JP2001188678A - Language case inferring device, language case inferring method, and storage medium on which language case inference program is described - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve such a problem that the degree of similarity of a detailed natural language including aspect expression of a sentence cannot be calculated and the case including the sentence which is syntactically and semantically similar to a retrieval sentence cannot be retrieved since calculation of the degree of similarity is executed only by using frequency information and category classification of a keyword extracted from the case. SOLUTION: When a keyword is extracted from at least one or more case sentences, a cluster to which each case sentence belongs is decided and a problem solving tree to be characterized by a set of each keyword is generated on the basis of a decision result of the cluster, the cluster to which each case sentence belongs is segmentised on the basis of the attribute of language expression of each case sentence which belongs to each cluster of the problem solving tree.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a language case inference apparatus, a language case inference method, and a language case inference program for classifying past case sentences, registering them in a case database, and searching for a case sentence similar to a search sentence. It relates to the described storage medium.

[0002]

2. Description of the Related Art There is a strong demand for reusing data accumulated by electronic means in exchanges of inquiries at a consultation desk and at a failure monitoring center at a failure monitoring center. In response to such a request, a case sentence expressed in natural language is registered in the case database in advance,
Conventionally, there is a language case inference apparatus that searches for a case sentence similar to a search sentence that is a new input sentence.

FIG. 28 is a block diagram showing a conventional language case inference apparatus disclosed in, for example, Japanese Patent Application Laid-Open No. 9-73364. In the drawing, reference numeral 1 denotes a past case described in a natural language sentence; Case database that registers case 1
Is an attribute information generating unit that extracts keywords from [problems] of the case 1 stored in the case database 2 and generates attribute information of the case 1 by referring to the case number table 7.

[0004] Reference numeral 4 denotes a keyword number table storing keyword numbers specifying keywords extracted by the attribute information generating means 3, and reference numeral 5 denotes a case number storing case numbers specifying case 1 and information specifying keywords appearing in each case. Is a keyword table, 6 is a category table for storing a category number indicating the classification result of each case,
7 is the number t _{ij of} keyword numbers corresponding to category numbers
A case number table for storing attribute information 9; an attribute database 8 for storing attribute information 9; 9; attribute information including a case number, a category number, a keyword number, and a weight for each keyword; Is a similarity generating means for generating the similarity of.

Next, the operation will be described. First, when the past case 1 is accumulated, the attribute information generating means 3 extracts a keyword from the [problem] of the case 1 stored in the case database 2 and refers to the case number table 7 to find the category C _i. to generate a weight ω _ij for example the total number S _i, the total number t _j of cases attribute of the attribute number j _appears, the attribute of the attribute number j of category C _i of belonging to, to store these attribute information in the attribute database 8 . ω _ij = (t _ij / t _j −S _i / S) + (t _ij / S _i
−t _j / S) S = S ₁ + S ₂ +... + S _n

Next, when searching for a case, the similarity generating means 10 extracts a keyword from the newly input search sentence and obtains a keyword number from the keyword number table 4. Next, the similarity generation means 10 searches the attribute database 8 from a list of keywords extracted from the search sentence, extracts attribute information including a matching keyword,
The similarity G is calculated. G = (sum of weights) × (number of matched keywords / length of keyword list extracted from attribute database 8) Then, similarity generating means 10 sorts the case numbers in descending order of similarity G, and Display or print the contents.

[0007]

Since the conventional language case inference apparatus is configured as described above, similarity calculation is performed using only frequency information and category classification of keywords extracted from the case. There is a problem that it is not possible to calculate a detailed natural language similarity including a sentence modal expression and the like, and it is not possible to search for a case including a sentence syntactically and semantically similar to a search sentence. Further, in general, if it is attempted to determine the similarity of all cases using only the syntax structure of a sentence, there has been a problem that a great deal of processing time is required.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and it is possible to efficiently classify cases to construct a problem solving tree, and to provide a search statement and syntactical
It is an object of the present invention to provide a language case inference apparatus, a language case inference method, and a storage medium in which a language case inference program is described, which can search for a case including a sentence having a semantically similar sentence.

[0009]

SUMMARY OF THE INVENTION A language case inference apparatus according to the present invention extracts a keyword from at least one or more case sentences, determines a cluster to which each case sentence belongs, and determines each keyword based on the determination result. Keyword clustering means for generating a problem solving tree characterized by a set of, and each case sentence belongs based on the attribute of the linguistic expression of each case sentence belonging to each cluster generated by the problem solving tree generated by the keyword clustering means. And a detailed clustering means for subdividing a cluster.

The language case inference method according to the present invention extracts a keyword from at least one case sentence, determines a cluster to which each case sentence belongs, and is characterized by a set of keywords based on the determination result. When the problem solving tree is generated, the cluster to which each case sentence belongs is subdivided based on the attribute of the linguistic expression of each case sentence belonging to each cluster of the problem solving tree.

In the language case inference method according to the present invention, when extracting a keyword from a case sentence, the case sentence is extracted from the digitized document, and a tag indicating the type of the case sentence is added.

In the language case inference method according to the present invention, when generating a problem solving tree, the number of hierarchical levels of a cluster is specified.

In the language case inference method according to the present invention, when subdividing the cluster to which each case sentence belongs, the minimum similarity between the case sentences constituting the cluster is specified.

In the language case inference method according to the present invention, when a keyword is extracted from a case sentence, a word having an appearance frequency higher than a set value among words in the case sentence is extracted as a keyword.

In the language case inference method according to the present invention, when a keyword is extracted from a case sentence, an ontology that holds knowledge of a word depending on a target area or a task and a relationship between the words is referred to, and the case sentence in the case sentence is referred to. Among the words, words held in the ontology are extracted as keywords.

In the language case inference method according to the present invention, when subdividing a cluster to which each case sentence belongs, the language refers to an ontology that holds knowledge about a word that depends on a target area or business and a relationship between the words, The knowledge about the relationship between words and words held in the ontology is used for detailed clustering.

In the language case inference method according to the present invention, the ontology has IS-A relation knowledge indicating a semantic upper-lower relation.

In the language case inference method according to the present invention, the ontology has HAS-A relation knowledge indicating a semantic part-whole relation.

In the language case inference method according to the present invention, the ontology holds case knowledge.

The language case inference method according to the present invention is such that the ontology holds paraphrase knowledge.

In the language case inference method according to the present invention, the ontology holds conflicting knowledge.

In the language case inference method according to the present invention, when performing similar sentence matching between a case sentence belonging to a cluster and a search sentence, matching of a semantic structure is executed based on attributes of syntactic elements. is there.

In the language case inference method according to the present invention, when performing similar sentence matching between a case sentence belonging to a cluster and a search sentence, the degree of detail of matching is specified.

In the language case inference method according to the present invention, the depth of a tree structure in a parse tree is specified as the level of detail of collation.

In the language case inference method according to the present invention, the similarity between each cluster is described in a database.

In the language case inference method according to the present invention, the conflicting relation between each cluster is described in a database.

The storage medium in which the language case inference program according to the present invention is described extracts keywords from at least one or more case sentences, determines a cluster to which each case sentence belongs, and determines each keyword based on the determination result. Based on the keyword clustering processing procedure for generating a problem solving tree characterized by a set of, and the attribute of the linguistic expression of each case sentence belonging to each cluster generated by the keyword clustering processing procedure. And a detailed clustering processing procedure for subdividing the cluster to which the belongs.

[0028]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a configuration diagram showing a language case inference apparatus according to Embodiment 1 of the present invention.
Is a digitized document such as an inquiry record including a case sentence described in a natural language, 12 is a sentence extracting unit for extracting at least one or more case sentences from the digitized document 11, and 13 is extracted by the sentence extracting unit 12. A sentence analysis unit that performs morphological analysis and syntax analysis of each of the case sentences and the search sentence 20 input by the search sentence input unit 21, and a keyword clustering unit 1.
5 and a detailed clustering unit 16 to classify each case sentence.

Reference numeral 15 denotes a problem solving tree characterized by extracting a keyword from each case sentence extracted by the sentence extraction unit 12, determining a cluster to which each case sentence belongs, and characterizing the set of each keyword based on the decision result. A keyword clustering unit (keyword clustering means) for generating
A detailed clustering unit (detailed clustering means) 16 subdivides the cluster to which each case sentence belongs based on the attribute of the linguistic expression of each case sentence belonging to each cluster generated by the keyword clustering unit 15. Is a case database that stores the results of clustering by the detailed clustering unit 16 and the results of analysis by the sentence analysis unit 13, 18 is an ontology that holds knowledge on words that depend on the target area or business and the correspondence between each word, and 19 is details This is a case cluster editing unit for editing a clustering result or the like by the clustering unit 16.

Reference numeral 20 denotes a search sentence, reference numeral 21 denotes a search sentence input unit for inputting the search sentence 20, and reference numeral 22 extracts a keyword from the search sentence 20 and searches a cluster having the keyword as an element from a problem solving tree. Is a similar case search unit (similar case search means) for executing similar sentence matching between the case sentences belonging to the search sentence 20 and the search sentence 20, and a search result display unit 23 for displaying the search results and the like of the similar case search unit 22. FIG. 2 is a flowchart showing a language case inference method (case construction processing) according to Embodiment 1 of the present invention. Note that all of the language case inference apparatus in FIG. 1 may be configured by software, and a storage medium in which a computer-readable language case inference program is described may be prepared.

Next, the operation will be described. First, in step ST1, the sentence extraction unit 12 extracts a sentence to be processed such as a case sentence from the digitized document 11 such as an inquiry record.

FIG. 6 is an explanatory diagram showing an example of the digitized document 11. In some cases, a plurality of documents are stored in one digitized document file. In the example of FIG. 6, a document start tag <DOC> and a document end tag </ DOC> indicate a document break. ing. One document is composed of a plurality of document fields. For example, as shown in FIG. 6, a fixed field portion in which contents such as <customer name>, <customer phone>, <model>, and <subject> can be expressed by a number or a simple character string, and a plurality of natural language sentences It consists of a non-standard field part such as <question> and <answer> to be described.

FIG. 3 is a flowchart showing the specific contents of the sentence extraction process (step ST1). Processing for reading the digitized document 11 as shown in FIG. 6 will be described. First, in step ST11, the first document start tag is extracted. In step ST12, it is determined whether or not the document start tag has been successfully extracted. If the document start tag can be taken out, step ST13
, A unique case number is assigned to process the document A as a “case”.

Next, in step ST14, a sentence extraction process is performed on one document, and when the sentence extraction process on one document is completed, a document end tag is checked in step ST15. If there is no document end tag, a warning is issued. And so on. Next, in step ST16, a document start tag for the next document B is extracted, and the same processing is repeated from step ST12. In step ST12, the above processing is repeatedly executed until the removal of the document start tag is not detected or the end of the document file is reached.

FIG. 4 is a flowchart showing the specific contents of the sentence extraction process (step ST14) in FIG.
In the sentence extraction process for one document (step ST14), when the digitized document 11 is received, first, in step ST21, the document field start tags (enclosed by the symbols “<” and “>”) in the digitized document 11 are received. A character string (hereinafter abbreviated as a start tag) is checked to see if it exists, and if there is a start tag, the character string is extracted as a field name. In the example of FIG. 6, the character string "customer name" is extracted from <customer name> as a field name. At this time, if there is no start tag in step ST22, the digitized document 1
The processing for 1 is completed.

If there is a start tag in step ST22, it is determined in step ST23 whether or not the field is a fixed field. If the field is a fixed field, step ST24
In FIG. 6, for example, in FIG. 6, when the field is <customer name>, the character string “Taro Yamada” is registered in the corresponding column of the case database 17 as it is. Subsequently, in step ST25, a document field end tag (a character string surrounded by a symbol “<//” and a symbol “>”),
For example, it is confirmed whether or not </ customer name> is present, and if not, a warning is displayed.

On the other hand, if it is determined in step ST23 that the field is an irregular field,
For example, in the case of the <question> tag or the like in FIG. 6, in step ST26, an atypical field process is executed by passing the case number and the field name (specific processing contents will be described later). Step ST25 or step ST26
Is completed, in step ST27, the start tag of the next document field is extracted, and the process proceeds to step ST2.
The same processing is repeated from step 2.

It should be noted that the determination of the fixed field and the non-fixed field in step ST23 is performed by referring to document field information as shown in FIG. In the document field information, field names (for example, customer names and questions) and standard / non-standard information indicating whether each field is standard or non-standard is defined in advance. In addition, document field information includes options such as "category attribute" for classifying a large set of case data and "title attribute" for identifying each case when displaying a list of case data. Information is defined as needed.

FIG. 5 shows an atypical field process (step S).
It is a flowchart which shows the specific content of T26). In step ST31, when the current case number and the current field name are passed, a period (.) Or a bullet is selected from a plurality of sentences described in the corresponding atypical field in the digitized document 11. One sentence is extracted by judging with (•), a character string representing an itemized list such as a chapter number, or the like.

In the example of FIG. 6, "sound from air conditioner" is first extracted as a sentence from the <question> field. The sentence extracted here is called a “case sentence” constituting the case data. If a case sentence can be extracted in step ST31, the case sentence is assigned a case number, a field name, and a sentence number indicating the number of the sentence in the field. In step ST32, it is checked whether or not a case sentence can be taken out. If the case sentence cannot be taken out, it is assumed that all the sentences in the field have been taken out, and the end tag of the document field is confirmed in step ST35. Then, the atypical field processing ends, and the process returns to step ST27 of FIG.

If a case sentence is extracted in step ST32, a "sentence tag", which is a label indicating the type of the sentence, is added to the case sentence in step ST33.
Register to 7. In step ST33, a sentence tag list as shown in FIG. 9 is referred to as information for adding a sentence tag. In the sentence tag list, a sentence tag corresponding to a key expression (condition part) in the sentence is defined in advance for each field.

In FIG. 9, for simplicity of explanation,
In this example, the sentence tag is determined using the character string of the surface layer (the characters included in the sentence). However, it is necessary to allow the specification of the sentence pattern, negation, guesswork, etc. in the condition part. Alternatively, a sentence tag may be determined by performing a sentence analysis, which will be described later, on the case sentence and then taking correspondence with the sentence tag list.

FIG. 7 is an explanatory diagram showing "cases" generated from the digitized document 11 shown in FIG. 6 by the sentence extraction step ST1. The case is, as shown in FIG.
Fixed fields such as Customer name> and their values, field names, statement numbers in the fields, statement tags, statement numbers for each statement tag,
It consists of an atypical field consisting of a case sentence.

FIG. 7 shows a case where the sentence extraction process in step ST31 and the sentence tag attaching process in step ST33 are all successful. However, these processes are performed due to sentence delimitation and ambiguity of sentence tag attaching conditions. There are mistakes. At that time, the sentence extraction unit 12 is provided with an editing / display function, and the case is displayed to the user in a form as shown in FIG.
Can be modified.

Next, the sentence analysis processing of FIG. 2 (step ST
In 2), the sentence analysis unit 13 performs a morphological analysis and a syntax analysis on each sentence of a set of sentences to be processed, and generates a syntax structure. FIG. 10 shows a sentence analysis process (step ST2).
3 is a flowchart showing the specific contents of FIG.

The morphological analysis in step ST41 refers to the analysis word dictionary 24, and the syntax analysis in step ST42 refers to the ontology 18. Also,
The analysis word dictionary 24 may be included in the ontology 18, and the morphological analysis process may refer to the ontology 18. In step ST43, the syntax structure (dependency structure) generated by the syntax analysis process in step ST42 is
As shown in FIG. 17, it is stored in the case database 17 in association with the case sentence.

FIG. 11 shows a morphological analysis process (step ST4).
It is a flowchart which shows the specific content of 1). In the morphological analysis process of step ST51, the analysis word dictionary 2
4. Referring to the auxiliary word dictionary 25 and the auxiliary word connection table 26,
The case sentence (for example, “timer does not enter in RC”) is divided into morpheme columns. Morphological analysis method, auxiliary word dictionary 25
The attached word connection table 26 is described in detail in many documents, and a description thereof will be omitted. In the following description,
As in the morphological analysis result of FIG. 11, morpheme delimiters are abbreviated with "/".

The result of the morphological analysis is, from the independent word part, heading information, part-of-speech information described in the analysis word dictionary 24, and “semantic symbols” which are concept information indicating the type of word.
Is configured to be able to be referred to. At this time, pointer information for referring to the dictionary may be held, or when it takes time to refer, for example, when the dictionary exists in the secondary storage, the information may be copied to the primary storage. Note that, in the following description, words such as <>
And represent a semantic symbol.

FIG. 12 shows a syntax analysis process (step ST4).
It is a flowchart which shows the specific content of 2). In the phrase structure generation processing of step ST61, the morphological analysis output result of FIG. 11 is input, and a phrase structure that is a basic unit for performing dependency analysis is generated. The phrase structure is composed of at least one independent word morpheme and 0 or more adjunct word morphemes connected to the independent word morpheme. FIG. 19 is an explanatory diagram showing an example of a phrase structure. The phrase structure in FIG. 19 includes a dependency attribute, a receiving attribute, independent word information, and attached word information. The independent word information is a pointer to the independent word morpheme information forming the phrase, and the attached word information is an array of pointers to zero or more attached word morpheme information.

In the dependency analysis processing in step ST62, the dependency analysis of the phrase structure is executed in accordance with the grammar rule 27, and the dependency structure is generated as a result of the syntax analysis. Since the dependency analysis method is described in many documents and the like, a detailed description is omitted here. In general, a large number of ambiguities are generated during the dependency analysis, but the ontology 18 is appropriately referred to in order to resolve the ambiguity.

The sentence analysis processing in FIG. 2 (step ST
When 2) is completed, next, in step ST3, a set of syntax analysis results of the processing target sentence is classified for each similar sentence while calling the keyword clustering unit 15 and the detailed clustering unit 16 (clustering), and shown in FIG. A problem solving tree composed of clusters including a set of similar sentences is generated. Each cluster of the problem solving tree has a set of similar case sentences, and each case sentence is stored in the case database 17 in a format as shown in FIG.

FIG. 13 shows a similar case classification process (step ST).
It is a flowchart which shows the specific content of 3). The similar case classification process (step ST3) includes a keyword clustering process (step ST71) and a detailed clustering process (step ST72). Each clustering process refers to the ontology 18 as needed, and stores the classification (clustering) result in the case database 17.

FIG. 14 is a flowchart showing the specific contents of the keyword clustering process (step ST71), and FIG. 15 is a flowchart showing the specific contents of the recursive function calling process of FIG. Prior to the description of FIGS. 14 and 15, symbols and mathematical expressions used in FIGS. 14 and 15 will be described.

For example, W _x , W _y ,..., W _z represent keywords selected from words appearing in the case, and J (W _x ∧W _y ∧... ∧W _z ) is the keyword W _x , W
_{y, ...,} W _z is assumed to represent a list of case number that appears at the same time. Also, the number of elements in the case number list R is | J
_{(W x ∧W y ∧ ... ∧W} z) | represented by the keyword _{W x,} W y, ..., mutual information indicating the strength of the binding of _{W z} I
(W _x , W _y ,..., W _z ) is defined by the following equation.

[0055]

(Equation 1)

Also, kwnum1 is the number of candidates for the first keyword, kwnum2 is the number of candidates for the second and subsequent keywords, minJirei is the minimum number of cases as a stop condition for stopping clustering by a set of keywords, and usedKW is used once. The variable for registering the keyword, scoreTH, is a threshold (cluster establishment condition) of mutual information constituting a cluster (a group of similar sentences) on the assumption that the degree of association of the set of keywords is high.

First, in step ST81, a keyword is extracted from the analysis result of the case sentence with reference to the case database 17. As the keyword, a word whose frequency of appearance of the word is higher than a set value may be extracted, or a word held in the ontology 18 may be extracted. For example, when a meaning symbol of <model> is held in the ontology 18, even if the appearance frequency is low, it can be used as a keyword.

Next, in step ST82, a correspondence table between keywords and case numbers is created. In the correspondence table, keywords are sorted in descending order of appearance frequency. In step ST83, an empty set is set to usedKW, which is a variable for registering the keyword once used, and the keyword is initialized. In step ST84, the counter i indicating the number of keywords of the first type is set to "1".
Set to.

In step ST85, i ≦ kwnu
If m1 (the maximum number of keywords of the first type) is satisfied, the process proceeds to step ST86; otherwise, the process proceeds to step ST90. In step ST86, step ST if it is not element of the keyword _{W i} is usedKW
The process proceeds to step ST89 if it is an element.

[0060] In step ST87, to register the keyword _{W i} to the elements of the usedKW. In step ST88, the number of keyword sets N = 1 as a parameter
Then, a process of calling a recursive function clstKWsub, which is a sub-process of keyword clustering, is passed by passing the keyword set {W _i } (details of the process will be described later). When the clustering sub-process is completed, in step ST89, the value of the counter i is incremented, and the process proceeds to step ST85.
Repeat the process from. In step ST90,
A problem solving tree (tree structure information) based on a keyword set as a result of keyword clustering is stored in the case database
7 is stored.

Referring to FIG. 15, the details of the recursive function calling process (step ST88), which is the keyword clustering sub-process, will be specifically described. First, in step ST91, the counter j is set to "1". Here, a description will be given assuming that the maximum value of j is kwnum2. It is assumed that the value of kwnum2 can be changed according to the cluster creation status, such as the size of the entire case and the desired processing time. The same applies to the value of kwnum1.

In step ST92, j ≦ kwnu
If m2 is satisfied, the process proceeds to step ST93, and if not, the keyword clustering sub-process ends. In step ST93, the keyword W _j is usedK
If it is not an element of W, the process proceeds to step ST94, and if it is an element, the process proceeds to step ST98. When j = 1, W ₁
Then proceeds to step ST98 because it is an element of UsedKW, it becomes a j = 2, since _{W 2} is not an element of UsedKW, the process proceeds to step ST94.

[0063] Subsequently, in step ST94, the condition _{| J (W i1 ∧W i2 ∧} ... ∧W iN) | ≧ minJir
ei (if the number of cases falls below minJirei, no further subdivision), and I (W _i1 , W
_{i2, ...,} W iN, if _{W j)} ≧ _scoreTH (mutual information threshold) is satisfied, the process proceeds to step ST95, does not hold the process proceeds to step ST98. For example, FIG.
If in step ST87 of i = 1 in the case of j = _{2, W i1} coincides with _{W 1, W} i2 _{~W iN} is absent.

In step ST95, the result storage array kw
For p [N + 1], the keyword set {W _i1 ,
_{W i2, ..., W iN,} W j} the degree of coupling is registered as a high set of keywords. For example, when j = 2, ｛W ₁ , W
₂ ) is registered as a set of keywords having a high degree of connection. Subsequently, in step ST96, to add a keyword _{W j} to usedKW. Further, in step ST97, N +
1 and the keyword set { _Wi1 , _Wi2 , ..., _WiN , W
_The keyword clustering sub-process clstKWsub is recursively called using _j } as a parameter. In step ST98, the value of the counter j is incremented by one, and the same processing is repeatedly executed from step ST92. In step ST92, if j> kwnum2, the keyword clustering sub-process clstKWsub ends.

FIG. 16 is an explanatory diagram showing an example of a hierarchical structure solution tree generated by the keyword clustering process (step ST71). In FIG.
The word “AMiTY” is a word adopted as a keyword because the ontology 18 has a semantic symbol of <model>.

In FIGS. 14 and 15, the keyword used once is registered in the usedKW and is not reused (in step ST83, the usedKW is used).
W is set to an empty set), a reusable keyword can be set in advance, and the keyword can be used again. In FIG. 16, the keyword “power” appears in two places, but the reusable designation has been made.

FIG. 18 is a flowchart showing the specific contents of the detailed clustering process (step ST72). First, in step ST101, the number N of clusters serving as leaves of the hierarchical problem solving tree generated by the keyword clustering process (step ST71)
Ask for.

In step ST102, a repetition counter i is initialized to "0". In step ST103, the counter i is compared with the number N of leaf clusters.
The process proceeds to T104, and if not smaller, proceeds to step ST106.

In step ST104, the detailed clustering sub-process is applied to the set of case sentences included in the corresponding leaf cluster i, and the set of case sentences is subdivided. In this sub-process, for example, pairs of two sentences that are most similar are sequentially grouped into one cluster. When the two found pairs are equal to or smaller than a certain threshold, the grouping ends. At this time, a predetermined number of hierarchies may be created. Although the detailed clustering of the leaf clusters has been described here for simplicity, a configuration may be adopted in which, for example, clustering is appropriately called for a word during the keyword classification process. Step S
At T105, the counter i is incremented by one, and at step ST10
The same processing is repeatedly executed from step 3. Step ST1
At 06, the cluster information of the case sentences classified in detail is stored in the case database 17.

Next, the similarity calculation in the detailed clustering sub-process (step ST104) will be described with reference to FIG. Calculation of sentence similarity in the detailed clustering process (step ST72) using the ontology 18 is as follows:
This is performed by similar sentence matching in which nodes in the syntax structure are associated with each other and weighted by attributes in the nodes (a process of calculating a modal expression such as negation or guesswork).

In the similarity calculation function Sim (A, B, D) for similarity between natural language sentences in similar sentence matching, arguments A and B are dependency structures which are the results of syntax analysis, and are shown in FIG. Suppose that it has such a tree structure. The argument D is
FIG. 24 is a detail level of the collation when the similarity calculation is performed.
In the calculation of the degree of similarity between the two tree structures (a) and (b) shown in FIG. 7, the value indicates the number of hierarchies from the root node to be processed. Here, for the sake of simplicity, description will be made assuming that D = 2.

First, an initial similarity of 1.0 is given. A similarity of 1.0 means that the two input sentences have exactly the same meaning. In the following processing, the information of each node is compared while tracing the tree structure, a “penalty” is given to a different part, the difference is reduced from 1.0, and when the similarity becomes “0” or a predetermined value, The comparison target is regarded as dissimilar and the similarity calculation is stopped. FIG. 26 is an explanatory diagram showing an example of a penalty calculation rule in similarity calculation.

First, comparison between first level nodes is performed. Here, the semantic symbols are equal in <detection operation>,
Since the words on the surface layer are different, a penalty value of −0.01 is given according to rule 01 in FIG. Next, a second level comparison is performed. At this time, no penalty is given to the left node (sensor <monitor device>) because the information completely matches. The node on the right has a different semantic symbol, so the penalty value −
Give 0.3. In this way, the similarity is calculated as 0.69.

If the argument D is given as "1", the similarity is calculated only at the first level, and the similarity is calculated as 0.99. As described above, since the accuracy of the similarity calculation can be controlled by the value of D, flexible processing according to the search situation is possible.

When the rule 02 of FIG. 26 is applied to the structure of FIG. 24A, the sentence "monitor recognizes clogging" by IS-A (upper-lower) knowledge 31 shown in FIG. Is 0.9. On the other hand, FIG.
Is applied to the dependency structure of the sentence "timer does not enter by RC", the similarity with "timer enters by RC" is 0.1 because the syntax analysis result includes negation as a node attribute. And are not similar. Similarly, if a sentence includes a guess expression such as “-Would be” and the attribute of the node of the parse result includes guesswork, the rule 0
6, the similarity obtained by subtracting -0.15 as a penalty is calculated.

In the same manner as described above, the rule “input of the timer button is not accepted” is applied to the rule 0 in FIG.
Applying No. 3 increases the similarity with the sentence “Remote control input is not accepted” using the HAS-A (partial-whole) knowledge 32 shown in FIG.

The HAS-A knowledge 32 of the ontology 18 can describe conflicting information. In the example of the HAS-A knowledge 32 relating to <timer display> in FIG. 23, the conflicting information 35 indicates that <timer display> has states of “none”, “on timer”, and “off timer”, but “none”. "When,
The "ON timer" or the "OFF timer" are not displayed at the same time, indicating that they are contrary. With this knowledge, when there is a sentence "No timer display" and a sentence "Timer display has an input timer" in the case sentences, the case database 17 as inter-cluster information.
Can be used to store information that the cluster containing each sentence is conflicting.

The paraphrase knowledge 34 of the ontology 18 includes:
You can write expressions that have the same meaning. The paraphrase knowledge 34 of FIG. 23 describes paraphrase knowledge that “<breaker> flies” and “<breaker> falls” have the same meaning, and by referring to the knowledge at the time of similarity matching, The sentence "breaker flies" and "breaker falls" can be treated as similarity 1.

The case relation knowledge 33 of the ontology 18 can describe the case relation between the body language and the declinable language of the natural language expression. For example, in the analysis word dictionary 24, the words "connect" and "connect" have a meaning symbol of <connection operation>, and the words "drain pipe" and "drain hose" have a meaning symbol of <drain pipe>, When the word “outdoor unit” has the meaning symbol “<outdoor unit>”, referring to the case knowledge 33, the sentence “Connect the drain pipe to the outdoor unit” and “connect the drain hose to the outdoor unit”
Can be determined as similar.

FIG. 20 shows a problem created by performing a detailed clustering process (step ST72) using a part of the problem solving tree based on the result of the keyword clustering of FIG. 16 and the accessory word information of the phrase structure illustrated in FIG. It is explanatory drawing which shows a solution tree.

In FIG. 2, when the similar case classification process (step ST3) is completed, next, in the case cluster editing process (step ST4), the hierarchy of the clusters of the case data stored in the case database 17 is shown in FIG. It is displayed in the form of a problem solving tree or the problem solving tree shown in FIG. FIG. 27 shows an example of a problem solving tree created from the set of digitized documents 11 shown in FIG. FIG. 27 shows an example of the case sentences extracted in the sentence extraction process (step ST1).
This is an example of a problem solving tree created by performing similar case classification on the first, second, and third sentences having [symptom] as a sentence tag from the <question> field. This problem solving tree is created for each value of the category shown in FIG.

In FIG. 27, the display of each cluster is composed of the number of cases (sentences) associated with the cluster, a check box used for searching, and a display label of the cluster set for display. When the cluster is located at the end of the problem solving tree, an end display is provided so that the cluster can be recognized as the end. At this time, the display in the case cluster editing unit 19 is displayed from top to bottom in the descending order of the number of case sentences included in the cluster. A cluster is composed of a set of similar case sentences, and information in the cluster can be referred to by designating a display label of the cluster and operating the mouse or the like. The case cluster editing unit 19 can set and edit the character string of the display label of the cluster by a user operation.

FIG. 17 is an explanatory diagram showing an example of the case sentence data stored in the case database 17. The output of the sentence extraction process (step ST1) includes the sentence analysis process (step S1).
After the dependency structure is added by T2), the cluster to which the case sentence data belongs is determined in the similar case classification process (step ST3), the cluster number is assigned, and the cluster is stored in the case database 17. At this time, information representing the hierarchy between clusters as shown in the problem solving tree of FIG. 27 is also stored in the case database 17 at the same time.

FIG. 25 is an explanatory diagram showing an example in which the inter-cluster information other than the hierarchical relation is stored in the case database 17 in step ST93 of the similar case classification process (step ST3). The inter-cluster information defines a relationship between two clusters represented by a cluster number.
Examples of relation types include -1 (reciprocal) and 1 (similar). The conflicting relationship is created with reference to conflicting information 35 exemplified in the HAS-A knowledge 32 of FIG. In the similarity relation, a representative sentence representing the cluster can be set in each cluster, and the similarity between the representative sentences is stored as inter-cluster information. FIG. 25 shows an example of the relationship between clusters shown in the problem solving tree of FIG.

FIG. 21 is a flowchart showing a language case inference method (case search process) according to Embodiment 1 of the present invention. Hereinafter, the case search process will be described with reference to FIG. First, in step ST111, the user inputs a search sentence (new problem description) for searching for a desired document using the search sentence input unit 21 of FIG. At this time, the search sentence input unit 21 may be a character recognition device or a voice recognition device as well as a keyboard.

Next, in step ST112, the search sentence input by the search sentence input unit 21 is analyzed. The sentence analysis process in step ST112 is, as described in the construction process, a morphological analysis process (step ST41).
Then, the search sentence is analyzed in the order of the syntax analysis process (step ST42), and a dependency structure for the search sentence is generated. Since the details of each analysis process are the same as those in the case construction process, the description is omitted here.

FIG. 22 shows a similar case search process (step ST).
It is a flowchart which shows the specific content of 113). First, in step ST121, the number M of primary search results is initialized to “0”, and in step ST122,
The counter k is initialized to “1”.

Next, in step ST123, a primary search of a case sentence is performed on the case database 17. That is, if the storage amount of the case database 17 is small, it is conceivable to perform similarity matching between the input search sentence and all the case sentences. However, since a large amount of case sentences are generally stored, all the case sentences When the collation processing is performed, there is a problem that a long processing time is required.

Then, for the dependency structure of the case sentence,
An index using independent words of the syntax element of the structure (see FIG. 17) and an index based on a set of words resulting from keyword clustering are provided, and a primary search process is performed using this index to determine the similarity matching target. Case sentences to be narrowed down in advance. Therefore, in the similar case search process (step ST113) of FIG. 21, the similar sentence collation described in the detailed clustering process (step ST72) is called, and the dependency analysis result of each case sentence in the primary search result and the input. Performs collation with the dependency analysis result of the search sentence.

Next, when the primary search processing of step ST123 is completed, steps ST124 to ST1 are executed.
The loop processing up to 26 is executed for each case sentence of the primary search result. In step ST125, the similarity between the k-th case sentence S _k (FIG. 24 (b): dependency structure) and the input search sentence S ₀ (FIG. 24 (a): dependency structure) of the primary search result is obtained. Degree calculation Sim (S ₀ , S _k ,
D) is performed, and the calculation result is output in step ST126. The value of D is set as needed, as described above.

Finally, when the end of the loop is detected in step ST124, in step ST127,
The primary search results are sorted in the order of similarity, and in step ST114 of FIG.
The result is output to the search result display section 23 of FIG. According to this method,
Even if they completely match at the keyword level included in the sentence, the similarity decreases if the meaning and content of the sentence are different, and even if the keywords included in the sentence are different, similar sentence matching using the ontology 18 , The similarity of sentences having semantically the same contents is increased.

In a conventional search using only keywords, such detailed processing is not performed. Therefore, if a negative sentence or modal expression is included, a sentence having a significantly different meaning from the search sentence is output as a search garbage at a higher position. However, according to this method, for example, a sentence whose similarity is equal to or less than the threshold value is set not to be displayed, so that the sentence can not be output as a similar sentence.

As described above, at the time of retrieval, in the sentence analysis process (step ST112), the dependency structure for the input search sentence is generated, and the similar case search process (step S112) is performed.
At T113), a search using the dependency structure is performed, and while referring to the ontology 18 in the similar sentence matching,
By calculating the similarity between the input search sentence and the search result, and outputting the search result to the search result display unit 23 in accordance with the similarity in the search result display process (step ST114), various natural language A case search with less search garbage is realized for expressions.

Search result display processing (step ST114)
In the method of displaying the search result in the search result display unit 23, the problem solving tree shown in FIG.
May be displayed by highlighting a cluster including a case sentence similar to the input sentence. FIG. 27 shows an example in which, for an input search sentence “I hear an unpleasant sound”, the cluster labeled “Sound of concern” includes the most similar case sentence “I hear a strange sound”. In the example shown in FIG.

Further, the search result display unit 23 displays the problem solving tree from the top in the descending order of the number of case sentences included in the cluster, similarly to the case cluster editing unit 19. It becomes easy to refer to. The user can designate each cluster, refer to the case corresponding to the case sentence included in the cluster, and display the case as shown in FIG. One cluster includes a plurality of case sentences, each of which corresponds to another case. To display a list of cases when referring to the cases, the contents of the cases corresponding to the title attribute fields shown in FIG. 8 may be displayed in a list.

As shown in FIG. 25, the case database 1
7 stores inter-cluster information. The inter-cluster information defines a reciprocal relationship and a similar relationship between two clusters represented by cluster numbers. By using this inter-cluster information, for example, when a cluster “body lamp not lit” in FIG. 27 is searched for an input sentence “no timer display”, the user determines that the state of the cluster is correct. When the check box is selected, the similar cluster “main body area display not lit” is highlighted, and the display of the “on timer operation” cluster with the on timer display is not highlighted. Or you can. As described above, by storing the case data including the relationship between the data in the case database 17 in advance when storing the cases,
It is possible to improve the efficiency of the search in consideration of the relationship between data.

[0097]

As described above, according to the present invention, keywords are extracted from at least one or more case sentences, the cluster to which each case sentence is determined is determined, and a set of each keyword is determined based on the determination result. The keyword clustering means for generating the problem solving tree to be characterized and the cluster to which each case sentence belongs are subdivided based on the attribute of the linguistic expression of each case sentence belonging to each cluster generated by the problem solving tree generated by the keyword clustering means. And a detailed clustering means for optimizing, it is possible to efficiently classify cases and construct a problem solving tree, and to search for cases including sentences syntactically and semantically similar to the search sentence. There is an effect that can be. Further, after the keyword clustering means determines the cluster to which each case sentence belongs, the detailed clustering means subdivides the cluster to which each case sentence belongs.

According to the present invention, a keyword is extracted from at least one or more case sentences, a cluster to which each case sentence is determined is determined, and a problem solving tree characterized by a set of each keyword is determined based on the determination result. When generated, the cluster to which each case sentence belongs is subdivided based on the attribute of the linguistic expression of each case sentence belonging to each cluster of the problem solving tree. As well as search statements and syntactic
There is an effect that a case including a sentence that is semantically similar can be searched. Further, after determining the cluster to which each case sentence belongs, the cluster to which each case sentence belongs is subdivided, so that there is an effect that the search can be efficiently performed at the time of search.

According to the present invention, when extracting a keyword from a case sentence, the case sentence is extracted from the digitized document.
Since the configuration is such that a tag indicating the type of the case sentence is added, there is an effect that only case sentences of the same type to which a specific tag has been added can be set as a processing target of collation.

According to the present invention, when the problem solving tree is generated, the number of hierarchical levels of the cluster is specified.
There is an effect that the balance between the matching accuracy and the processing speed can be appropriately changed.

According to the present invention, when subdividing the cluster to which each case sentence belongs, the minimum similarity between the case sentences constituting the cluster is specified, so that the processing time is allowed within the allowable range. Within this, there is an effect that a desired collation accuracy can be set.

According to the present invention, when a keyword is extracted from a case sentence, a word whose appearance frequency is higher than a set value among words in the case sentence is extracted as a keyword. There is an effect that important things can be selected and used.

According to the present invention, when a keyword is extracted from a case sentence, an ontology that has knowledge about a word that depends on a target area or a task and a relationship between the words is referred to, and among words in the case sentence, Since the keywords held in the ontology are extracted as keywords, even words with low frequency of appearance can be set as keywords with high importance, realizing fine-grained matching processing There is an effect that can be.

According to the present invention, when subdividing the cluster to which each case sentence belongs, the ontology that holds knowledge about the relationship between words and words that depend on the target area or task and each word is referred to, and the ontology that is held in that ontology is referred to. Since the knowledge about the relation between the words and the words is used for the detailed clustering, there is an effect that a fine-grained matching process can be realized.

According to the present invention, since the ontology has the IS-A relation knowledge indicating the semantic superior-subordinate relation, it is possible to identify the semantic superior-subordinate relation between the linguistic expressions. There is an effect that can be done.

According to the present invention, since the ontology is configured to have the HAS-A relation knowledge indicating the semantic part-whole relation, it is possible to identify the semantic part-whole relation between the linguistic expressions. There is an effect that can be done.

According to the present invention, since the ontology is configured to hold the case relation knowledge, there is an effect that the case relation between sentences can be identified.

According to the present invention, since the ontology is configured to hold paraphrase knowledge, there is an effect that sentences having the same meaning and different expressions can be treated as the same sentence.

According to the present invention, since the ontology is configured to hold the conflicting knowledge, there is an effect that the conflicting relation between sentences can be identified.

According to the present invention, when similar sentence matching between a case sentence belonging to a cluster and a search sentence is executed, matching is performed on a semantic structure based on the attribute of a syntactic element. This has the effect of identifying the difference between the two and improving the matching accuracy.

According to the present invention, when performing similar sentence matching between a case sentence and a search sentence belonging to a cluster, a configuration is adopted in which the degree of detail of matching is specified, so that a desired matching accuracy can be set. There is.

According to the present invention, the detailed level of the collation is
Since the configuration is such that the depth of the tree structure in the parse tree is specified, there is an effect that a desired collation accuracy can be set with emphasis on syntactically important elements.

According to the present invention, since the similarity between the clusters is described in the database, there is an effect that the similarity between the clusters can be specified.

According to the present invention, since the reciprocal relationship between the clusters is described in the database, the reciprocal relationship between the clusters can be specified.

According to the present invention, a keyword is extracted from at least one or more case sentences, a cluster to which each case sentence is determined is determined, and a problem solving tree characterized by a set of each keyword is determined based on the determination result. Detailed clustering processing for subdividing the cluster to which each case sentence belongs based on the keyword clustering processing procedure to be generated and the attribute of the linguistic expression of each case sentence belonging to each cluster in the problem solving tree generated by the keyword clustering processing procedure Since it is configured to describe procedures, it is possible to efficiently classify cases and build problem solving trees, and to search for cases that include sentences that are syntactically and semantically similar to search sentences. effective. Further, after the keyword clustering processing procedure determines the cluster to which each case sentence belongs, the detailed clustering processing procedure subdivides the cluster to which each case sentence belongs.

[Brief description of the drawings]

FIG. 1 is a configuration diagram illustrating a language case inference device according to Embodiment 1 of the present invention;

FIG. 2 is a flowchart illustrating a language case inference method (case construction processing) according to Embodiment 1 of the present invention;

FIG. 3 is a flowchart showing specific contents of a sentence extraction process (step ST1).

FIG. 4 is a sentence extraction process in FIG. 3 (step ST1);
It is a flowchart which shows the specific content of 4).

FIG. 5 is an atypical field process (step ST26).
3 is a flowchart showing the specific contents of FIG.

FIG. 6 is an explanatory diagram showing an example of the digitized document 11;

FIG. 7 is an explanatory diagram showing “cases” generated from the digitized document 11 shown in FIG. 6 by a sentence extraction step ST1.

FIG. 8 is an explanatory diagram showing document field information and the like.

FIG. 9 is an explanatory diagram showing a sentence tag list.

FIG. 10 is a flowchart showing specific contents of a sentence analysis process (step ST2).

FIG. 11 is a flowchart showing specific contents of a morphological analysis process (step ST41).

FIG. 12 is a flowchart showing specific contents of a syntax analysis process (step ST42).

FIG. 13 is a flowchart showing specific contents of a similar case classification process (step ST3).

FIG. 14 is a flowchart showing specific contents of a keyword clustering process (step ST71).

FIG. 15 is a flowchart showing specific contents of a recursive function calling process of FIG.

FIG. 16 is an explanatory diagram showing an example of a hierarchical structure solution tree generated by a keyword clustering process (step ST71).

FIG. 17 is an explanatory diagram showing an example of case sentence data stored in the case database 17;

FIG. 18 shows a detailed clustering process (step ST7).
It is a flowchart which shows the specific content of 2).

FIG. 19 is an explanatory diagram showing an example of a clause structure.

FIG. 20 shows a detailed clustering process (step ST7).
FIG. 4 is an explanatory diagram showing a problem solving tree created by classification in 2).

FIG. 21 is a flowchart illustrating a language case inference method (case search process) according to Embodiment 1 of the present invention;

FIG. 22 is a similar case search process (step ST113).
3 is a flowchart showing the specific contents of FIG.

FIG. 23 is an explanatory diagram showing knowledge held by the ontology 18.

FIG. 24 is an explanatory diagram showing a structure of a case sentence and the like.

FIG. 25 is an explanatory diagram showing an example in which inter-cluster information other than a hierarchical relationship is stored in the case database 17 in step ST93 of the similar case classification process (step ST3).

FIG. 26 is an explanatory diagram showing an example of a penalty calculation rule in similarity calculation.

FIG. 27 is an explanatory diagram showing an example of a problem solving tree created by performing similar case classification.

FIG. 28 is a configuration diagram showing a conventional language case inference apparatus.

[Explanation of symbols]

11 digitized document, 12 sentence extraction unit, 13 sentence analysis unit,
14 similar case classification unit, 15 keyword clustering unit (keyword clustering means), 16 detailed clustering unit (detailed clustering means), 17 case databases, 18 ontology, 19 case cluster editing unit, 20 search sentences, 21 search sentence input unit, 22 Similar case search section (similar case search means), 23 search result display section, 24 word dictionary for analysis, 25 attached word dictionary, 26
Annex connection table, 27 grammar rules, 31 IS-A knowledge, 32 HAS-A knowledge, 33 case relation knowledge, 34
Paraphrase knowledge, 35 rebellious information.

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification mark FI Theme coat ゛ (Reference) G06F 15/403 330C 350C (72) Inventor Katsushi Suzuki 2-3-2 Marunouchi 2-chome, Chiyoda-ku, Tokyo Mitsubishi Electric F term in reference (reference) 5B075 ND03 NK32 NR12 PP24 PR04 PR06 QM08

Claims (19)

[Claims] 1. Keyword clustering for extracting a keyword from at least one or more case sentences, determining a cluster to which each case sentence belongs, and generating a problem solving tree characterized by a set of each keyword based on the determination result. Means, detailed clustering means for subdividing the cluster to which each case sentence belongs based on the attribute of the linguistic expression of each case sentence belonging to each cluster generated by the keyword clustering means, and keywords from the search sentence Language case inference apparatus, comprising: extracting a cluster having the keyword as an element from the problem solving tree; and performing similar sentence matching on the case sentences belonging to the cluster and the similar sentence of the search sentence. . 2. Extracting a keyword from at least one or more case sentences, determining a cluster to which each case sentence belongs, and generating a problem solving tree characterized by a set of keywords based on the determination result. Based on the attribute of the linguistic expression of each case sentence belonging to each cluster of the problem solving tree, the cluster to which each case sentence belongs is subdivided, and a keyword is extracted from the search sentence, and the cluster having the keyword as an element is described above. A language case inference method that searches from a problem solving tree and performs similar sentence matching between case sentences belonging to the cluster and the search sentence. 3. The language case inference method according to claim 2, wherein, when extracting a keyword from the case sentence, the case sentence is extracted from the digitized document and a tag indicating a type of the case sentence is added. 4. The language case inference method according to claim 2, wherein when generating the problem solving tree, the number of hierarchical levels of the cluster is specified. 5. The language case inference method according to claim 2, wherein at the time of subdividing the cluster to which each case sentence belongs, a minimum similarity between the case sentences constituting the cluster is specified. 6. The language case inference method according to claim 2, wherein, when extracting a keyword from the case sentence, a word having an appearance frequency higher than a set value among words in the case sentence is extracted as a keyword. 7. When extracting a keyword from a case sentence, referencing an ontology that holds knowledge of a relationship between each word and a word that depends on a target area or a task, and among the words in the case sentence, holding the ontology in the ontology 3. The language case inference method according to claim 2, wherein the extracted words are extracted as keywords. 8. When subdividing the cluster to which each case sentence belongs, refer to an ontology that has knowledge about the relationship between each word and a word that depends on the target area or task, and refer to the ontology that holds the ontology. 3. The language case inference method according to claim 2, wherein knowledge about relations between words is used for detailed clustering. 9. The language case inference method according to claim 7, wherein the ontology has IS-A relation knowledge indicating a semantic upper-lower relation. 10. The language case inference method according to claim 7, wherein the ontology has HAS-A relation knowledge indicating a semantic part-whole relation. 11. The language case inference method according to claim 7, wherein the ontology has case-related knowledge. 12. The language case inference method according to claim 7, wherein the ontology has paraphrase knowledge. 13. The language case inference method according to claim 7, wherein the ontology holds conflicting knowledge. 14. The language case inference according to claim 2, wherein when performing similar sentence matching between the case sentence belonging to the cluster and the search sentence, matching of the semantic structure is executed based on the attribute of the syntactic element. Method. 15. The language case inference method according to claim 2, wherein when performing similar sentence matching between the case sentence belonging to the cluster and the search sentence, the degree of detail of the matching is specified. 16. The method according to claim 1, wherein the depth of the tree structure in the parse tree is specified as the level of detail of the collation.
5. The language case inference method according to 5. 17. The method according to claim 2, wherein the similarity between the clusters is described in a database.
The language case inference method according to any one of the above. 18. The method according to claim 2, wherein the conflicting relation between each cluster is described in a database.
The language case inference method according to any one of the above. 19. A keyword is extracted from at least one or more case sentences, a cluster to which each case sentence belongs is determined,
A keyword clustering processing procedure for generating a problem solving tree characterized by a set of keywords based on the determination result, and a linguistic expression of each case sentence belonging to each cluster generated by the above-described keyword clustering processing procedure A detailed clustering processing procedure for subdividing the cluster to which each case sentence belongs based on the attribute, and extracting a keyword from the search sentence, searching for a cluster having the keyword as an element from the problem solving tree, and belonging to the cluster A storage medium in which a language case inference program including a case sentence and a similar case search processing procedure for executing similar sentence matching of the search sentence is described.