ends in the middle of the word. When the part-of-speech arrangement is determined, the third solution candidate is excluded because it includes a verb (kaku).

When the part-of-speech arrangement is determined, the fourth solution candidate includes technical terms, but the category thereof is "diagnostic device", so it is excluded from the solution candidates. Through the above processing, the first solution candidate remains among the solution candidates.

On the other hand, if no extracted information is obtained as a result of the above processing, it is assumed that the information of that item does not exist. Then, the extracted information determined for each sentence is output.

The fact data DB storage processing unit 4 stores the contents of the fact data extraction result in the designated "medical fact database" (fact DB 9). An example of the extraction result stored in FIG. 11 is shown. In FIG. 11, the number after the symbol “!” Indicates a layer-intra-layer number, which is currently given in order.

The above "medical fact database"
Is realized by a commercial relational database. We conducted an experiment on information extraction of a medical fact data extraction subsystem, which extracts fact data, from abstract papers of the Japan Society of Medical Radiology.

As the accuracy evaluation items, seven items were selected from the disease name and the diagnosis result, and the extraction accuracy was verified. The result is shown in FIG. As shown in FIG. 12, recall (the number of correctly extracted information / the number of extracted items recorded in all the abstracts) and precision (the number of correctly extracted information / the number of extracted information) The accuracy seen is very high, which shows that the fact data extraction device of the present embodiment is very effective. As described above, the fact data can be automatically extracted from the technical paper by using the fact data extraction device of the present embodiment.

According to the present embodiment configured as described above, the technical document file into which the technical document text is input,
By using the technical term dictionary file in which technical terms and their categories are entered, technical document texts can be tagged with technical terms, and a template pattern for extracting fact data from the tagged technical document texts can be created. Can be created.

Then, with respect to the technical document text and the extraction template pattern from which the fact data is to be extracted,
The fact data corresponding to the extraction item can be extracted by performing the matching operation of the character string. The extracted fact data can be stored and accumulated as a fact database.

Further, the tagging processing section (tagging means) 1
Can classify terms in a text into categories of part-of-speech and jargon by morphological analysis incorporating jargon and perform basic tagging. In addition, basic tags can be used to combine and tag compound words.

Further, a tagging processing section (tagging means)
1 can tag the structure of the text by using the structure information of the sentences in the text.

Further, the fact extraction processing section (fact extraction means) 3 uses the morphological analysis result based on the morphological analysis result from the solution candidates obtained by performing the matching operation between the technical document text as the extraction target of the fact data and the extraction pattern. Extracted data are those that satisfy the constraint conditions regarding character delimiters and technical term categories.

Further, a fact extraction section (fact extraction means)
When the matching operation is performed from the technical document text to be the extraction target of the fact data and the extraction pattern, the numerical value 3 and the unit or the numerical value can be further divided to obtain the extracted data.

With the above configuration, many fact data can be automatically extracted from technical documents such as papers, accumulated as a fact database, and used as basic data for diagnosis and control. It can also be stored as a tagged file.

[0092]

As described above, according to the present invention, the text of the technical document is tagged by using the technical term, the extraction pattern of the fact data is created from the tagged text of the technical document, and the fact data is extracted. Since the fact data is extracted by the matching operation from the text of the technical document to be extracted and the extraction pattern created above, the specific fact data is automatically taken out from a large number of technical papers and technical documents. A database can be constructed and it can be used in various fields.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention and showing a configuration of a fact data extraction device.

FIG. 2 shows an embodiment of the present invention and is a diagram showing term classification items of a medical term specialized dictionary.

FIG. 3 is a flowchart illustrating processing in a tagging processing unit according to the embodiment of this invention.

FIG. 4 is a diagram showing an embodiment of the present invention and showing an example of a structure tag in an abstract paper collection.

FIG. 5 is a diagram showing an embodiment of the present invention and showing an example of a description content start instruction character dictionary file.

FIG. 6 is a diagram showing a list of extraction items and tags according to the embodiment of the present invention.

7 is a diagram showing an embodiment of the present invention and showing a list of extraction items and tags following FIG. 6. FIG.

FIG. 8 is a diagram showing a list of tags to be left finally according to the embodiment of the present invention.

9 is a diagram showing the embodiment of the present invention and showing a list of tags to be left finally, following FIG.

FIG. 10 shows an embodiment of the present invention and is a flowchart illustrating a fact extraction process using a template in a fact extraction processing unit.

FIG. 11 is a diagram showing an embodiment of the present invention and showing an example of an extraction result of fact data.

FIG. 12 is a diagram showing an embodiment of the present invention and showing an example of extraction precision of fact data.

[Explanation of symbols]

1 Tagging Processing Section 2 Extraction Pattern Match Template Creation Processing Section 3 Fact Extraction Processing Section 4 Fact Data DB Storage Processing Section 5 Technical Term Dictionary File 6 Learning Text File 7 Tagged File 8 Template File 9 Fact Database (Fact DB) 10 Technical Term Data 11 Learning Text 12 Target Text for Fact Extraction 13 Tagging Processing by Morphological Analysis Using Technical Term Dictionary 14 Tagging Processing by Line Structure Analysis 15 Tag Compositing Processing 16 Tagging Extracted Items 17 Tags Shaping processing 18 Template change processing by surface expression 19 Pattern matching processing by template 20 Extraction result division processing 21 Morphological analysis processing 22 Constraint processing 23 Description content start dictionary file 90 Fact data

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Yuiko Soumura             3-3-Minato Mirai, Nishi Ward, Yokohama City, Kanagawa Prefecture             1 Nippon Steel Solutions Co., Ltd. Sys             Tem R & D Center (72) Inventor Tatsuya Iwashita             3-3-Minato Mirai, Nishi Ward, Yokohama City, Kanagawa Prefecture             1 Nippon Steel Solutions Co., Ltd. Sys             Tem R & D Center F-term (reference) 5B009 NA05 VA02 VA09 VC01                 5B075 ND03 ND23 NK02 NK10 NK32                       NS10 UU06 UU26

Claims (5)

[Claims] 1. A fact data extraction device for extracting fact data contained in a technical document from a technical document text, which is prepared by using a technical term dictionary file of a relevant technical field of the technical document prepared in advance for learning. Tagging each sentence of the text file with respect to the extraction item to create a tagged file, the tagged sentence selected from the tagged file, and the fact data to be extracted from the specified extraction information Extraction pattern matching template creating means for creating a template for inputting the text to be the extraction target of the fact data, performing matching using the created template, the characters of the input sentence corresponding to the extraction item of the template The columns are set as solution candidates, the constraint judgment of the solution candidates is performed, and the fact data is A fact data extraction device comprising: a fact extraction means for extracting; and a database storage means for storing the extracted fact data in a database. 2. The tagging means classifies the terms in the text into categories of technical terms by means of morphological analysis incorporating technical terms to perform basic tagging, and synthesizes them into compound words based on the basic tags to tag them. The fact data extraction device according to claim 1, wherein the fact data extraction device is attached. 3. The fact data extraction device according to claim 1, wherein the tagging unit tags the structure of the text based on structure information of a sentence in the text. 4. The constraint determination of the solution candidate in the fact extraction means determines a constraint condition based on a morphological analysis result of the solution candidate, according to any one of claims 1 to 3. Fact data extraction device. 5. The fact extraction means further has an extraction result division function of dividing the solution candidates of the expression spanning a plurality of extraction items into individual extraction items when determining the constraint of the obtained solution candidates. The fact data extraction device according to any one of claims 1 to 4, which is characterized.