JP2009258870A - Information extraction device and program, and information extraction rule generation device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To perform high-accuracy information extraction from a document in an individual field, e.g., a medical document such as an image reading report. <P>SOLUTION: A medical term extraction part 11 extracts a medical term by use of an extraction rule or a medical vocabulary dictionary stored in a medical vocabulary resource holding part 12. An initial rule application part 14 takes out a medical term relation extraction rule stored in an initial rule holding part 13, applies it to the medical document, and extracts a vocabulary of a TARGET portion in each rule. An expansion rule creation part 15 acquires a combination between an acquired medical category and the vocabulary having some certain frequency or above among the vocabularies, and extracts an expansion rule. An information extraction part 17 performs the information extraction by use of a set of an initial rule stored in the initial rule holding part 13 and the expansion rule held in an expansion rule holding part 16. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to an information extraction technique for extracting linguistic knowledge from a document in a specific field, such as an electronic medical record or an interpretation report in the medical field, and an information extraction rule generation technique.

  In the medical information field, methods for automatically extracting linguistic knowledge such as ontology from medical records and interpretation reports are being actively researched in order to make use of electronic medical records and automatic interpretation report support for system development. Extraction of information in a specific domain (hereinafter also referred to as an individual field or a specific field) requires a device specific to the domain. Information extraction technology in the medical field is being actively researched now that medical documents are electronically stored on computers. The purpose of information extraction is to extract broader terms and lower-word pairs from medical documents to help build ontologies, extract the disease and findings from medical documents, and make inference knowledge from disease to findings There are things that try to get. Non-Patent Document 1 proposes information extraction in a domain specific situation in the medical field. Further, it has been proposed to extract information by using a plurality of rules corresponding to the extraction location for the purpose of extracting opinions from users regarding products outside the medical field (Patent Document 1, etc.). In the proposal of Patent Document 1, four slots are considered for a sentence expressing an opinion, and if there is no keyword corresponding to the extracted part, an extraction target omitted from other extracted keywords is estimated. The device is equipped with various rules.

  In the method of extracting broader terms and narrower terms from the actual medical document such as interpretation report, and the information extraction method such as obtaining the pair of disease and finding part from the parsing result, the medical treatment spans multiple sentences. There was a problem that extraction patterns and methods could not be applied to document descriptions.

  In a medical document such as an interpretation report, there are cases where a part describing the characteristics of an affected part and a part describing a doctor's judgment are divided into a plurality of sentences. Or, even if it is described in one sentence, the sentence structure may be complicated. If you want to extract the relationship between the noun of the disease name that the doctor made a decision and the description about the findings from the sentences in the interpretation report such as the following, the method using existing patterns and syntax analysis results is supported Can not.

[Sentence A]
--- Lymph node-like nodules are seen in the left hilar region. This change is not seen in previous CT. The possibility of metastatic lymph nodes is high. ----

  In the existing system, after annotating (annotating) a medical vocabulary with respect to each sentence in the text, a set of a disease and a finding is extracted using a pattern or a parsing result. In the technique according to Non-Patent Document 1, a verb representing a doctor's claim representing a finding part is detected, medical knowledge is detected from a subtree of a series of medical noun parts related to the verb, and a disease name obtained by another method Associate with. However, with these existing extraction methods, knowledge cannot be extracted from descriptions that straddle sentences. Therefore, it cannot cope with the extraction of the relationship such as “I think lymph node-like nodules are metastatic lymph nodes” as described in the above text. Further, it is assumed that the sentence A is converted into the following form.

[Sentence B]
---- Lymph node-like nodules are seen in the left hilar region, a change not seen in previous CT, and the possibility of metastatic lymph nodes seems high. ----

  However, in the above case, a plurality of verbs exist in one sentence, the syntax structure becomes complicated, and it is difficult to extract knowledge by pattern. If the doctor insists that "probably likely" is the part that the doctor claims, that part and "the lymph node-like nodule is seen" will be a juxtaposed structure, and a different pattern will be used for the process of extracting the flow of arguments It is necessary to describe many. In order to extract information from a medical document described in various sentences by a doctor, it is considered that some device specific to the medical document is necessary.

  For example, a technique for extracting medical knowledge from a complicated structure that exists across a plurality of sentences is desired.

Note that the above-described prior art and its problems are described only to explain a part of the background of the present invention. It should be noted that the present invention is not limited to the above-described prior art and problems.
JP 2004-157841 A Takeshi Imai, Eiji Aramaki, Masayuki Hagio, Kengo Miyo, Kazuhiko Oe, “Attempt to construct ontology of diagnostic imaging findings using syntactic information and medical term attributes”, Medical Informatics, 25 (6), 2005, pp. 395-403

  An object of the present invention is to perform highly accurate information extraction from a document in an individual field, for example, a medical document such as an interpretation report.

  According to this invention, in order to achieve the above-mentioned object, the configuration as described in the claims is adopted. Here, prior to describing the invention in detail, supplementary explanations of the claims will be given.

  That is, according to the first aspect of the present invention, in order to achieve the above-mentioned object, the information extracting device includes: an individual field vocabulary specifying means for specifying a vocabulary used in an individual field from a document in the individual field; Rule applying means for applying the first rule to the vocabulary used in the individual field specified in the individual field document and extracting the relationship between the vocabularies; and the rule applying means for each first rule From the vocabulary forming a predetermined term or the set of categories of the vocabulary included in each of the plurality of relationships extracted in step 1, the extraction of the first rule is applied as an activation condition to extract a new relationship between the vocabularies. Rule creating means for creating a second rule; information for extracting a vocabulary relationship from a document in the individual field by using the first rule and the second rule together Out it is provided and means.

  According to the invention of claim 2, in the invention of claim 1, the first rule and the second rule extract a relationship between two vocabularies.

  According to the invention of claim 3, in the invention of claim 2, the first rule includes a first part and a second part, immediately before the first sentence element satisfying the first part. The first rule is satisfied when a second sentence element that satisfies the second part appears in the first sentence element, and the vocabulary forming the first term in the first sentence element and the second sentence element in the second sentence element The relationship between the vocabulary that constitutes the first term is extracted, and the second rule is the first part of the corresponding first rule, and the vocabulary that constitutes the first term or the category of the vocabulary. The first part of the corresponding first rule, including the one containing the one, the first part of the corresponding vocabulary or the other category of the vocabulary. 2 part and triggered when the corresponding first rule is satisfied, When the second sentence element satisfies the first part of the second rule, and the third sentence element that satisfies the second part of the second rule appears immediately before the first sentence element, The relationship between the vocabulary forming the one term in the second sentence element and the vocabulary forming the predetermined one term in the third sentence element is extracted.

  According to a fourth aspect of the present invention, in the first, second or third aspect of the invention, the individual field is a medical field, and the first rule is a vocabulary indicating that a medical condition is determined and the determination Including a rule for extracting a relationship between a vocabulary that forms a basis, and the second rule is based on a rule for extracting a relationship between a vocabulary representing that the medical condition is determined and a vocabulary that forms the basis of the determination. It is created from a vocabulary that forms the basis for the above-described determination of a plurality of extracted relationships or a set of vocabulary categories of the vocabulary.

  Further, according to the invention of claim 5, in the invention of claim 1, 2, 3 or 4, the vocabulary or category of the vocabulary used to create the second rule for each of the first rules. Is limited to a vocabulary or a category of the vocabulary whose frequency extracted by applying the first rule exceeds a predetermined threshold.

  According to a sixth aspect of the present invention, there is provided a computer: an individual field vocabulary specifying means for specifying a vocabulary used in an individual field from an individual field document; one or more first rules in the individual field document Rule applying means for extracting a relationship between vocabularies by applying to the specified vocabulary used in the individual field; a predetermined number included in each of the plurality of relationships extracted by the rule applying means for each of the first rules Rule creating means for creating a second rule for extracting a new relationship between words by applying the extraction of the first rule as a trigger condition from a vocabulary forming one term or a set of categories of the vocabulary; Information extraction for functioning as information extraction means for extracting vocabulary relationships from documents in the individual fields by using both the first rule and the second rule To implement the program.

  According to the invention of claim 7, the information extraction rule generation device includes: individual field vocabulary specifying means for specifying a vocabulary used in an individual field from an individual field document; and one or more first rules as the individual rule A rule applying means for extracting a relationship between the vocabularies by applying to the vocabulary used in the individual field specified in the document of the field; a plurality of relationships extracted by the rule applying means for each of the first rules; A second rule for extracting a new relationship between vocabularies is created by applying the extraction of the first rule as a triggering condition from a vocabulary forming a predetermined term included in each or a set of categories of the vocabulary. Rule creation means is provided.

  According to the invention of claim 8, a computer: an individual field vocabulary specifying means for specifying a vocabulary used in an individual field from an individual field document; one or more first rules in the individual field document Rule applying means for extracting a relationship between vocabularies by applying to the specified vocabulary used in the individual field; a predetermined number included in each of the plurality of relationships extracted by the rule applying means for each of the first rules From the vocabulary constituting one term or the set of categories of the vocabulary, the extraction of the first rule is applied as an activation condition to function as rule creation means for creating a second rule for extracting a new relationship between the vocabularies. An information extraction rule generation program is realized.

  These and other aspects of the invention are set forth in the appended claims and will be described in detail below with reference to examples.

  According to the first and sixth aspects of the present invention, the extraction of the first rule is applied as the trigger condition from the relationship extracted by providing the first rule to the document in the individual field, and a new relationship between the vocabularies is established. A second rule to be extracted is created to enable highly accurate information extraction from the individual field document. Note that the new relationship between vocabularies extracted by applying the second rule refers to a relationship newly extracted by applying the second rule, which is a rule different from the first rule. It does not indicate that the relationship itself is new.

  According to the inventions of claims 2 and 3, information can be extracted from a simple relationship called a relationship between two vocabularies. In the case where the second rule applied with the first rule as an activation condition is composed of a single individual rule, the number of the first rule and the second rule is N1 and N2, and the number of N1 × N2 However, in the invention of claim 2, the number of rules is N1 + N2, and the number of rules is suppressed.

  According to the invention of claim 4, the medical condition / causal relationship between the symptom and the basis of judgment can be extracted from the document in the medical field, and information for supplementing it can be further extracted by the second rule.

  According to the invention of claim 5, the second rule can be extracted reflecting the statistical information of the actual document.

  According to the inventions of claim 7 and claim 8, the extraction of the first rule is applied as a trigger condition from the relationship extracted by providing the first rule to a document in an individual field, and a new relationship between vocabularies is established. A second rule to be extracted can be created.

  Hereinafter, an embodiment in which the present invention is applied to an information extraction system for extracting information from a medical document will be described.

  In the following, an interpretation report is assumed as a medical document, but the present invention is not limited to this. The text in the interpretation report is basically divided into a sentence that describes radiographic features such as MRI and CT of the affected area and describes the findings, and a sentence that describes the judgment of the doctor. There is a medical term in each sentence. Information extraction in a medical document is generally to extract a relationship between medical terms or between a medical term and a general term such as a verb.

  In this embodiment, a medical domain dictionary and a plurality of medical vocabulary extraction rules are used in combination to expand an existing medical vocabulary relationship extraction rule, and the relationship between medical vocabulary existing in a plurality of sentences or sentences of complex structures. To be able to extract.

  In this embodiment, the relationship between medical vocabularies is extracted from a large amount of medical documents by an existing method, and the original extraction rule is extended based on the obtained statistical information. The expanded rule can extract the relationship between medical vocabulary even from a more complex sentence or multiple sentences than the original rule. In the following, with reference to an embodiment, an exemplary procedure for extending a rule according to the present invention will be described together with a typical form of the present invention.

  FIG. 1 shows an information extraction system 100 according to an embodiment of the present invention. This information extraction system 100 is typically realized by a computer system, for example, a personal computer 200. As is well known, the personal computer 200 includes a CPU, a main memory, an external storage device, a bus, various input / output devices, and the like, and the information extraction system 100 is constructed by installing the program 201 therein. The information extraction system 100 may be constructed from a single computer system or may be constructed from a plurality of computer systems connected via a network.

  In FIG. 1, an information extraction system 100 includes a medical term extraction unit 11, a medical vocabulary resource holding unit 12, an initial rule holding unit 13, an initial rule application unit 14, an extended rule creation unit 15, an extended rule holding unit 16, and an information extraction unit. 17 etc. are comprised.

  The medical term extraction unit 11 extracts a medical vocabulary in a medical document using the medical care vocabulary stored in the medical vocabulary resource holding unit 12 and its extraction rule. Medical vocabulary categories (also referred to as types) include disease names, site names (anatomies), disease states (semi-disase), areas (areas), nouns (generals) frequently used in interpretation reports, Examples include affixes that appear characteristically in medical vocabulary such as mon and sei. The following is an extraction example. The portion surrounded by the tag portion corresponds to the medical vocabulary, and the medical vocabulary category is distinguished by the tag name. Tags are indicated by “<” and “>”. The following example shows that a vocabulary between tags in an example sentence is extracted as a medical vocabulary.

[Example]
<Area> left </ area><mon> hilar </ mon><area> part </ area> has <anatomy> lymph node-like </ anatomy><anatomy> nodule </ anatomy>. <General> change </ general> not seen in previous tests. <Sem-disase> metastasis </ semi-disase><anatomy> lymph node </ anatomy><sei> possibility </ sei> seems to be high.

  The medical vocabulary resource holding unit 12 stores a medical vocabulary dictionary and extraction rules used by the medical term extraction unit 11. The medical term extraction unit 11 extracts medical vocabulary in a medical document using these dictionaries and rules.

  The initial rule holding unit 13 stores medical term relationship extraction rules (also referred to as initial rules). The medical term extraction unit 11 applies these rules to a document to extract a relationship between medical vocabulary or sentence parts. For example, the relationship between the “part representing the judgment that the doctor is a medical condition” and “the cause” are extracted according to the following rules. The rule is composed of two parts, a part for detecting “a part representing a doctor's judgment” and a part for detecting “a part of the cause”. The initial rule holding unit 13 stores such rules in advance.

  An example of a disease state cause extraction rule is as follows. In an actual implementation, the program corresponding to the initial rule application unit 14 reads and interprets the following lines for processing.

[Example of information cause extraction rule]
Extraction unit: {<medical-term>. +? {Seen | some} | {seen | some} <medical-term>}
@ TARGET = <medica-term>
Condition part: {<Semi-disease><general> | <Semi-disease>}. +? {Think | think}}

  In this example, processing is executed as follows.

[Description of information cause extraction rule example]
"[<Semi-disease><general>]" or [Semi-disease] is followed by [[Thinking | thinking]], or immediately before that "[seen->thinking]" The medical vocabulary part to be extracted is taken out as TARGET. The TARGET part is extracted as a “cause” of Semi-dissease. At this time, the part of the rule including TARGET and “[See | The part including the judgment of the doctor is set as the condition part of the rule.

  This initial pathological cause extraction rule cannot be extracted well when a complicated sentence or a plurality of sentences describe “relationship between doctor's judgment and its grounds”. For example, in the following sentence A, the relationship between “lymph node-like nodules” and “metastatic lymph nodes” stops applying the rule only with the sentence “changes not seen in previous CT”. Therefore, it cannot be extracted. In the above rule, TARGET = “change” is only extracted as the cause of the disease state “metastatic lymph node”. Furthermore, as assumed in many existing information extraction systems, even when applying a rule to a single sentence, even the rule cannot be applied.

[Sentence A]
--- Lymph node-like nodules are seen in the left hilar region. This change is not seen in previous CT. There seems to be a high possibility of metastatic lymph nodes ----

  In addition, it is not realistic to describe a rule that covers all situations including the above case. Therefore, in this embodiment, in addition to the initial rule application unit 14, the extended rule creation unit 15 and the extended rule holding unit 16 are used to extend the rules.

  The initial rule application unit 14 extracts the medical term relationship extraction rule stored in the initial rule holding unit 13 and applies it to the medical document, and extracts the vocabulary of the TARGET part for each rule. At this time, the type of medical vocabulary at the extraction location is acquired simultaneously. When the disease state cause extraction rule is applied to the sentence A, “change” is acquired as a TARGET with a medical category <general>. Such extraction is performed on the entire medical document, and the type of medical vocabulary obtained as TARGET is acquired. At this time, the surface layer is acquired if it appears with a frequency exceeding a certain threshold value, and the type of medical vocabulary is acquired if it falls below the threshold value. For example, “change”, which is a noun frequently appearing in the interpretation report, belongs to the medical vocabulary category <general> in this embodiment, and “change” obtains a surface character string. On the other hand, “frontal white matter” and the like are acquired as <situ> in the medical category. In medical documents, compound words with various forms of affixes appear, and even words with the same meaning appear in various forms. The vocabulary whose surface vocabulary fluctuates is acquired as the medical category itself.

  The extended rule creation unit 15 acquires a combination of the acquired medical category and vocabulary that has a certain frequency or higher. Furthermore, an extended rule is extracted from the vocabulary combination according to the following procedure.

[Procedure 1]: A vocabulary or medical category with a certain frequency or higher is acquired.
For example, the site <Anatomy>, features, changes, findings,... Are acquired.
[Procedure 2]: Create the above vocabulary set.
For example, a site <Anatomy> -feature / site <Anatomy> -change / site <Anatomy> -finding /... Is created.
[Procedure 3]: Create a rule that matches the situation like the condition part of the initial rule with two vocabularies in succession, so that one of the vocabulary sets is the rule extraction part and the other is the condition part. A new disease cause expansion rule is created from the set. For example, the following rule is created from a set of part <Anatomy>-“change”. An example of the operation for creating the disease cause expansion rule will be described later with reference to FIG.

  Examples of the disease cause expansion rules are as follows. In actual implementation, the program corresponding to the information extraction unit 17 reads and interprets the following lines for processing.

[Example of disease cause expansion rules]
Extraction unit: {{seen | some} {<Anatomy>} | {<Anatomy>}. +? {Seen} |
@ TARGET2 = <Anatomy>
Condition part: {{seen | already} {change} | {change}. +? {Seen} |

[Explanation of examples of disease cause expansion rules]
Immediately before the part that constitutes the [modification] related to “[seen | already]” or the constituent modification clause, the site <Anatomy> is related to “[seen | always]” or constitutes the structural modification clause When there is a part, the part of “part <Anatomy>” is taken out as TARGET2.

  The above-mentioned disease state cause expansion rule matches a situation in which a set of vocabulary appears continuously with a predicate of “seen” or “present”. By using the pathological cause extraction extended rule and the pathological cause extraction initial rule in combination, if the vocabulary representing the cause of the medical condition appears in multiple locations, the vocabulary that causes the medical condition is extracted simultaneously. be able to. The extended rules as described above are stored in the extended rule holding unit 16.

  The information extraction unit 17 performs information extraction using a set of the initial rule stored in the initial rule holding unit 13 and the extended rule held in the extended rule holding unit 16. The extended rule is applied when the relationship is extracted by the initial rule. For example, when the initial rule and the extended rule are applied to the sentence A in the above process, first, “metastasis” is obtained as “disease state” <semi-dissease> by the initial rule, and the doctor's judgment on the part is shown. As a vocabulary, “change” is obtained. Furthermore, if the extended rule is used, a “lymphoid nodule” is obtained corresponding to the “change”. Therefore, a group of “lymph node nodule, change” is obtained as a judgment part of the doctor for “metastasis”. It is possible to extract information up to the “lymph node-like nodule” portion that could not be extracted.

  In this embodiment, the number of rules can be reduced. For example, the rule provided by this embodiment includes a predicate that there are 1000 types of rules for extracting “parts judged to be medical conditions by a doctor” and a “part name-change” pair is “seen | If there are 1000 types of rules for extracting situations such as being linked together, the existing rule is linked in association with the predicate that “part name-change” is “seen | In order to match a situation where “the part judged to be” is continuous, 1000 times as many rules as 1000 are required. On the other hand, the information extraction system according to this embodiment can match with 2000 types of rules.

  FIG. 2 shows an example of the overall operation of the above embodiment, and FIG. 3 shows an example of the operation relating to the rule generation procedure.

  The overall operation example illustrated in FIG. 2 is as follows.

[Step S21]: A medical vocabulary in a medical document is detected.
[Step S22]: The initial rule is applied to the medical document, and the medical term relationship for each rule is extracted.
[Step S23]: An extended rule is created for each initial rule and stored for each initial rule.
[Step S24]: Information extraction processing is performed using the initial rule and the extended rule. Extended rules are applied when the corresponding initial rules are met.

  An operation example of the rule generation procedure illustrated in FIG. 3 is as follows.

[Step S31]: It is confirmed whether or not the processing has been completed for all the initial rules. If the processing has been completed, the processing is terminated. If the processing has not been completed, the processing from step S32 is repeated.
[Step S32]: One initial rule is extracted unprocessed.
[Step S33]: The initial rule is applied to the medical document, the vocabulary and its medical category are extracted for each rule, and the frequency is counted.
[Step S34]: It is determined whether or not the vocabulary frequency is equal to or higher than a threshold value. If the vocabulary frequency is equal to or higher than the threshold value, the process proceeds to step S36, and if not, the process proceeds to step S35.
[Step S35]: The medical category is added to the vocabulary set. Proceed to step S37.
[Step S36]: The vocabulary is added to the vocabulary set. Proceed to step S37.
[Step S37]: Create a pair of vocabulary and medical category in the vocabulary set.
[Step S38]: It is determined whether the frequency of the pair is equal to or higher than a threshold value. If the frequency is equal to or higher than the threshold value, the process proceeds to step S37.
[Step S39]: An additional rule is created from the pair. Then, it returns to step S31 and repeats a process.

  In addition, this invention is determined based on description of a claim, and is not limited to the specific structure of the Example, a subject, and an effect. The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention. For example, in the above embodiment, domain specific information in the medical field is extracted. However, if a domain specific vocabulary in another field or an initial rule for extracting a relationship is prepared instead of the vocabulary in the medical field, that field is prepared. The present invention can be applied to the extraction of information. Further, the present invention may be realized as an apparatus or program for generating rules. For example, the rule generation device can be configured from a portion excluding the information extraction unit 17 of FIG.

It is a block diagram which shows the structure of the Example of this invention as a whole. It is a flowchart which shows the operation example of the above-mentioned Example as a whole. It is a flowchart which shows the operation example of extended rule preparation of the above-mentioned Example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Medical term extraction part 12 Medical vocabulary resource holding part 13 Initial rule holding part 14 Initial rule application part 15 Extended rule creation part 16 Extended rule holding part 17 Information extraction part 100 Information extraction system 200 Personal computer 201 Program

Claims (8)

Individual field vocabulary identifying means for identifying vocabulary used in individual fields from documents in individual fields;
Rule applying means for applying one or more first rules to the vocabulary used in the individual field specified in the individual field document, and extracting a relationship between the vocabularies;
For each of the first rules, the extraction of the first rule is applied as a trigger condition from a vocabulary forming a predetermined term or a set of categories of the vocabulary included in each of a plurality of relations extracted by the rule applying means. A rule creating means for creating a second rule for extracting a new relationship between the vocabulary and
An information extraction apparatus comprising: an information extraction unit that extracts a vocabulary relationship from a document in the individual field by using the first rule and the second rule together.   The information extraction apparatus according to claim 1, wherein the first rule and the second rule extract a relationship between two vocabularies.   The first rule includes a first part and a second part, and a second sentence element that satisfies the second part appears immediately before the first sentence element that satisfies the first part. The first rule is satisfied, and a relationship between the vocabulary forming the first term in the first sentence element and the vocabulary forming the first term in the second sentence element is extracted, and the second The first rule includes the first part of the corresponding first rule, the first part including one of the vocabulary forming the one term or the category of the vocabulary, and the corresponding first When the first part of the rule includes a vocabulary that constitutes one of the terms or one of the categories of the vocabulary as a second part, and the corresponding first rule is satisfied And the second sentence element satisfies the first part of the second rule, When the third sentence element that satisfies the second part of the second rule appears immediately before the first sentence element, the vocabulary forming the first term in the second sentence element and the third sentence element The information extracting device according to claim 2, wherein a relationship between a vocabulary forming a predetermined term in the sentence element is extracted.   The individual field is a medical field, and the first rule includes a rule for extracting a relationship between a vocabulary indicating that it is determined as a medical condition and a vocabulary as a basis for the determination, and the second rule includes A vocabulary that forms the basis for the above determination of a plurality of relationships extracted from a rule that extracts the relationship between the vocabulary that represents the medical condition and the vocabulary that forms the basis for the determination, or a set of vocabulary categories of the vocabulary The information extraction device according to claim 1, 2 or 3, which is created from   For each of the first rules, the vocabulary or category of the vocabulary used to create the second rule is a vocabulary whose frequency extracted by applying the first rule exceeds a predetermined threshold or 5. The information extraction apparatus according to claim 1, 2, 3 or 4 limited to the vocabulary category. Computer
Individual field vocabulary identification means for identifying vocabulary used in individual fields from documents in individual fields,
A rule applying means for applying one or more first rules to the vocabulary used in the individual field specified in the individual field document to extract a relationship between the vocabularies;
For each of the first rules, the extraction of the first rule is applied as a trigger condition from a vocabulary forming a predetermined term or a set of categories of the vocabulary included in each of a plurality of relations extracted by the rule applying means. A rule creating means for creating a second rule for extracting a new relationship between the vocabulary
An information extraction program for functioning as an information extraction unit that extracts a vocabulary relationship from a document in the individual field by using the first rule and the second rule together. Individual field vocabulary identifying means for identifying vocabulary used in individual fields from documents in individual fields;
Rule applying means for applying one or more first rules to the vocabulary used in the individual field specified in the individual field document, and extracting a relationship between the vocabularies;
For each of the first rules, the extraction of the first rule is applied as a trigger condition from a vocabulary forming a predetermined term or a set of categories of the vocabulary included in each of a plurality of relations extracted by the rule applying means. And an information extraction rule generation device having rule creation means for creating a second rule for extracting a new relationship between the vocabularies. Computer
Individual field vocabulary identification means for identifying vocabulary used in individual fields from documents in individual fields,
A rule applying means for applying one or more first rules to the vocabulary used in the individual field specified in the individual field document to extract a relationship between the vocabularies;
For each of the first rules, the extraction of the first rule is applied as a trigger condition from a vocabulary forming a predetermined term or a set of categories of the vocabulary included in each of a plurality of relations extracted by the rule applying means. An information extraction rule generation program for functioning as rule creation means for creating a second rule for extracting a new relationship between vocabularies.

Images