JP2001229177A - Method and device for structuring instance base, and recording medium with recorded instance base structuring program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem that a word having the same meaning property is not always selected and then the reference of classification of documents is inconsistent since a key word as the reference of classification is selected by using as a clue the distribution of words between documents without considering the meaning property of words. SOLUTION: A term extracting means 17 which extracts a term from a text 11 and a classification pattern generating means 18 which generates the classification pattern of the text 11 from the term are included and on the basis of the classification pattern, the text 11 is clustered.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is applied to a system for accumulating past inquiries in a help desk business or the like, and when a new inquiry is received, an operator searches for past cases accumulated and uses them for problem solving. The present invention relates to a case-based construction method, a case-based construction apparatus, and a recording medium on which a case-based construction program is recorded.

[0002]

2. Description of the Related Art FIG.
FIG. 1 is a configuration diagram showing a conventional case-based construction apparatus shown in Japanese Patent Application Publication No. H10-207, in which 1 is a document classification apparatus, and 2 is a document D.
When the data management unit 3 manages B8 and the keyword DB 9 and manages the input and output of documents and keyword groups, and receives the document from the data management unit 2, the data management unit 3 refers to the word dictionary 6 or thesaurus dictionary 7 and searches for the word keyword. A word detector to generate,
4 receives a keyword group of each document from the data management unit 2 and generates a hierarchical classification system. 5 receives the hierarchical classification system from the document classification unit 4 and displays an interface screen on the terminal 10. This is a classification result output unit to be displayed.

[0003] Reference numeral 6 denotes a word dictionary containing general terms, 7 denotes a thesaurus dictionary containing upper / lower relations between terms and synonym information, and 8 denotes a document D storing documents to be classified.
B and 9 are keyword DBs for storing keywords manually assigned in advance, and 10 is a terminal equipped with a CRT, a keyboard, and the like. FIG. 35 is a flowchart showing the processing of the document classification unit 4.

Next, the operation will be described. First, when a data is input from the document DB 8, the data management unit 2 outputs the document to the word detection unit 3.

When receiving a document from the data management unit 2, the word detection unit 3 executes a morphological analysis process with reference to the word dictionary 6, and extracts one or more words from the document. Then, when one or more words are extracted from the document, the word detection unit 3 refers to the thesaurus dictionary 7 and generates a synonym group from the word group. The word detection unit 3 generates a word keyword by combining the word group and the synonym group, and outputs the word keyword to the data management unit 2.

Thus, the data management unit 2 additionally registers the word keyword in the keyword group of each document in the keyword DB 9. In addition, the word detection unit 3 determines the appearance frequency of each word in the total appearance frequency of all words, the title,
The keyword importance of each word is calculated based on the appearance position of the heading, paragraph, etc. in the document.

[0007] When the document classifying unit 4 receives the keyword group of each document from the data management unit 2, it generates a hierarchical classification system. That is, when the document classifying unit 4 receives, from the data management unit 2, a document associated with the manual keyword or the word keyword generated by the word detection unit 3, for each keyword associated with the document, A process of combining documents including keywords (hereinafter, referred to as "single keyword classification process") is executed (step ST1). A set of documents compiled by the single keyword classification process is called a single keyword folder.

Next, the document classifying unit 4 executes a related keyword classifying process, compares documents in the single keyword folder, combines the single keyword folders including similar document groups, and associates A keyword folder is created (step ST2). Whether or not the combination is possible is determined by executing a distance calculation from the word vectors of the documents in the single keyword folder and comparing the distance distribution in the single keyword folder.

Next, the document classifying unit 4 determines whether a set of related keyword folders including a group of similar documents can be integrated with respect to the related keyword folder created by the related keyword classification process, and the integration is possible. During the interval, a process of repeating integration of the related keyword folder is executed (step ST3).

The document classifying unit 4 further checks whether or not the single keyword folder or the related keyword folder can be subdivided (step ST4), and executes a process of repeating the classification hierarchically while the subclassification is possible. (Step ST5). In step ST5, hierarchical classification is performed for all folders, and when there are no unprocessed folders, the process ends (step ST6).

When receiving the hierarchical classification system from the document classification unit 4, the classification result output unit 5 displays an interface screen on the terminal 10.

[0012]

The conventional case-based construction apparatus is configured as described above.
There were two issues.

First, since a keyword serving as a reference for classification is selected based on the distribution of words between documents without considering the semantic properties of the words, words having the same semantic properties are selected. However, as a result, there has been a problem that the standards for document classification become inconsistent. For example, when classifying failure response records, documents containing the same "company name" are collected in a certain folder, and the same "failure name" is stored in a certain folder.
It can happen that documents with are gathered. As a result, as shown in FIG. 36, clusters having different concepts of words serving as classification criteria are mixed, and it becomes difficult to efficiently search for a target document. In the example of FIG.
Cluster including "Company A" and "F failure" or "G failure"
Clusters containing Ideally, as shown in FIG. 37, it is desirable that classification be possible based on words having a common concept according to the purpose of the user.

Second, by omitting the process of automatically extracting a keyword from a document and setting only keywords assigned to a document to manually assigned keywords, it is possible to set a keyword as a reference for classification. However, for that purpose, it is necessary for the user to specify each keyword as a criterion of classification accurately, and there is a problem that the operation burden on the user increases. For example, when "company name" is used as a criterion for classification, it is necessary to specify individual "company names" (for example, "company A", "company B", and "company C") appearing in a document. Similarly, when “fault name” is used as a criterion for classification, each “fault name” (“D fault”, “E fault”
"F failure" and "G failure"). In order to enable the designation, the user needs to know in advance all the company names and failure names that appear in the document.

[0015] The present invention has been made to solve the above-described problems, and it is possible to perform classification based on words having a common concept according to the purpose of a user, and to provide an example of the purpose. It is an object of the present invention to obtain a case-based construction method, a case-based construction device, and a recording medium on which a case-based construction program can be efficiently searched. In addition, the present invention provides a case-based construction method and a case-based construction that can automatically extract a keyword serving as a reference for classification from a document and classify the document without specifying individual keywords serving as a reference for classification. An object is to obtain a recording medium in which an apparatus and a case base construction program are recorded.

[0016]

A case base construction method according to the present invention includes a term extraction step of extracting terms included in text, and a classification pattern generation step of generating a text classification pattern from the terms. The clustering of the text is executed based on the classification pattern.

In the case-based construction method according to the present invention, the synonym of a term is determined from the similarity of the notation of the term extracted in the term extracting step, and the fluctuation of the notation related to the synonym is eliminated. .

The case-based construction method according to the present invention refers to a synonym dictionary, determines the synonymity of the term extracted in the term extraction step, and eliminates the fluctuation of the notation related to the synonym. is there.

In the case-based construction method according to the present invention, clustering is executed with reference to an ontology dictionary in which vocabulary conceptual relationships are described.

In the case-base construction method according to the present invention, an IS-A dictionary in which the relation between a superordinate concept and a subordinate concept of a word is described is used as an ontology dictionary.

In the case-based construction method according to the present invention, a HAS-A describing a relationship between a higher-level configuration and a lower-level configuration of a device is described.
The dictionary is used as an ontology dictionary.

In the case-based construction method according to the present invention, a classification target sentence extracting step of extracting a classification target sentence to be processed in a term extraction step, a classification pattern generation step and a cluster generation step from a text input in the input step. Is provided.

In the case-based construction method according to the present invention, when a plurality of classification target sentences are extracted from one text in the classification target sentence extracting step, clusters are generated in order from the first classification target sentence of the text. This is to create a hierarchical structure.

In the case-based construction method according to the present invention, when a text classification pattern is generated, a classification pattern description table in which a vocabulary pattern serving as a reference for classification is described, while the classification pattern description table includes The vocabulary pattern is described for each level of the same level.

In the case-based construction method according to the present invention, of the clusters generated in the cluster generation step,
A display step is provided for displaying a cluster in descending order of the number of belonging texts.

In the case-based construction method according to the present invention, when a cluster to be displayed is designated, the entire text of the text assigned to the cluster is displayed.

In the case-based construction method according to the present invention, when a cluster to be displayed is designated, a reference target sentence assigned to the cluster is displayed.

In the case-based construction method according to the present invention, when a plurality of classification target sentences are extracted from one text in the classification target sentence extracting step, identification information for identifying the type of each classification target sentence is displayed. It was done.

In the case-based construction method according to the present invention, when a plurality of classification target sentences are extracted from one text in the classification target sentence extracting step, an arbitrary classification target sentence is displayed in an initial stage of the hierarchical structure, Another classification target sentence is displayed from the middle of the structure.

In the case base construction method according to the present invention, the clustering is executed while associating the classification target sentence extracted in the classification target sentence extraction step with the existing cluster stored in the case base. .

The case base construction method according to the present invention is provided with an editing step of editing a cluster stored in the case base.

In the case base construction method according to the present invention, when the cluster generated in the cluster generation step is stored in the case base, the similarity between the cluster and the existing cluster stored in the case base is smaller than a predetermined threshold. When it is high, the display processing of the display step and the edit processing of the edit step are omitted.

In the case-based construction method according to the present invention, when the cluster generated in the cluster generation step is stored in the case base, the similarity between the cluster and the existing cluster stored in the case base is smaller than a predetermined threshold. When the number is low, the display processing of the display step and the edit processing of the edit step are omitted, and the cluster is not stored in the case base but is stored in a temporary storage file.

The case base construction apparatus according to the present invention is provided with term extraction means for extracting terms included in text, and classification pattern generation means for generating a classification pattern of the text from the terms, with the classification pattern as a reference. This is to perform text clustering.

The recording medium storing the case base construction program according to the present invention is provided with a term extraction processing procedure for extracting terms included in text and a classification pattern generation processing procedure for generating a text classification pattern from the terms. The text clustering is executed based on the classification pattern.

[0036]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a block diagram showing a case-based construction apparatus according to Embodiment 1 of the present invention.
Reference numeral 1 denotes a text to be processed in case construction, reference numeral 12 denotes a basic word dictionary used for morphological analysis of the text 11, and the basic word dictionary 12 includes two fields of notation and part of speech (see FIG. 2). Reference numeral 13 denotes a term pattern description table in which a term and a pattern used to extract a concept of the term from the text 11 (see FIG. 3) are described. Reference numeral 14 denotes a vocabulary pattern (see FIG. ) Is described.

15 is an input means for inputting the text 11;
Reference numeral 16 denotes a sentence analyzing unit for executing morphological analysis of the text 11 with reference to the basic word dictionary 12, and reference numeral 17 refers to a term pattern description table 13 for extracting a term included in the text 11 and a concept of the term. Means 18 refers to the classification pattern description table 14 and refers to the term extraction means 17
Is a classification pattern generating means for generating a classification pattern of the text 11 from the terms and concepts extracted by the above.

Reference numeral 19 denotes a cluster generating means for performing clustering of the text 11 on the basis of the classification pattern generated by the classification pattern generating means 18, and 20 denotes an output means for storing the cluster generated by the cluster generating means 19 in the case base 23. , 21 are editing means for editing the cluster stored in the case base 23, and 22 is the input means 1
5. Control the operations of the sentence analysis unit 16, the term extraction unit 17, the classification pattern generation unit 18, the cluster generation unit 19, the output unit 20, and the editing unit 21 and execute the transfer of data that is input / output of each unit. Control means, 23
Is a case base for storing clusters. FIG. 5 is a flowchart showing the case base construction method according to the first embodiment of the present invention.

Next, the operation will be described. First, the input unit 15 reads the text 11 (step ST11). However, in the first embodiment, it is assumed that the text 11 indicating the failure handling record regarding the network is read. FIG. 6 shows a text 11 showing a failure response record.
This is an example.

When the input means 15 reads the text 11, the sentence analyzing means 16 extracts a classification target sentence from the text 11 (step ST12). In this example, among the sentences recorded in the text 11, a description related to the symptom of the failure is set as a classification target sentence.

In the process of extracting a classification target sentence, a sentence tag determination table is referred to (see FIG. 7). The sentence tag determination table has a sentence tag definition section and a keyword definition section,
A sentence including a character string defined in the keyword definition part is represented by a description of the content defined in the sentence tag definition part. For example, a sentence containing the character string "detected" becomes "symptom"
It is a description about. Therefore, in each sentence of the failure response record in FIG. 6, “detection”, “occurrence”, “ON”, “O”
A sentence tag of “symptom” is added to a sentence including a character string of either “FF” or “restoration”, and the sentence is determined as a classification target sentence.

Next, the sentence analyzing means 16 extracts a reference target sentence from the text 11 (step ST13). In this example, among the sentences recorded in the text 11, a description relating to an action coping with the failure is set as a reference target sentence. The sentence tag determination table is referred to in the same manner as the processing in step ST12. In each of the sentences of the failure response record in FIG. 6, a sentence tag including "Act" is added to a sentence including any of the character strings of "execute", "contact", and "request", and is determined as a reference target sentence.

FIG. 8 is an explanatory diagram showing a failure response record to which a sentence tag has been added by the processing of steps ST12 and ST13. However, in FIG. 8, the sentence tags are numbered so as to be easily identified in the following description, and the order in which the "symptom" or "action" appears in the same failure response record is shown.

Next, the sentence analyzing means 16 executes a morphological analysis of the sentence to be classified (step ST14). Morphological analysis is, for example, "morphological analysis of Japanese sentence including unregistered words"
(Kenji Yoshimura, Mitsuno Takeuchi, Kenzo Tsuda, Kimiaki Shuto, Information Processing Society of Japan, Vol. 30, No. 3, 1989,
Hereinafter, it is executed based on the minimum cost method disclosed in “Document A”. Document A can also analyze an input sentence that includes an unregistered word that is not in the morphological analysis dictionary.
At that time, unregistered words in the katakana string or alphabet string are treated the same as independent words. FIG. 9 shows word breaks obtained as a result of performing morphological analysis on the classification target sentence of FIG.

When the sentence analysis means 16 executes the morphological analysis of the sentence to be classified, the term extraction means 17 refers to the term pattern description table 13 to extract the terms contained in the text 11 (step ST15). Here, the description form of the term pattern description table 13 will be described. FIG.
Changes the format of the term pattern description table 13 to BNF (Ba
ckus Naur form) notation.

The term pattern is composed of a pattern definition section and a term determination section, and both are connected by "→". The pattern definition part contains a morphological description that represents the characteristics of the morphological analysis result.
Consists of one or more sequences. The morpheme description is an AND, OR of a notation pattern, a part-of-speech pattern, a character type pattern, or a character string length pattern (collectively referred to as a “morpheme pattern”).
And repeated.

In FIG. 10, string represents a character string, and number represents a number. A character string is taken as an argument of a notation pattern, a part of speech pattern, a character type pattern, and a vocabulary concept pattern, and a number is taken as an argument of a character string length pattern. For example, s (occurrence) is a morpheme whose notation is “occurrence”, p (noun) is a morpheme whose noun is part of speech, t (kata) is a morpheme of katakana character type, and l (1) is a morpheme whose character string is one character. The morpheme c (device) is a morpheme whose vocabulary concept is “device”. Each morpheme description is separated by “,”. FIG. 11 shows a list of character types used in the first embodiment, and FIG. 12 shows an example of a morpheme description.

On the other hand, the term deciding section is represented by a sequence of a term range deciding section, a term part of speech deciding section, and a term concept deciding section. The term range deciding unit designates the range of the morpheme description of the pattern definition unit by a numeral. For example, range (1,
2) are the first and second morpheme patterns in the pattern definition section.
This means that the range where the th is connected is acquired as one term. The term part-of-speech determining unit specifies the part of speech of the acquired term. For example, pos (noun) indicates that the part of speech of the extracted term is a noun. The term concept determination unit specifies the concept of the acquired term. For example, concept (device) indicates that the concept of the extracted term is a device. Further, a character string determined by the term range determination unit can be specified as an argument of the term concept determination unit. For example, concept (ran
ge (1,2) + device) is a character string that connects the first and second morpheme patterns in the pattern definition unit, and “device”
Is extracted as a concept.

FIG. 13 is a flowchart showing the morphological analysis result and the collation processing of the pattern definition part in the term extraction processing of step ST15. Here, as an example,
Operation for collating “frame / monitoring device / at Major / alarm / occurrence /.” Which is the result of morphological analysis of symptom 1 in failure response record 1 of FIG. 9 with the term pattern of (1) of FIG. explain.

First, in step ST31, a process of extracting the first morpheme description "(t (kata) | | t (al) | | t (kanji)) +" is executed. Next, in step ST32, it is determined whether or not there is a repetition symbol in the morpheme description. In this example, since there is a repetition symbol in the morpheme description, the process proceeds to step ST33 to extract a “frame” which is a morpheme of the analysis result. Execute the process.

Next, in step ST34, FIG.
To determine whether the evaluation value is “TRUE”. That is, the process B is called in step ST61 of FIG.
Is called, and in step ST69, it is determined whether or not the first character is “(”. “(T (kata) || t”
Since the first character of ((Al) || t (Kanji)) + ”is“ (”, the process proceeds to step ST70 to execute a process for calling process A.

Next, in step ST61, process B is called, and in step ST65, process C is called, and in step ST69, it is determined whether or not the first character is "(". So that
The determination result is “No”, and in step ST71, it is evaluated whether or not the morpheme pattern “t (kata)” matches the morpheme “frame”.

FIG. 16 is a flowchart showing the evaluation processing in step ST71. Since the morpheme pattern “t (kata)” is a character type pattern, step ST81
Is "No" and the determination in step ST85 is "N".
o ", the determination in step ST89 is" Yes ". In step ST90, since the “frame” is a katakana character string, the evaluation value is “T” in step ST91.
RUE ", the flow in FIG. 16 ends, and the process returns to step ST71.

Next, step ST6 in which the process C is called
Returning to step ST66, in step ST66, it is checked whether or not the next character string is "&&". In this example,
Since the next character string is “||”, the determination is “No”, and the process returns to step ST61 in which the process B is called.

Next, in step ST62, it is checked whether or not the next character string is "||".
es ”, it is evaluated in step ST63 whether the next morpheme description“ t (al) ”matches the morpheme“ frame ”. The evaluation value becomes “FALSE” by the same processing as the above-described process of matching the morpheme description “t (kata)” with the morpheme “frame”.
In 4, by ORing the expressions ("TRUE"
OR "FALSE"), and the evaluation value is "TRU
E ".

Then, the process proceeds to a step ST62 to check whether or not the next character string is "||". If the judgment is "Y
es ", and evaluates whether the next morpheme description" t (kanji) "matches the morpheme" frame ". Although the evaluation value becomes “FALSE” by the above-described processing, in step ST64, the OR of the expressions is obtained (“T
RUE ”and“ FALSE ”), and the evaluation value becomes“ TRUE ”. By the above processing, the routine of FIG. 15 ends with the evaluation value “TRUE”, and returns to step ST34 of FIG.

Since the evaluation value is "TRUE", the determination in step ST34 is "Yes", and step ST3
Go to 5. In step ST35, the morpheme description after the repetition part is collated. FIG. 14 is a flowchart showing the collation processing in step ST35.

In step ST51, a process of extracting the morpheme description of the term pattern "(s (at) || s (at))" is executed. Next, in step ST52,
A process for extracting the morpheme “monitoring device” of the analysis result is executed. In step ST53, the morpheme description of the term pattern and the morpheme of the analysis result are collated, and the determination in the process of FIG. 15 becomes “FALSE”. As a result, in step ST54, it is determined that the collation has failed, and the process returns to step ST35.

The process of extracting the morpheme description in step ST51 of FIG. 14 and the process of extracting the morphological analysis result in step ST52 do not affect the process of FIG.
That is, the morpheme description or morphological analysis result to be taken out next after execution of step ST35
This is the same as the morpheme description or morphological analysis result to be taken out next before the execution of 35.

Next, in step ST36, it is determined whether or not the analysis result has been completed. Since there is an analysis result subsequent to the “frame”, the determination is “No”, the process proceeds to step ST33, and a process of extracting the morpheme “monitoring device” of the next analysis result is executed. Next, step ST3
In step 4, the morpheme description “(t (kata) || t (al) || t (kanji)) +” is compared with the analysis result “monitoring device” according to the processing procedure of FIG. By executing the above-described collation processing, the evaluation value becomes “TRUE”.

Next, in step ST35, the morpheme description after the repetition part is collated. Step S in FIG.
In T51, the morpheme description “(s (at) || s (at))” of the term pattern is extracted, and at step ST52, a process of extracting “at” is executed. Next, in step ST53, the two are collated, and the determination is “TRU
E ". As a result, in step ST55, the collation is determined to be successful, and the process returns to step ST35.

Next, in step ST36, it is determined whether or not the analysis result has been completed. However, since there is an analysis result subsequent to the "monitoring device", the determination is "No", and the process proceeds to step ST33. The morpheme “at” of the analysis result
Execute the process of extracting the. Next, in step ST34, the morpheme description “(t (kata) || t (al) || t (kanji)) +” is compared with the analysis result “at” according to the processing procedure of FIG. By performing the above-described collation processing, the evaluation value becomes “FALSE”.

Next, in step ST43, it is checked whether or not there is a character string that has been successfully collated.
(Kata) || t (Al) || t (Kanji)) +], the matching between the "frame" and the "monitoring device" was successful, so the determination is "Yes". Next, in step ST44, it is checked whether or not the pattern definition section has been completed, and "(t (kata) | | t (al) | | t (kanji))
+ ”Followed by the morpheme description“ (s (at) || s
()) ”, The determination is“ No ”.

Next, the process proceeds to step ST31, in which a process for extracting the next morpheme description "(s (at) || s (at))" is executed. In step ST32, since there is no repetition symbol in the morpheme description, the determination is “No”, and the process proceeds to step ST37. Next, in step ST37, a process of extracting the next morpheme “at” of the “frame” and the “monitoring device” that have been successfully collated is executed.

Next, in step ST38, FIG.
Is performed and the result becomes "TRUE", so that the procedure goes to step ST39. Next, step S
At T39, it is checked whether or not the pattern definition part has been completed. The morpheme description “(t (kata) || t (al) |” following “(s (at) || s (at))” is checked. | T
(Kanji)) + ”, the determination is“ No ”, and the process proceeds to step ST40. Next, in step ST40, it is checked whether or not the analysis result has been completed. Since there is a morpheme “Major” following “at”, the determination is “No”, and the process proceeds to step ST31.

When the same processing is continued, the correspondence shown in FIG. 17 is established between the morphological analysis result and the pattern definition part of the term pattern. Next, terms are extracted according to the description of the term determination unit. The description of the term range determination unit "rang
e (1,1) ”as the notation of the term“ (t (kata) |
| T (Al) || t (Kanji)) + ”is extracted, and the description“ po
“Noun” is extracted as a part of speech from “s (noun)”, and “monitoring device” is extracted as a concept from the description “concept (monitoring device)” of the term concept determination unit.

"Router / disconnection / detection /." Which is the result of the morphological analysis of symptom 2 in the failure response record 1 of FIG.
18 is established between the morphological analysis result and the pattern definition unit for the term pattern of (3) in FIG. At this time, the extracted term “router disconnection” is extracted from the description “range (1, 2)” of the term range determination unit, and the description “pos
“Noun” is extracted as a part of speech from “(Noun)”. In addition, the description “concept (rang)
e (1,1) + failure), the concept of the term is “r
A “router fault” that connects the “router” corresponding to “ange (1, 1)” and the “fault” is extracted. FIG. 19 is an explanatory diagram showing terms extracted from the failure response record of FIG. 9 in step ST15.

When the term extraction processing by the term extraction means 17 is completed, the classification pattern generation means 18 eliminates the swaying of all words which are the results of analysis by the sentence analysis means 16 (step ST16). Here, the sway of the notation due to the presence / absence of the long sound "-" at the end of the katakana character string and the sway of the notation due to the difference between uppercase / lowercase in the alphabet character string are eliminated. Therefore, long sounds at the end of the katakana character string are deleted, and all lowercase letters of the alphabet are converted to uppercase.

For a notation that cannot be resolved by regular conversion, such as the presence / absence of a long sound and the uppercase / lowercase letters of the alphabet, the synonym dictionary shown in FIG. 20 is used. The synonym dictionary includes a heading field in which a character string appearing in a text is described, and a regular notation field for storing a regular notation of the heading. For example, "frame"
And "frame relay" can be recognized as words having the same meaning by being converted to "FR". FIG.
FIG. 1 is an explanatory diagram showing a term extraction result in which a spelling elimination process has been executed.

When the classification pattern generating means 18 completes the processing for eliminating the fluctuation of the notation, it generates a classification pattern by referring to the classification pattern description table 14 (step ST17). The classification pattern description table 14 is represented by a combination of AND or OR of notations or concepts (hereinafter, referred to as “classification elements”) of words used for classification. For the classification element, predicates “string”, “concept”, and “subconcept” can be specified. A character string for which no predicate is specified represents the notation of the character string. FIG. 22 shows “string”, “concept
It is an explanatory view showing an example of how to use “t” and “subconcept”.

"String" takes a concept as an argument and represents each notation for the concept shown in the argument. For example, if the terms shown in FIG.
The value of (alarm name) is expressed as “MAJOR alarm” and “MINOR alarm”.

“Concept” takes a concept as an argument,
Represents the concept itself shown in the argument. For example, "conc
The value of “ept (alarm name)” is a concept of “alarm name”. In the classification element, the notation of a character string and the concept are mixed, so to avoid confusion, the concept is “<
>>. For example, "Alarm name"
Notation is expressed as "alarm name" and "alarm name"
Is expressed as “<alarm name>”.

"Subconcept" takes a concept as an argument and represents a lower concept of the concept shown in the argument. Here, the relationship between the superordinate concept and the subordinate concept is determined based on whether or not the end of the concept is included as a partial character string. The superordinate concept is a partial character string whose end matches the subordinate concept. For example, "<failure>"
Is a partial character string that matches the end of “<router fault>”, so the concept “<failure>” is a superordinate concept of the concept “<router fault>”. Conversely, the concept “<router fault>” is a subordinate concept to the concept “<failure>”. Therefore, when the terms shown in FIG. 21 are extracted, the value of the description “subconcept (fault)” is “<router fault>” and “<modem fault>”
It is.

According to the classification pattern (1-1) shown in FIG. 4, each of a plurality of sentences has words having the concept of "<monitoring device>" and "<alarm name>", and the notation of those words. Are the same, they are classified into the same cluster. Further, according to the classification pattern (2-1), sentences including words having the same concept as a lower concept of “disorder” are classified into the same cluster. Furthermore, according to the classification pattern (2-2), the same “survey”
Sentences that contain words with the subordinate concept of, and that contain the notation “abnormal” and that contain the notation “there are” are classified into the same cluster, and contain words that have the same subordinate concept of “investigation”. , And contains the character string "abnormal", and
Sentences containing the character string “None” indicate that they are classified into the same cluster. As shown in FIG. 4, the classification pattern can be described for each layer. FIG. 23 is an explanatory diagram showing a classification pattern generated for the morphological analysis result of FIG.

When the classification pattern generation unit 18 generates a classification pattern, the cluster generation unit 19 executes clustering based on the classification pattern to generate a cluster (step ST18). FIG. 24 is a flowchart showing the cluster generation processing. The cluster is generated hierarchically from the head classification target sentence. First, in step ST101, the current hierarchy i is changed to the maximum hierarchy layer
er is determined. However, the initial value of i is “1”, and in this example, the maximum hierarchy layer is “4”.
And Therefore, the determination here is “Yes”, and the process proceeds to step ST102.

In step ST102, a process of extracting the classification pattern of the first hierarchy is executed. Here, the classification patterns corresponding to “symptom 1” of the failure response record 1, the failure response record 2, the failure response record 3, the failure response record 4, and the failure response record 5 are extracted. The extracted classification patterns are as follows.・ For the fault response record 1, (FR monitoring device, MAJO
(R alarm) ・ For failure response record 2, (FR monitor, MAJO
(R alarm) ・ For failure response record 3, (FR monitor, MAJO
(R alarm) ・ For failure response record 4, (FR monitoring device, MAJO
(R alarm) ・ For the fault response record 5, (modem monitoring device, MIN
OR alarm)

Next, in step ST103, a cluster is generated from the classification target sentences having the same classification pattern. Since the classification patterns of the failure response record 1, the failure response record 2, the failure response record 3, and the failure response record 4 are common, the following statement forms one cluster (hereinafter, referred to as "cluster C1").

The sentence of the symptom 1 of the failure response record 1 → “A major alarm has occurred in the frame monitoring device.” The sentence of the symptom 1 of the failure response record 2 → “M
AJOR alarm detection.・ Symptom 1 sentence of failure response record 3 → “Ma with FR monitoring device”
Jor alarm detection. -The sentence of symptom 1 in trouble shooting record 4 → "Ma with FR monitoring device"
Jor alarm detection. "

Further, the sentence of the symptom 1 of the failure handling record 5 “Minor alarm has occurred in the modem monitoring device” becomes one cluster (hereinafter, referred to as “cluster C2”).

Next, in step ST104, a process of allocating the reference target sentence to the cluster is executed. In this example, the description of “action” is a reference target sentence. The reference target sentence to be associated with a cluster is a description of an "action" that appears from the description of the "symptom" belonging to the cluster to be associated with the description of the next "symptom". The description of “action” that satisfies this condition is as follows.

-Failure response record 1-> Action 1 "Contact customer."-> Action 2-Request for on-site investigation.-Failure response record 2-> Action 1-Report the situation to the customer. Request. ”• Failure response record 3 → Action 1“ Contact customer. ”→ Action 2“ Contact the site. ”• Failure response record 4 → None. • Failure response record 5 → Action 1“ Contact customer Absence. "→ Action 2" Contact the site. "

Therefore, the sentence to be referred to which is associated with the cluster C1 is as follows. -Action 1 of the failure response record 1 "Contact the customer."-Action 2 of the failure response record 1 "Request a field survey."-Action 1 of the failure response record 2 "Notify the customer to the situation." Action 2 of 2 “Request for field survey.” ・ Action 1 of failure response record 3 “Contact customer.” ・ Action 2 of failure response record 3 “Contact the site.”

The reference target sentence associated with the cluster C2 is as follows.・ Act 1 of the failure response record 5 “Contact the customer, but there is no person in charge.” ・ Act 2 of the failure response record 5 “Contact the site.”

Next, in step ST105, a process of associating the entire text with the cluster is executed. The full text of the failure response record 1, the failure response record 2, the failure response record 3, and the failure response record 4 is associated with the cluster C1, and the full text of the failure response record 5 is associated with the cluster C2. . Note that the number k of clusters generated in step ST103 is “2”, and a set of clusters is C1 and C2.

Next, in each of the clusters C1 and C2, a cluster of the next hierarchy is generated. In step ST106, j indicating the order of the cluster to be processed is designated as “1”. In step ST108, until it is determined in step ST107 that child clusters have been generated for all clusters, (i +
1) Generate a cluster of the tier. Here, C1
A process of generating a child cluster of will be described.

First, since the value of i indicating the current hierarchy is “2”, the determination in step ST101 is “Yes”.
And the process proceeds to step ST102. Next, step S
At T102, a process of extracting the classification pattern of the second hierarchy from the text belonging to the cluster C1 is executed. Here, a classification pattern corresponding to the sentence of the symptom 2 in the failure response record 1, the failure response record 2, the failure response record 3, and the failure response record 4 is extracted. The extracted classification patterns are as follows.

For the failure response record 1, (<router failure>) For the failure response record 2, (<router failure>) For the failure response record 3, (<modem failure>) Failure For correspondence record 4, (<High-speed modem failure>)

Next, in step ST103, clusters are generated from the classification target sentences having the same classification pattern. Here, an example will be described in which an ontology dictionary that describes conceptual relationships of vocabulary is used when generating a cluster. The ontology dictionary includes an IS-A dictionary that describes the relationship between a higher concept and a lower concept of a vocabulary, and a HAS-A dictionary that describes the relationship between a higher configuration and a lower configuration of devices and the like. FIG.
5 shows an example of the IS-A dictionary, and FIG. 26 shows an example of the HAS-A dictionary. FIGS. 25A and 26A show the superordinate / subordinate relation in a tree structure.
26, and the upper / lower relation of each word is described as a pair of two words.

Here, since "high-speed modem" is a subordinate concept of "modem", "<high-speed modem failure>" can be determined as a kind of "<modem failure>". Therefore, in the process of step ST103, the following statement forms one cluster (hereinafter, referred to as “cluster C11”). -The sentence of the symptom 2 of the failure response record 1 → "Router disconnection detected."-The sentence of the symptom 2 of the failure response record 2 → "Router failure detected."

The following statement forms one cluster (hereinafter, referred to as "cluster C12"). -The sentence of the symptom 2 in the failure response record 3 → "Modem failure is also detected."-The sentence of the symptom 2 in the failure response record 4 → "High-speed modem disconnection is also detected."

Next, in step ST104, a process of associating a reference target sentence with a cluster is executed. When the above-described processing is similarly executed, the reference target sentence associated with the cluster C11 becomes “none”, and the reference target sentence associated with the cluster C12 is as follows. -Action 3 of failure response record 3 "Perform modem replacement."-Action 1 of failure response record 4 "Perform modem replacement."

Next, in step ST105, processing for associating the entire text with the cluster is executed. The full text of the failure response record 1 and the failure response record 2 is associated with the cluster C11, and the full text of the failure response record 3 and the failure response record 4 is associated with the cluster C12.

Next, a cluster of the next hierarchy in the clusters C11 and C12 is generated. Step ST1
At 08, the (i + 1) th, that is, the third cluster is generated for the cluster C11. Step ST
Since the judgment of 101 is "Yes",
Proceeding to T102, a process of extracting a classification pattern corresponding to the sentence of the symptom 3 of the failure response record 1 and the failure response record 2 belonging to the cluster C11 is executed. The extracted classification patterns are as follows. -For the failure response record 1, (<power switch>, O
N)-For the failure response record 2, (<router switch>, O
N)

Here, referring to the HAS-A dictionary of FIG. 26, since the “router” has “power” as a lower configuration, it can be determined that the “router switch” and the “power switch” are the same. Therefore, in the process of step ST103, the following sentences form one cluster. -The sentence of the symptom 3 in the failure response record 1 → "Check that the power switch is turned on."-The sentence of the symptom 3 in the failure response record 2 → "The router switch is O.
N. "

Hereinafter, by performing the same processing,
FIG. 27 shows clusters generated from the failure response record 1, the failure response record 2, the failure response record 3, the failure response record 4, and the failure response record 5.

FIG. 24 shows a process for generating a cluster when an existing cluster does not exist.
If an existing cluster exists in the case base 23, a process of associating a text to be newly added with the existing cluster is required in step ST103. When the cluster in FIG. 27 already exists, the failure handling record 6 in FIG.
Is added.

FIG. 29 shows the text of FIG.
Extraction result of the classification target sentence in step ST12, extraction result of the reference target sentence in step ST13, step ST13
FIG. 14 is an explanatory diagram showing a morphological analysis result by step 14, a term extraction result by step ST15, and a classification pattern generation result by step ST17. A process of generating a cluster according to the classification pattern of FIG.

In step ST101 in FIG. 24, the initial value of i indicating the current hierarchy is “1”, and the maximum hierarchy l
ayer is "4", the step ST101 is executed.
Is "Yes", and then in step ST102, the classification pattern (FR
Take out the monitoring device, MAJOR alarm).

Next, in step ST103, clusters are generated based on the common classification pattern.
Since the classification pattern is the same as that of the cluster C1, the image is associated with the cluster C1. Next, in step ST104, the reference target statement is associated with the cluster,
In step ST105, the entire text is associated with the cluster, but the description is omitted because it is the same as the above-described processing.

Next, the process for the child hierarchy is performed. In this case, the comparison target is the classification patterns included in the clusters C11 and C12 which are the child clusters of the cluster C1. In step ST102, a classification pattern (<modem failure>) for the second classification target sentence is extracted, and in step ST103, a cluster is generated.
It is associated with the cluster C12.

For the classification pattern (<modem switch>, OFF) for the third classification target sentence of the failure response record 6, the cluster C, which is a child cluster of the cluster C12,
Although the classification process is performed with the sentence included in 121 as a comparison target, a new cluster is generated because the classification pattern is different from that of the cluster C121. In addition, since there is no cluster to be compared with the classification pattern (recovery) in the fourth classification target sentence of the failure response record 6, a new cluster is generated. When there is an existing cluster as shown in FIG. 27, the cluster when the failure handling record 6 of FIG. 28 is added is shown in FIG.

In the above description, the condition to be assigned to the same cluster is only when the classification patterns completely match. However, the concept of “similarity” is provided, and when the similarity of the classification pattern is high, the same cluster is assigned to the same cluster. Assignment processing may be performed. For example, the similarity is calculated as (number of matching classification elements)
/ (The number of classification elements described in the classification pattern), there is a method in which a sentence having a similarity equal to or greater than a certain value is regarded as matching the corresponding classification pattern.

As described above, step ST1 in FIG.
When the cluster generation processing in 8 is completed, the cluster generation means 19 sorts the clusters according to the number of cases included in the cluster (step ST19). The output unit 20 displays the clustering results sorted by the cluster generation unit 19 (Step ST20).
FIG. 31 is a display example showing a clustering result of a tree structure. In addition to the classification target sentence (description of symptoms), the cluster is associated with the reference target sentence (description of behavior) and the entire text.

The user can search for a target case by checking the contents of clusters in order from the top of the hierarchical structure of the displayed screen. Further, “action” for “symptom” represented by the cluster can be referred to.
For example, C2 cluster "modem monitoring device, MINOR
Guidance on the action to be taken by the operator by referring to the statement of "action" associated with C2 for the "symptom" of "alarm", "contacting customer, but not in charge" Can be. Further, by making it possible to refer to past cases in which the same “symptom” has occurred, it is possible to assist the operator in solving a problem when a failure occurs.

The user can edit the clustering result using the editing means 21 (step ST2).
1). FIG. 32 is an example of an editing screen, which is a screen for allocating one newly generated text to the existing case base 23. In the figure, 101 is an OK button, 102 is N
G button 103, clustering result display unit 104
Denotes a classification target sentence display unit, 105 denotes a sentence tag display unit, and 106 denotes a text display unit.

The clustering result display unit 103 shows how a new text has been assigned to the hierarchical structure in the existing case base 23, and the hatched clusters in FIG. A cluster,
Displayed with sibling clusters of the existing clusters associated with them. In the extraction processing of the classification target sentence in step ST12 and the extraction processing of the reference target sentence in step ST13, when the determination of the program is incorrect, the sentence tag display unit 105 is edited, and the "symptom" or "action" is correct. It can be modified by adding a sentence tag. The classification target sentence to which the tag of “symptom” is added is displayed on the classification target sentence display unit 104.

When the program is associated with an erroneous existing cluster in the cluster generation processing in step ST18, the sentence displayed on the classification target sentence display unit 104 is associated with the program as shown in FIG. Drag and drop to the cluster you want. At this time, for the child cluster of the associated cluster, the calculation is performed again according to the flow of FIG. 24, and the association is performed.

When the OK button 101 is clicked, the display of the clustering result display section 103 is regarded as correct,
In step ST22 of FIG. 5, the clustering result is registered in the case base 23. When the NG button 102 is clicked, the display on the clustering result display unit 103 is incorrect and is not registered in the case base 23 but is stored in a temporary storage file (not shown). The temporary storage file is a file for storing the text 11 whose registration in the case base 23 is suspended, and is used later when collectively performing the registration processing.

The above is the processing in the case of sequentially adding cases each time a text such as a failure response record is generated. At this time, the text classification pattern to be added is
When assigning to an existing cluster, the similarity is defined by the following expression. Similarity = (the number of matching classification elements) / (the number of classification elements described in the classification pattern) If the similarity is higher than a preset threshold T1 in all the hierarchies, the cluster allocation processing It may be determined that the reliability is high, and the display of the clustering result and the editing of the clustering result may be omitted and automatically registered in the case base 23, and the display and editing may be performed only in other cases.

When assigning a text classification pattern to be added to an existing cluster, if the similarity is lower than a preset threshold T2, it is determined that the reliability of the cluster assignment process is low, and The display of the clustering result in ST20 and the editing of the clustering result in step ST21 may be omitted, automatically stored in a temporary storage file, and later collectively registered in the case base 23.

In the first embodiment, a hierarchical structure is created with the description of "symptoms" as a classification target sentence and the description of "action" as a reference target sentence. May be used as the classification target sentence, and when displaying the clustering result, the hierarchical structure may be displayed so that both can be distinguished, for example, by changing the color of the cluster. Instead of displaying all "symptoms" and "behaviors", only "symptoms" may be displayed at the beginning of the hierarchical structure, and "behaviors" may be displayed after a certain hierarchical level.

The case-based construction program for realizing the operation of the case-based construction apparatus described above may be recorded on a computer-readable recording medium such as a floppy disk or a CD-ROM and operated. can get.

As is clear from the above, the first embodiment
According to this, a term extraction unit 17 for extracting a term included in the text 11 and a classification pattern generation unit 18 for generating a classification pattern of the text 11 from the term are provided, and the clustering of the text 11 is performed based on the classification pattern. Since it is configured to execute, it is possible to perform classification based on words having a common concept according to the user's purpose, and as a result, the effect of efficiently searching for the target case can be obtained. Play. In addition, there is no need to specify individual keywords that serve as classification criteria, and even if a user does not know that a keyword exists in a document, the target document can be classified as a keyword that serves as a classification standard. An effect is provided that a possible case-based construction method can be provided.

[0114]

As described above, according to the present invention, a term extraction step for extracting a term included in a text and a classification pattern generation step for generating a text classification pattern from the term are provided. Is configured to perform text clustering on the basis of, so that it is possible to perform classification based on words that have a common concept according to the user's purpose, and as a result, the target case can be efficiently There is an effect that can be searched.
In addition, there is an effect that a keyword serving as a reference for classification can be automatically extracted from a document and a document can be classified without designating individual keywords serving as a reference for classification.

According to the present invention, the synonym of a term is determined from the similarity of the notation of the term extracted in the term extracting step, and the fluctuation of the notation related to the synonym is eliminated, so that the accuracy of the clustering result is improved. There is an effect that can be increased.

According to the present invention, the synonym dictionary is referred to to determine the synonymity of the term extracted in the term extraction step, and the fluctuation of the notation related to the synonym is eliminated. The conversion has the effect of eliminating the fluctuation of the term that cannot be corrected.

According to the present invention, clustering is performed with reference to an ontology dictionary in which vocabulary concept relationships are described. Therefore, when clusters are generated, different classification patterns are regarded as the same and appropriate There is an effect that a proper clustering can be executed.

According to the present invention, since the IS-A dictionary in which the relation between the superordinate concept and the subordinate concept of a word is described is used as an ontology dictionary, for example, "modem failure" and "high-speed modem failure" There is an effect that clustering can be executed by recognizing that the failures are of the same kind.

According to the present invention, since the HAS-A dictionary in which the relationship between the higher-level configuration and the lower-level configuration of the device is described is used as the ontology dictionary, for example, the "router switch" and the "power switch" are the same. There is an effect that the clustering can be executed by recognizing the object.

According to the present invention, a classification target sentence extraction step for extracting a classification target sentence to be processed in the term extraction step, the classification pattern generation step, and the cluster generation step from the text input in the input step is provided. Because it is composed, for example, "symptoms" and "behavior"
For example, it is possible to classify only sentences having a certain specific meaning, and as a result, there is an effect that an example can be efficiently searched according to the purpose of the user.

According to the present invention, when a plurality of classification target sentences are extracted from one text in the classification target sentence extraction step, clusters are generated in order from the classification target sentence at the head of the text to create a hierarchical structure. As a result, the user can check the contents of the cluster in order from the top of the hierarchical structure, and as a result, there is an effect that the target case can be efficiently searched.

According to the present invention, when a text classification pattern is generated, a classification pattern description table in which a vocabulary pattern serving as a reference for classification is described, and the classification pattern description table has the same level of hierarchy. Since the vocabulary pattern is described for each word,
It is possible to specify on what basis each layer is to be clustered, and as a result, there is an effect that cases can be efficiently searched according to the purpose of the user.

According to the present invention, among the clusters generated in the cluster generation step, the display step is provided for displaying the clusters with the largest number of texts in order, so that the user has a large number of texts. As a result of being able to confirm from the contents of the cluster, there is an effect that a target case can be efficiently searched.

According to the present invention, when the cluster to be displayed is specified, the entire text of the text assigned to the cluster is displayed. Therefore, in the process of the user sequentially confirming the contents of the cluster, The effect is that the entire text can be confirmed directly from the cluster.

According to the present invention, when a cluster to be displayed is specified, the reference target sentence assigned to the cluster is displayed. For example, when a certain "symptom" occurs, It is possible to present useful information to the user, such as whether the user can take an appropriate "action".

According to the present invention, when a plurality of classification target sentences are extracted from one text in the classification target sentence extracting step, the identification information for identifying the type of each classification target sentence is displayed. For example, there is an effect that a case base in which the causal relationship between “symptoms” and “behavior” can be confirmed at a glance is provided.

According to the present invention, when a plurality of classification target sentences are extracted from one text in the classification target sentence extracting step, an arbitrary classification target sentence is displayed in an initial stage of the hierarchical structure, and a middle stage of the hierarchical structure is displayed. Since other classification target sentences are configured to be displayed, for example, there is an effect that a case base that presents “actions” to be dealt with after narrowing down “symptoms” to some extent is effective.

According to the present invention, clustering is performed while associating the classification target sentence extracted in the classification target sentence extraction step with the existing cluster stored in the case base. There is an effect that it can be added while maintaining consistency with the existing classification system.

According to the present invention, since an editing step for editing a cluster stored in the case base is provided, even when a program performs an incorrect classification, it can be corrected and registered in the case base. effective.

According to the present invention, when storing the cluster generated in the cluster generation step in the case base,
When the similarity between the cluster and the existing cluster stored in the case base is higher than a predetermined threshold, the display process of the display step and the edit process of the edit step are configured to be omitted. There is an effect that the work load can be reduced.

According to the present invention, when storing the cluster generated in the cluster generation step in the case base,
If the similarity between the cluster and the existing cluster stored in the case base is lower than a predetermined threshold, the display process of the display step and the edit process of the edit step are omitted, and the cluster is not stored in the case base. , Because it is configured to save in a temporary storage file, every time a text such as a failure response occurs, the trouble of adding the case is released, and the case-based construction work can be performed efficiently. There is.

According to the present invention, there is provided a term extracting means for extracting a term included in a text, and a classification pattern generating means for generating a text classification pattern from the term, and performs text clustering based on the classification pattern. Because it is configured to execute, it is possible to perform classification based on words that have a common concept according to the user's purpose, and as a result, the effect of efficiently searching for the target case can be obtained. is there. In addition, there is an effect that a keyword serving as a reference for classification can be automatically extracted from a document and a document can be classified without designating individual keywords serving as a reference for classification.

According to the present invention, a term extraction processing procedure for extracting terms included in text and a classification pattern generation processing procedure for generating a text classification pattern from the terms are provided. Since it is configured to execute clustering, it is possible to perform classification based on words having a common concept according to the purpose of the user, and as a result, the target case can be efficiently searched. effective. In addition, there is an effect that a keyword serving as a reference for classification can be automatically extracted from a document and a document can be classified without designating individual keywords serving as a reference for classification.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing a case-based construction device according to a first embodiment of the present invention.

FIG. 2 is an explanatory diagram showing an example of a basic word dictionary.

FIG. 3 is an explanatory diagram showing an example of a term pattern description table.

FIG. 4 is an explanatory diagram showing an example of a classification pattern description table.

FIG. 5 is a flowchart showing a case-based construction method according to the first embodiment of the present invention.

FIG. 6 is an explanatory diagram showing an example of a text.

FIG. 7 is an explanatory diagram showing an example of a sentence tag determination table.

FIG. 8 is an explanatory diagram showing a failure response record to which a sentence tag is added.

FIG. 9 is an explanatory diagram showing word breaks obtained as a result of performing morphological analysis on the classification target sentence of FIG. 8;

FIG. 10 shows the format of the term pattern description table as BNF.
It is explanatory drawing represented by the notation.

FIG. 11 is an explanatory diagram showing a list of character types.

FIG. 12 is an explanatory diagram showing an example of a morpheme description.

FIG. 13 is a flowchart showing a matching process between a morphological analysis result and a pattern definition unit.

FIG. 14 is a flowchart showing a collation process in step ST35.

FIG. 15 is a flowchart showing a determination process in step ST34.

FIG. 16 is a flowchart showing an evaluation process in step ST71.

FIG. 17 is an explanatory diagram showing a correspondence between a morphological analysis result and a pattern definition unit.

FIG. 18 is an explanatory diagram showing a correspondence between a morphological analysis result and a pattern definition unit.

FIG. 19 is an explanatory diagram showing terms extracted from the failure handling record of FIG. 9;

FIG. 20 is an explanatory diagram showing an example of a synonym dictionary.

FIG. 21 is an explanatory diagram illustrating a term extraction result in which a spelling elimination process is executed.

FIG. 22 is an explanatory diagram showing an example of how to use “string”, “concept”, and “subconcept”.

FIG. 23 is an explanatory diagram showing a classification pattern generated for the morphological analysis result of FIG. 9;

FIG. 24 is a flowchart illustrating cluster generation processing.

FIG. 25 is an explanatory diagram showing an example of an IS-A dictionary.

FIG. 26 is an explanatory diagram showing an example of a HAS-A dictionary.

FIG. 27 is an explanatory diagram showing a cluster generation result.

FIG. 28 is an explanatory diagram showing a failure handling record to be added.

FIG. 29 is an explanatory diagram showing various processing results.

FIG. 30 is an explanatory diagram showing a cluster generation result after a failure response record has been added;

FIG. 31 is an explanatory diagram showing a display example of a clustering result of a tree structure.

FIG. 32 is an explanatory diagram illustrating an example of an edit screen.

FIG. 33 is an explanatory diagram showing a cluster editing process.

FIG. 34 is a configuration diagram showing a conventional case-based construction apparatus.

FIG. 35 is a flowchart illustrating processing of a document classification unit.

FIG. 36 is an explanatory diagram showing words serving as reference for classification.

FIG. 37 is an explanatory diagram showing words serving as reference for classification.

[Explanation of symbols]

11 text, 12 basic word dictionary, 13 term pattern description table, 14 classification pattern description table, 15
Input means, 16 sentence analyzing means, 17 term extracting means,
18 Classification pattern generation means, 19 Cluster generation means, 20 Output means, 21 Editing means, 22 Control means, 23 Case base, 101 OK button, 102
NG button, 103 clustering result display section, 10
4 Classification target sentence display unit, 105 Sentence tag display unit, 106
Text display section.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Katsushi Suzuki 2-3-2 Marunouchi, Chiyoda-ku, Tokyo F-term in Mitsubishi Electric Corporation (reference) 5B075 ND03 ND20 ND35 NK31 NK34 NK35 NK54 NR02 NR12 NR20 PR06 QP03 UU05 UU40

Claims (20)

[Claims] An input step of inputting a text; a term extraction step of extracting terms included in the text; a classification pattern generation step of generating a classification pattern of the text from the terms extracted by the term extraction step. A case base comprising: a cluster generation step of performing the text clustering based on the classification pattern generated by the classification pattern generation step; and an output step of storing the cluster generated by the cluster generation step in a case base. How to build. 2. A classification pattern generating step for determining a synonym of a term from the similarity of the notation of the term extracted in the term extracting step when generating a classification pattern of a text, and eliminating a fluctuation of the notation related to the synonym. The case-based construction method according to claim 1, wherein 3. A classification pattern generating step includes: referring to a synonym dictionary when generating a text classification pattern;
2. The method according to claim 1, wherein the synonym of the term extracted in the term extracting step is determined, and the fluctuation of the description of the synonym is eliminated. 4. The case according to claim 1, wherein the cluster generation step performs clustering with reference to an ontology dictionary in which vocabulary conceptual relationships are described. Base construction method. 5. The case-based construction method according to claim 4, wherein in the cluster generation step, an IS-A dictionary in which a relation between a superordinate concept and a subordinate concept of a word is described is used as an ontology dictionary. 6. The case-based construction method according to claim 4, wherein in the cluster generation step, a HAS-A dictionary in which a relationship between a higher-order configuration and a lower-order configuration of a device is described is used as an ontology dictionary. 7. A classification target sentence extraction step for extracting a classification target sentence to be processed in a term extraction step, a classification pattern generation step, and a cluster generation step from text input in the input step. The case-based construction method according to any one of claims 1 to 6. 8. A cluster generating step includes: when a plurality of classification target sentences are extracted from one text in the classification target sentence extracting step, clusters are generated in order from the top classification target sentence of the text to form a hierarchical structure. The case-based construction method according to claim 7, wherein the case-based construction method is created. 9. The classification pattern generation step refers to a classification pattern description table in which a vocabulary pattern serving as a classification reference is described when generating a text classification pattern, and the classification pattern description table has the same level. 9. The case-based construction method according to claim 8, wherein a vocabulary pattern is described for each hierarchy. 10. The display device according to claim 1, further comprising a display step of displaying, from among the clusters generated by the cluster generation step, a cluster to which a large number of texts belong, in order. The case-based construction method according to claim 1. 11. The case-based construction method according to claim 10, wherein in the display step, when a cluster to be displayed is designated, the full text of the text assigned to the cluster is displayed. 12. The case-based construction method according to claim 10, wherein in the display step, when a cluster to be displayed is specified, a reference target sentence assigned to the cluster is displayed. 13. The display step, wherein when a plurality of classification target sentences are extracted from one text in the classification target sentence extraction step, identification information for identifying a type of each classification target sentence is displayed. Item 13. The case-based construction method according to Item 10. 14. A display step, wherein when a plurality of classification target sentences are extracted from one text in the classification target sentence extraction step, an arbitrary classification target sentence is displayed in an initial stage of the hierarchical structure, and a middle stage of the hierarchical structure is displayed. The case-based construction method according to claim 10, wherein another classification target sentence is displayed from. 15. The method according to claim 7, wherein the cluster generation step executes clustering while associating the classification target sentence extracted in the classification target sentence extraction step with an existing cluster stored in the case base. The case-based construction method according to claim 14. 16. The case base construction method according to claim 1, further comprising an editing step of editing a cluster stored in the case base. 17. An output step, when storing a cluster generated by the cluster generation step in a case base, when a similarity between the cluster and an existing cluster stored in the case base is higher than a predetermined threshold value, 17. The case-based construction method according to claim 16, wherein the display process of the display step and the edit process of the edit step are omitted. 18. The output step, when storing the cluster generated by the cluster generation step in the case base, when the similarity between the cluster and the existing cluster stored in the case base is lower than a predetermined threshold, 17. The case base construction method according to claim 16, wherein the display processing of the display step and the edit processing of the edit step are omitted, and the cluster is stored in a temporary storage file without being stored in the case base. 19. Input means for inputting text, term extraction means for extracting terms included in the text, classification pattern generation means for generating a classification pattern for the text from the terms extracted by the term extraction means A case base comprising: cluster generation means for performing clustering of the text based on the classification pattern generated by the classification pattern generation means; and output means for storing the cluster generated by the cluster generation means in a case base. Construction equipment. 20. An input processing procedure for inputting a text,
A term extraction processing procedure for extracting terms included in the text, a classification pattern generation processing procedure for generating a classification pattern of the text from the terms extracted by the term extraction processing procedure, and a classification pattern generation procedure for generating the classification pattern Recording a case-based construction program having a cluster generation processing procedure for executing the clustering of the text based on the classified pattern and an output processing procedure for storing the cluster generated by the cluster generation processing procedure in a case base Medium.