WO2010125707A1 - Search system, and medium containing a search program - Google Patents

Abstract

Disclosed is a search system provided with: a target data storage unit which stores target data; a dimension-separated matching table storage unit which stores dimension-separated matching tables having different weights depending on dimensions, which comprise combinations of conditions; a base matching table storage unit that stores a base matching table, in which weights for when there are no conditions are set; and a dimension priority storage unit. When conditions are set from a search value, dimension-separated matching tables that match said conditions are extracted from the dimension-separated matching table storage unit; a multidimensional matching table is then generated by overwriting the base matching table with the contents of each dimension-separated matching table, in ascending priority order according to dimension priorities, and the multi-dimensional matching table is used to perform matching.

Description

Search system and search program storage medium

The present invention relates to a search system and a search program storage medium. For example, by repeatedly asking questions in a dialog, the user's needs and values are extracted, and a variety of services and contents are included. Therefore, the present invention can be applied to an information search system that searches for items that match user needs and values.

Communication through IT (information technology) will change from text information to moving images and conversations stored by individuals, broadcast images and movie images, product information and sensor information, production history and traffic history, Expanded to various information media such as climate information, and its usage phase is dynamic to various aspects such as information home appliances, in-vehicle terminals, electronic tags, public transportation, corporate production / distribution sites, market stores, etc. It is expected to expand.

The next generation of information retrieval and analysis technology that responds to new communication needs in every aspect of personal life, business, etc., is not the only text-based information retrieval on the homepage of the current personal computer. It is.

Conventional information retrieval / analysis technology generally accepts keyword input, searches information on the network for information that matches the input keyword, and introduces them in descending order of the number of hits. In addition, there are a keyword automatic expansion technology for displaying keywords used together with input keywords, a technology for recommending products and the like through personal communication by many users, and the like.

However, according to the technique as described above, the user can search a lot of information including the input keyword, but these information statistically summarizes the results searched by many other users. Information, not the real needs or desired personalized information of the user.

Also, in an era of information overload, there is an increasing number of people who are unable to cope with the explosive increase in the amount of information and are easily losing information selection criteria. There is a need for information retrieval / analysis technology that allows the user who has the problem of not being able to decide the information and behavior that he / she wants to search together with the system, not just for searching, but also for thinking.

On the other hand, the inventor of the present application makes use of a laddering technique that draws out the needs and values of the user by repeating questions that are gradually delved into the dialogue, and the user and the system repeat the dialogue. Proposed information retrieval / analysis technology that retrieves the needs of services and contents that were difficult for users to express, and searches for a wide variety of services and contents that match them. (JP 2009-193532, JP 2009-193533).

In this interactive search system, a question sentence is generated based on domain knowledge and presented to the user, the answer from the user is interpreted at a semantic level by intention analysis, information extracted from the user and the search target in advance A search is realized by matching an analysis (for example, an analysis result to be interpreted from a user's answer) and a result of analyzing the search target data.

By the way, in the interactive search system, the following points are important in order to draw out the true needs and desires of users by repeatedly asking the questions gradually.
(1) support the decision-making process;
(2) Encouragement of “awareness” that cannot be fully verbalized by the user alone,
(3) Accurately recognize information selection criteria.

In other words, in the interactive search system, personalization according to each user is performed dynamically, the response that gives a close feeling to the user's remarks, the flow of dialogue, and various situation changes Since the importance of matters to be asked changes every moment, it is required to present a question sentence and perform matching according to this.

The following document describes the significance of ontology.
Riichiro Mizoguchi and two others, “Ontology Engineering Fundamentals: Ontological Semantics of Semantic Links, Classes, Relationships, and Roles”, Journal of the Japanese Society for Artificial Intelligence, Vol. 14, pp. 1019-1032, 1999 (Reference 1)
Ontology defines and systematizes knowledge about a specific domain, knowledge across multiple domains, or general-purpose knowledge. Here, the term “knowledge” refers to a term / vocabulary and its meaning, and the relationship between other terms / vocabulary, and the domain refers to a field such as medicine, engineering, real estate, automobile, and finance.

Ontology can systematically express the concept of words and the relationship between words, so in an interactive search system, it is possible to use what is described in the ontology for each domain as domain knowledge. .

This makes it possible to present a question sentence that uses an ontology, and to associate an answer to this question sentence, so that the user's answer (input) can be understood, and the previous conversation model Simplification.

However, current ontologies can only express the concept of words and the relationship between words from a single point of view when expressing the concept of words and the relationship between words.

Therefore, there is a problem that information selection criteria are accurately analyzed, the selection criteria are switched according to the user's intention and situation, and the search cannot be performed correctly using the switched selection criteria.

For example, if a person who searches for information is a human, information from the user is ignored and no matching condition is presented. For example, it is assumed that there is a qualification / skill that the job seeker requests is “3 years or more of Java® experience”. In this case, even if the user has the possession qualification regarding the program language “Java (registered trademark)”, the job offer is not presented unless the years of experience satisfy the condition of three years or more.

This is because humans correctly analyze the user's intention and switch the selection criteria according to the user's intention and situation. However, the conventional interactive search system cannot dynamically switch the selection criteria according to the user's intention and situation.

The present invention uses the concept of multidimensional ontology in view of the above problems. When the target data is searched from the user's answer, a matching table is required in which the search key of the target data corresponding to the user's answer and the degree of correspondence are entered. This matching table is expressed as a network indicating the relationship between words, and can be expressed by an ontology. Therefore, by considering this matching table as an ontology and dynamically switching the matching table according to the user's intention and situation, the user's intention is accurately analyzed from the user's remarks, and according to the user's intention and situation. A search system and a search program capable of dynamically switching conditions and selection criteria necessary for a search are provided.

In order to solve such a problem, the search system according to the first aspect of the present invention uses a matching table having a correspondence relationship between a value to be searched for and a value of target data corresponding to this value and a weight for the correspondence. In a search system for searching for target data corresponding to a value to be searched from target data of interest,
A target data storage unit for storing one or more target data to be searched;
A dimension-specific matching table storage unit that stores one or more dimension-specific matching tables having different weights depending on the dimension of the combination of conditions;
A basic matching table storage unit that stores a basic matching table in which the above weights are set when there is no condition;
A priority storage unit by dimension for storing priority by dimension indicating the priority for each dimension;
When a value to be searched is set as a condition, a dimension-specific matching table extraction unit that extracts each dimension-specific matching table from the dimension-specific matching table storage unit,
Of the extracted matching tables for each dimension, the values entered by overwriting the basic matching table with the contents of each dimension matching table in descending order of dimension priority according to the dimension priority in the dimension priority storage unit. A multidimensional matching table generation unit for generating a multidimensional matching table for searching for target data from
And a matching unit that searches for one or a plurality of target data that matches a value to be searched from the target data storage unit using the multi-dimensional matching table generated by the multi-dimensional matching table generation unit.

The search program of the second aspect of the present invention uses a matching table having a correspondence relationship between a value to be searched for and a value of target data corresponding to this value and a weight of the correspondence, from among target data to be searched. In a medium storing a search program for searching for target data corresponding to a value to be searched,
The search program is
A target data storage unit for storing one or more target data to be searched;
A dimension-specific matching table storage unit that stores one or a plurality of dimension-specific matching tables having different weights depending on the dimension of a combination of conditions;
A basic matching table storage unit that stores a basic matching table in which the above weights are set when there is no condition;
A priority storage unit by dimension for storing priority by dimension indicating the priority for each dimension, and
On the computer,
When a value to be searched is set as a condition, a dimension-specific matching table retrieval process for retrieving each dimension-specific matching table that meets the condition from the dimension-specific matching table storage unit,
Of the extracted matching tables for each dimension, the contents of the matching table for each dimension are overwritten on the basic matching table in order from the lowest dimension priority according to the dimension priority in the dimension priority storage unit. Multi-dimensional matching table generation processing for generating a multi-dimensional matching table for searching for target data from input values,
Using the multidimensional matching table generated by the multidimensional matching table generation unit to perform a matching process for searching for one or more target data that matches a value to be searched from the target data storage unit .

According to the present invention, by using a multidimensional ontology, the user's intention is accurately analyzed from the user's remarks, and the conditions and selection criteria necessary for the search are dynamically switched according to the user's intention and situation. be able to.

1 is an overall configuration diagram illustrating an overall configuration of an interactive search system according to a first embodiment. It is explanatory drawing explaining the basic concept of a multidimensional ontology (the 1). It is explanatory drawing explaining the basic concept of a multidimensional ontology (the 2). It is a flowchart which shows the whole flow of the dialog control system of 1st Embodiment. It is explanatory drawing explaining a property definition (table format). It is explanatory drawing explaining a property definition (network format). It is a flowchart which shows the multidimensional ontology creation process of 1st Embodiment. It is explanatory drawing explaining the user data which the domain knowledge manager acquired. It is explanatory drawing explaining the dimension list | wrist which shows the priority of a dimension. It is explanatory drawing explaining a dimension priority definition table. It is explanatory drawing explaining the production | generation process of a multidimensional ontology. It is a flowchart which shows the dialogue control process linked with the search process of 1st Embodiment. It is a flowchart which shows the matching process using the multidimensional ontology of 1st Embodiment. It is explanatory drawing explaining the user data which the domain knowledge manager acquired. It is explanatory drawing explaining the matching table of a base dimension. It is explanatory drawing explaining the matching table of a "/ person / personality / current occupation = cook" dimension. It is explanatory drawing explaining the matching table made multi-dimensional. It is explanatory drawing explaining the user data which the domain knowledge manager acquired. It is explanatory drawing explaining the matching table of a dimension of "/ person / personality / current occupation = SE". It is explanatory drawing explaining the matching table made multi-dimensional. It is explanatory drawing explaining the object data and similarity which concern on a matching process. It is explanatory drawing explaining the object data and similarity which concern on a matching process. It is a flowchart which shows the process in the intention analysis part which concerns on the multidimensional intention analysis process (dynamic switching process of a domain knowledge dictionary) of 1st Embodiment. It is a flowchart which shows the process in the domain knowledge management part which concerns on the multidimensional intention analysis process of 1st Embodiment. It is explanatory drawing explaining the intention analysis result in a multidimensional intention analysis process. It is explanatory drawing explaining the user data which the domain knowledge dictionary manager acquired. It is explanatory drawing explaining the domain knowledge dictionary multi-dimensionalized by the domain knowledge dictionary manager. It is explanatory drawing explaining the domain knowledge dictionary of each dimension. It is explanatory drawing explaining the intention analysis result by a multidimensional intention analysis process. It is explanatory drawing explaining the object data which the intention analysis part acquired in the multidimensional intention analysis process (dynamic switching process of an analysis object sentence) of 1st Embodiment. It is explanatory drawing explaining the domain knowledge dictionary multi-dimensionalized by the domain knowledge dictionary manager. It is explanatory drawing explaining the multidimensional syntax dictionary of each dimension. It is explanatory drawing explaining the intention analysis result by a multidimensional intention analysis process. It is a flowchart which shows the process of the dialog control part in the dialog control process linked with the matching result of 1st Embodiment. It is a flowchart which shows the matching process in the dialogue control process linked with the matching result of 1st Embodiment. It is explanatory drawing explaining the user data which the dialogue control part acquired. It is explanatory drawing explaining the object data which the matching part acquired. It is explanatory drawing explaining the matching table of a dimension of "/ person / present occupation = programmer". It is explanatory drawing explaining the matching with user data and object data. It is explanatory drawing explaining the user data which the dialogue control part acquired. It is explanatory drawing explaining the matching with user data and object data.

(A) First Embodiment Hereinafter, a first embodiment of a search system and a search program of the present invention will be described in detail with reference to the drawings.

In the first embodiment, an embodiment in which the present invention is applied to an interactive search system using laddering that repeatedly asks the user questions and derives the user's true needs and desires will be described as an example.

(A-1) Basic Concept of Multidimensional Ontology Since the matching table used in the first embodiment can be expressed by a network structure representing the relationship between words, it can be regarded as an ontology. Therefore, in the first embodiment, the concept of multidimensional ontology is applied to the matching table.

First, the concept of this multidimensional ontology will be described with reference to the drawings.

The multi-dimensional ontology refers to an ontology that holds a large number of semantic spaces in a multi-dimensional manner and can instantly switch the semantic space to be used according to user data or target data / situations obtained through dialogue.

As described above, the current ontology only expresses the concept of words and the relationship between words from a certain point of view. Therefore, even if the current ontology is used for an interactive search system, static data Can only be obtained.

Therefore, in the first embodiment, an ontology can be maintained in a multidimensional manner, and a mechanism for automatically moving between dimensions whenever the information extracted from the user changes is proposed.

This makes it possible to instantly switch the semantic space or switch information for dialogue according to user data and situations obtained through dialogue, and realize the dynamic nature of dialogue and search as a system. In addition, by enabling data to be inherited between dimensions, it was possible to achieve descriptiveness and labor saving (minimum description) at the same time. You can describe in detail only what you want to describe, which is the greatest feature of multidimensionalization.

2 and 3 are explanatory diagrams for explaining the basic concept of the multidimensional ontology. 2 and 3 exemplify a case where the interactive search / analysis system 1 of the first embodiment is used for an occupation introduction system.

As shown in FIG. 2 and FIG. 3, in the multidimensional ontology, the dimension in which the property is set is positioned according to the priority, and the base dimension and a special dimension other than the base (hereinafter referred to as a sub dimension). ).

∙ Various ontologies are defined for the base dimension and sub dimension. The base dimension is the dimension with the lowest priority, and defines all classes and properties defined in the sub dimensions. Further, only necessary classes and properties are defined for each sub dimension, and the higher the number of conditions, the higher the priority. When the number of conditions is the same, for example, priorities are assigned in advance by the developer.

For example, FIGS. 2 and 3 illustrate a 6-dimensional ontology including the base dimension. The sub-dimension of “Industry = IT”, the sub-dimension of “Occupation = Sales”, and the sub-dimension of “Occupation = SE” each have one condition and the same number of conditions, but “Industry = IT” Assume that the priority is set in the order of the sub-dimension, the “job type = sales” sub-dimension, and the “job type = SE” sub-dimension.

Also, a sub-dimension of “job type = sales && industry = IT (this notation indicates that the job type is sales and the industry is IT, and so on)”, “job type = SE && industry = IT” Each sub-dimension has two conditions and the same number of conditions. However, the sub-dimension of “Occupation = Sales && Industry = IT” and the priority set in the order of “Occupation = SE && Industry = IT”. To do.

Therefore, when these dimensions are overlapped, the “base dimension” has the lowest priority, so the priority of the dimension is “base dimension” → “industry = IT” sub dimension → “job type = sales” sub dimension. The dimension becomes higher in the order of sub-dimension of “job type = SE” → sub-dimension of “job type = sales && industry” → “job type = SE && industry = IT” sub-dimension.

In FIGS. 2 and 3, circles on each sub dimension indicate classes and properties defined in each sub dimension.

FIG. 2 shows “when nothing (data) is acquired” from the user. In this case, the dialogue has not yet progressed. At this time, the dialogue proceeds using classes and properties (not shown in FIG. 2) defined in the base dimension.

Suppose that the dialogue has progressed and the user has obtained the condition “industry = IT”. In this case, as shown in FIG. 3, the applied dimension is a sub-dimension of “industry = IT”, and the conversation is switched to a dialogue using classes and properties defined in the sub-dimension of “industry = IT”.

Suppose further that the dialogue has progressed and the user has obtained a condition of “occupation = sales and industry = IT”. In this case, as shown in FIG. 3, the applied dimension is a sub-dimension of “Occupation = Sales && Industry = IT”, and the classes and properties defined in the sub-dimension of “Occupation = Sales && Industry = IT”. Switch to dialogue using.

Here, the dimensions applied to the dialogue are two sub-dimensions of “industry = IT” and “occupation = sales and industry = IT”, and the class and property of each applied subdimension is the base dimension. It is overwritten by and inherited. At this time, the sub-dimensions with higher priorities are stored in order from the lower-priority sub-dimensions so that the sub-dimensional classes and properties are stored. That is, the sub-dimension class and property of “industry = IT” are stored in the base dimension, and then the sub-dimension class and property of “occupation = sales and industry = IT” are overwritten.

By using a multidimensional ontology in this way, data can be inherited between dimensions, and values such as correspondence and importance of target data and user data can also be inherited between dimensions in the matching table. It becomes.

(A-2) Configuration of First Embodiment Hereinafter, the overall configuration of the interactive search system of the first embodiment will be described with reference to FIG.

In FIG. 1, the interactive search system 1 according to the first embodiment includes at least a dialogue management unit 10, an intention analysis unit 20, a domain knowledge management unit 30, a matching unit 40, and a matching target analysis unit 50. .

The interactive search system 1 is connected to a data providing server 80 that provides data to the user, acquires the target data held by the data providing server 80, and stores it in the expansion target data 54 of the matching target analyzing unit 50.

The user starts the interactive search system 1 by accessing the Web server 70 using the browser 90. The user interface can be widely applied to a PC having a communication function, a portable terminal, a dedicated terminal, or the like, but may be provided with a voice synthesizing / recognizing unit that converts a voice uttered by the user into a text. Good.

The dialog management unit 10 manages the progress of a dialog with a user who wants to search, and includes a dialog control unit 11 that controls the progress of the dialog.

The dialogue control unit 11 receives a multidimensional ontology from the domain knowledge management unit 30 and creates a question sentence (also referred to as a system utterance sentence) related to the progress of the dialogue based on the multidimensional ontology.

Further, the dialogue control unit 11 gives an answer sentence (also referred to as a user utterance sentence) from the user to the question sentence to the intention analysis unit 20 to analyze the contents of the user utterance sentence. When the dialogue control unit 11 receives the analysis result of the user utterance, the dialogue control unit 11 stores the analysis result of the user utterance as user data in the extended user data 53 of the matching target analysis unit 50.

Further, upon receiving the intention analysis result from the intention analysis unit 20, the dialogue control unit 11 gives the intention analysis result to the domain knowledge management unit 30 to generate a multidimensional ontology used for the progress of the dialogue, and the domain knowledge management unit Using the multidimensional ontology from 30, the next question sentence is determined and the dialogue proceeds.

Further, the dialogue control unit 11 instructs the matching unit 40 to match user data and target data. Various timings for instructing the matching are set, for example, when a matching instruction is received from the user, when an answer sentence for each question sentence is obtained, or when all question items are completed. be able to.

The intention analysis unit 20 receives a user utterance sentence answered by the user in response to the question sentence acquired from the dialogue control unit 11, and analyzes the contents of the user utterance sentence. In addition, the intention analysis unit 20 gives the analyzed intention analysis result to the dialogue control unit 11 in an ontology format. Furthermore, the intention analysis unit 20 performs dynamic switching processing of domain knowledge dictionary and dynamic switching processing based on analysis target sentence information as multidimensional intention analysis processing.

The intention analysis unit 20 includes an intention analysis execution unit 21, an intention analysis dictionary 22, and a domain knowledge dictionary manager 23.

The intention analysis execution unit 21 refers to the intention analysis dictionary 22 and performs a morpheme analysis unit 211 that performs morphological analysis on the user utterance sentence, and refers to the intention analysis dictionary 22 to construct a syntax for the user utterance sentence. It has a syntax analysis unit 212 that performs analysis.

The intention analysis execution unit 21 acquires a class defined in the property definition of the domain knowledge DB 32 by performing morphological analysis and syntax analysis on the user utterance sentence, and gives the acquired class to the domain knowledge management unit 30.

The intention analysis dictionary 22 is a dictionary group for analyzing the contents of a user utterance sentence. For example, a morpheme dictionary (for example, a Japanese morpheme dictionary), a syntax dictionary (for example, a Japanese syntax dictionary), a base dimension, and each dimension This includes domain knowledge dictionaries that are automatically generated from dimension class definitions.

The domain knowledge dictionary manager 23 receives the extended user data from the user data management, rearranges the priorities of the dimension-specific domain dictionaries to be looked up according to the extended user data, This is set in the analysis dictionary 22.

Further, the domain knowledge dictionary manager 23 receives the analysis result of the target data (target sentence information) from the data providing server 80 from the intention analysis execution unit 21, and creates a dictionary for each dimension according to the contents of the target data. Sorting is performed, and the sorted syntax dictionary by dimension is set in the intention analysis dictionary 22.

The domain knowledge dictionary manager 23 includes a feature extraction unit 231, a condition rearrangement unit 232, and a dictionary setting unit 233.

The feature extraction unit 231 extracts a value of a dimension condition when a value that is a dimension condition is included from the intention analysis result. The condition rearrangement unit 232 rearranges the dimension-specific domain dictionary or the dimension-specific syntax dictionary according to a predetermined priority order with respect to the dimension condition extracted by the feature extraction unit 231. Further, the dictionary setting unit 233 sets the domain domain dictionary or the dimension syntax dictionary sorted by the condition sorting unit 232 in the intention analysis dictionary 22.

The domain knowledge management unit 30 manages domain knowledge with knowledge of multidimensional ontology for each domain. The domain knowledge management unit 30 includes a domain knowledge manager 31 and a domain knowledge DB 32.

In the domain knowledge DB 32, the base dimension for constructing the multidimensional ontology and the property definition of each dimension, the class definition and the inference definition, the dimensional condition for constructing the multidimensional ontology, and the priority of the dimensional condition are set. A dimension priority definition table to be defined, a generated multidimensional ontology, and a matching table are stored.

The domain knowledge manager 31 receives the intention analysis result from the dialogue control unit 11, extracts features from the intention analysis result, and generates a multidimensional ontology while referring to the property definition and dimension priority definition table of the domain knowledge DB 32. Is. The domain knowledge manager 31 stores the generated multidimensional ontology in the domain knowledge DB 32.

The domain knowledge manager 31 regards the matching table as a kind of ontology, and generates a matching table using the multidimensional ontology (hereinafter also referred to as a multidimensional matching table). The matching unit 40 performs matching processing while referring to the multidimensional matching table.

The domain knowledge manager 31 of the domain knowledge management unit 30 includes a feature extraction unit 311, a condition matching unit 312, and a multidimensional ontology generation unit 313.

The feature extraction unit 311 confirms whether or not a feature (a value corresponding to the dimension condition) as a dimension condition is set in the intention analysis result obtained by analyzing the contents of the user utterance sentence. The feature is extracted.

Here, the user data DB 53 managed by the user data management unit 51 stores user information obtained from the user utterance by the dialogue control unit 11. This user data DB 53 holds a user data path indicating the storage location (storage location) of user information and a user data class name indicating the value of user information.

The condition collation unit 312 collates the dimension value extracted by the feature extraction unit 311 with the contents of the dimension priority definition table.

The multi-dimensional ontology generation unit 313 generates a multi-dimensional ontology or a multi-dimensional matching table by superimposing the dimensional ontology on the basic ontology according to the priority of the dimension priority definition table.

The matching unit 40 performs matching between user data stored in the user data management unit 51 and target data stored in the target data management unit 50, and matches conditions desired by the user.

The matching unit 40 gives the multi-dimensional matching table received from the dialogue control unit 11 to the matcher 42, or gives the matched information to the dialogue control unit 11 to the match control unit 11, user data and target data And a matcher 42 for performing the matching process.

The matching target analysis unit 50 stores user data and target data to be matched, expands to a form that can be easily matched, and stores the expanded user data and target data. The matching target analysis unit 50 includes a user data management unit 51, a target data management unit 52, extended user data 53, and expansion target data 54.

(A-3) Operation of the First Embodiment Next, the operation of the intention analysis system in the interactive search system 1 of the first embodiment will be described with reference to the drawings.

(A-3-1) Overall Operation First, the overall flow in the interactive search system 1 will be described with reference to the drawings.

FIG. 4 is a flowchart showing the overall flow of the interactive search system 1.

The user who uses the interactive search system 1 accesses the designated URL using the browser 90 and activates the interactive search system 1 through the Web server 70.

When the interactive search system 1 is activated, the dialog control unit 11 refers to the property definition stored in the domain knowledge DB 32 through the domain knowledge manager 31 and has the highest priority among the properties having the current pointer in the domain. Select.

Here, FIG. 5 is a configuration diagram showing the configuration of the property definition, and FIG. 6 is a diagram showing a part of the property definition in a network format.

The property definition shown in FIG. 5 is an example of a property definition serving as a multidimensional base. For example, the property relationship of each class is described with “human” as a vertex. In addition, it describes option information related to dialogue strategies and system utterances that each property has, and defines what questions the system asks in what order.

In FIG. 5, the property definition has items of a definition area, a property, a value area, and option information. In addition, the option information is provided with an importance indicating the degree of determining whether the question is an essential question and a priority indicating the degree of determining the order of questions (that is, the flow of dialogue).

Furthermore, basic question text generated by the system 1 is defined in the option information. In FIG. 5, the option information defines various functions for the system to smoothly proceed with the conversation, such as a connection sentence generated before the question sentence and various received sentences generated based on the intention analysis result. Yes.

Also, property information is described in option information, for example, group and acquisition target. In addition, although the case where there are two types of properties is illustrated here, the property can be set without being limited to this.

When the property type is “acquisition target”, the domain knowledge manager 31 gives the basic question sentence set in the item of the basic question sentence to the dialog control unit 11, and the dialog control unit 11 uses this as a system utterance sentence. It transmits to the user side through the Web server 70.

On the other hand, when the property type is “group”, the domain knowledge manager 31 searches and selects a property having the range class in the domain. At this time, when there are a plurality of properties having this range class as the domain, the domain knowledge manager 31 looks at the priority of the option information and selects the property with the highest priority. If the selected property is “acquisition target”, the domain knowledge manager 31 gives the basic question sentence to the dialogue control unit 11, and the dialogue control unit 11 transmits this to the user side as a system utterance sentence. .

In this way, when the property type is a group, the priority can be used to achieve the transition to a property having a range class in the domain.

FIG. 6 is a configuration diagram in which a part of the property definition shown in FIG. 5 is expanded in a network format.

In FIG. 6, circles indicate classes (that is, domain and value range in FIG. 5), and lines connecting between classes indicate properties.

For example, in FIG. 6, the range when “human” is the vertex and the property is “personality” is “personality”, and the range when the property is “strength” is “strength that can be used for work”. Yes, the value range when the property is “an opportunity to change jobs” is “reason for job change”, and the value range when the property is “work experience” is “current job”.

Also, for example, when “personality” is defined as a domain, as in the above, for example, the value range when the property is “site visit purpose” has a relationship such as “site visit purpose”. “Personality” and “site visit purpose” have a has part relationship, and “site visit purpose” and user response “want to know suitable job” have an is A relationship.

∙ When the system is started, the definition area to be selected as the initial value is set in the property definition. A pointer is set for the first domain domain class “human” in FIG.

Therefore, the domain knowledge manager 31 sees the range class “personality” in the case of setting “personality” having the highest priority among the properties of the domain class “human” pointed to by the initial pointer.

At this time, since the type of the property of the range class “personality” is “group”, the domain knowledge manager 31 has the domain class “with the highest priority among the properties having the range class“ personality ”as the domain. Select the property of “Personality”-Property “Site visit purpose”-Range class “Site visit purpose”.

Then, the domain knowledge manager 31 gives the basic question message of this property “Please tell us the purpose of coming to this site” to the dialog control unit 11, and the dialog control unit 11 uses this basic question message as a system utterance Send to the user side. In this way, a question from the system is started (step S101).

The user who has received the system utterance sentence replies, for example, “I want to know what his / her job is”, and transmits this as a reply sentence to the dialog control unit 11 through the Web server 70 (step S102).

When the dialog control unit 11 receives the user utterance, the dialogue control unit 11 gives the user utterance to the intention analysis unit 20. In the intention analysis unit 20, the intention analysis execution unit 21 uses the intention analysis dictionary 22 to perform morphological analysis and syntactic analysis on the user utterance sentence, and as a result, the intention of the user “I want to know the right job” is obtained. Analyzing and giving the intention analysis result to the dialogue control unit 11 (step S103).

At this time, the intention analysis result of the intention analysis unit 20 is given to the domain knowledge manager 31, a multidimensional ontology is generated by the domain knowledge manager 31, and the generated multidimensional ontology is stored in the domain knowledge DB 32.

Upon receiving the intention analysis result from the intention analysis unit 20, the dialogue control unit 11 inquires of the domain knowledge manager 31 about the next question and determines the next question (step S104).

At this time, the domain knowledge manager 31 refers to the property definition and shifts to the domain class “personality” —the property “current job type” —the range class “job type”, which has the next highest priority after the range class “site visit”. Then, the basic question sentence “What is your current job type?” Is given to the dialogue control unit 11, and the dialogue control unit 11 determines this as a system utterance and transmits it to the user side (step S105).

Thus, the dialogue between the system and the user is realized by repeatedly presenting the next question sentence and receiving the answer sentence from the user.

The search process is processed in the dialog control process (S104), and when the dialog control unit 11 receives a search request from the user, the search result is displayed (details are described in A-3-4).

Note that if the next question is not determined in step S104, that is, if all the questions are finished, the operation of the system is finished.

(A-3-2) Domain Knowledge Manager Processing Next, details of multidimensional ontology creation processing in the domain knowledge manager 31 will be described with reference to the drawings.

The example here explains the processing of the domain knowledge manager 31 in dialog control, but the basic operation of this operation is the same when processing the matching table as an ontology.

FIG. 7 is a flowchart showing the operation of the multi-dimensional ontology creation process in the domain knowledge manager 31.

The domain knowledge manager 31 receives the intention analysis result from the dialogue control unit 11. Moreover, the domain knowledge manager 31 acquires the user data of the user from the user data management unit 51 (step S201).

For example, the intention analysis result from the dialogue control unit 11 is “/ person / personality / current job = sales”.

Further, it is assumed that the user data of the user from the user data management unit 51 has the contents shown in FIG. As shown in FIG. 8, the user data is described in an ontology format. In the case of FIG. 8, as the user data of the user, the class of the property “nickname” is the class “○ -chan”, and similarly, the property “stage of job change activity” —the class “sent resume” and the property “life event” ”—Class“ restructuring / bankruptcy ”, property“ current industry ”—class“ pharmaceutical ”, property“ current job type ”—class“ sales ”.

For the sake of convenience, the user data indicates a case where the class for the property obtained from the user's answer is used as the data. For example, the user data may be a new class inferred from the user's answer class. The contents may be expanded data.

The feature extraction unit 311 of the domain knowledge manager 31 compares the user data with the conditions of all dimensions registered in the domain knowledge DB 32, and extracts the dimension conditions that meet the conditions (step S202). That is, the feature extraction unit 311 confirms whether or not the user data includes values that serve as conditions for all dimensions, and if there is a value that serves as a condition, extracts a value that serves as a condition for that dimension. .

For example, as the dimension conditions stored in the domain knowledge DB 32, “/ human / personality / current job = sales”, “/ human / personality / current industry = pharmaceutical”, “/ human / personality / life event = Restructuring, Bankruptcy && / Human / Personality / Current Job Type = Sales ”.

The feature extraction unit 311 matches with the above dimensional condition, and “/ person / personality / current job = sales”, “/ person / personality / current industry = pharmaceutical”, “/ person / personality / The characteristics of “life event = restructuring / bankruptcy && / human / personality / current job type = sales” are extracted from user data.

Next, the condition matching unit 312 rearranges the dimension condition values extracted by the feature extraction unit 311 in descending order of predetermined priority (step S203), and checks whether there is a dimension condition value. And stored in the dimension priority definition table (step S204).

At this time, for example, the one with a large number of dimension conditions is given higher priority. When the number of dimension conditions is the same, priorities for each dimension are set in advance, and rearrangement is performed based on the priorities.

For example, among the dimension condition values extracted by the feature extraction unit 311, “/ person / personality / life event = restructuring / bankruptcy && / person / personality / current job = sales” has two conditions. The highest priority.

Next, “/ human / personality / current job = sales” and “/ human / personality / current industry = pharmaceutical” have the same number of dimension conditions.

Therefore, for example, the condition matching unit 312 refers to the dimension list shown in FIG. 9, confirms the priority of the condition values of each dimension, and sorts the condition values of each dimension according to the dimension list.

In the dimension list of FIG. 9, the higher the priority is the one described above. For example, “/ human / personality / life event” has a higher priority than “/ human / personality / current job”. If the paths representing the dimensions are the same, “/ person / personality / current job = sales” has a higher priority than “/ person / personality / current job = SE”.

Therefore, for “/ human / personality / current job = sales” and “/ human / personality / current industry = pharmaceutical”, the condition matching unit 312 indicates that “/ human / personality / current job = sales” is “/ Sort so that the priority is higher than "Human / Personality / Current industry = Pharmaceuticals".

As a result, as shown in FIG. 10, “/ human / personality / life event = restructuring / bankruptcy && / human / personality / current job = sales”, “/ human / personality / current job = sales”, “/ Sorted in the order of “human / personality / current industry = pharmaceutical”.

Next, the multidimensional ontology generation unit 313 generates a multidimensional ontology by superimposing the dimensional ontology on the basic ontology according to the priority of the dimension priority definition table (step S205).

FIG. 11 is an explanatory diagram for explaining multidimensional ontology generation processing. FIG. 11A shows job class and class options defined in the base dimension. FIG. 11B shows the definition of job class and class option in the dimension of “industry class = pharmaceutical”.

The multi-dimensional ontology generation unit 313 performs the occupation class and class in the “business category = pharmaceutical” dimension of FIG. 11B in response to the occupation class and class option defined in the base dimension of FIG. Overwrite option contents.

At this time, in the job category of the base dimension in FIG. 11A, the option acceptance sentence of the class “job type” -class “sales” is defined as “You are doing sales work”. As a result of the overwriting of step S205, the option acceptance sentence of class “job type” -class “sales” is rewritten to “sales business even in a highly competitive industry”.

Thus, according to the multidimensional ontology, the contents of the class option definition defined in the base dimension can be changed. Thereby, for example, an appropriate acceptance sentence can be output according to the job type of the user.

The multidimensional ontology generation unit 313 outputs the generated multidimensional ontology to the dialogue control unit 11 (step S206). The generated multidimensional ontology is stored in the domain knowledge DB 32.

(A-3-3) Dialog Control Process in Interactive Search Process Next, the dialog control process in the interactive search process will be described in detail with reference to the drawings.

As a premise for performing the dialog control process in the interactive search process, the domain knowledge manager 31 performs the multidimensional ontology creation process described with reference to FIG. 7 on the target data acquired from the data providing server 80 to determine the intention of the target data. Analyzed and stored in the expansion target data 54.

FIG. 12 is a flowchart showing the dialog control process in the interactive search process.

First, the intention analysis unit 20 performs intention analysis on the user utterance, and when the intention analysis result is given to the dialogue control unit 11 (step S301), the dialogue control unit 11 converts the intention analysis result into the extended user data 53. (Step S302).

For example, the dialogue control unit 11 stores the intention analysis result “/ human / personality / current job / year of experience = 5 years” in the extended personal data 53.

Next, the matching unit 20 starts the matching process. Specifically, the domain knowledge manager 31 generates a multidimensional matching table, and uses it to search for extension target data from the extended user data, and rearranges the target data in order from the highest search similarity (step S303). .

In this way, through the dialogue with the user, the similarity between all the information obtained from the user and all the target data is calculated and presented to the user in order from the target data having the highest similarity.

At this time, when a matching result display request is received from the user side (step S304), the dialogue control unit 11 receives the matching result from the matching unit 40 and transmits the matching result to the user side (step S305).

On the other hand, when there is no display request from the user side (step S304), the dialogue control unit 11 does not transmit the matching result and proceeds to step S306.

In step S306, the dialogue control unit 11 receives the multidimensional ontology from the domain knowledge manager 31, and determines the next question (property). Then, the dialogue control unit 11 transmits the next question sentence to the user side, and ends the process (step S307).

(A-3-4) Matching Process Using Multidimensional Ontology Next, the detailed operation of the matching process using the multidimensional ontology in step S303 in FIG. 12 will be described with reference to the drawings.

FIG. 13 is a flowchart showing the operation of matching processing using a multidimensional ontology. The domain knowledge manager 31 is processing for generating a multidimensional matching table by regarding the matching table as a kind of ontology.

First, the matching unit 40 requests the domain knowledge manager 31 to create a matching table, and the domain knowledge manager 31 regards the matching table as a kind of ontology and generates a multidimensional matching table (step S401). .

Upon receiving the matching table creation request, the domain knowledge manager 31 acquires all user data from the extended user data 53.

The operation flow of the domain knowledge manager 31 is represented by a flow in which S204 in FIG. 7 does not exist. In the domain knowledge manager 31, the feature extraction unit 311 compares the user data with the conditions of all dimensions registered in all the matching tables, and extracts the dimension conditions that meet the conditions (S202).

Next, in the domain knowledge manager 31, the condition matching unit 312 arranges the dimension conditions extracted by the feature extraction unit 311 in descending order of priority according to a certain standard, and stores them in the dimension priority definition table (S203). ).

Then, the multidimensional ontology generation unit 313 generates a multidimensional matching table by superimposing the dimension-specific matching table on the base dimension matching table according to the priority of the dimension priority definition table (S205). (S206).

For example, FIG. 14 shows user data (network format) of the nickname “Chan” whose current job type is cook. FIG. 15 shows a configuration example of a base dimension matching table.

FIG. 15A shows the base dimension matching table in a table format, and FIG. 15B shows the base dimension matching table in a network format.

15, the base dimension matching table includes “user data”, “importance”, and “target data” as items. The item “user data” indicates a property (condition) in the user data. The item “target data” indicates a property (condition) in the target data. The item “importance” indicates a weight for a property in matching.

In the example of FIG. 15, “user data = current job type” and “target data = experience job type” have a corresponding relationship, and the importance is “0.5”.

The property name is different depending on the relationship between the user data and the target data. This is because the condition name (property name) of the recruitment guidelines (data) of the company recruiting the job change and the condition name extracted from the corresponding user This is because (property name) is different.

FIG. 16 shows a dimension matching table referred to when the feature extraction unit 311 extracts the features.

The matching table in FIG. 16 is a “/ human / personality / current job = cook” dimension matching table. Note that FIG. 16A shows a table format, and FIG. 16B shows a network format.

In the dimension matching table shown in FIG. 16, since “current job = cook”, it is important to have a qualification related to cooking, so the importance of the qualification owned is “1.0”. Highly defined.

Therefore, the multidimensional ontology generation unit 313 overwrites the base dimension matching table shown in FIG. 15 with the dimension matching table shown in FIG. As a result, as shown in FIG. 17, it is possible to generate a multidimensional matching table in which “the importance of qualification” is changed to “1.0”.

As another example, a case will be described in which user data of “Δ-kun” whose current job type is “SE” as shown in FIG. 18 is acquired.

In this case, FIG. 19 shows a dimension matching table referred to when the feature extraction unit 311 extracts the features.

The matching table in FIG. 19 is a “/ human / personality / current job type = SE” matching table. Note that FIG. 19A shows a table format, and FIG. 19B shows a network format.

In the dimension matching table shown in FIG. 19, since “current occupation = cook”, the qualification to own is not emphasized, but since the years of experience are emphasized, the importance of the qualification to possess is “0.1”. It is defined as low and the importance of years of experience is defined as high as “0.8”.

Therefore, the multidimensional ontology generation unit 313 overwrites the base dimension matching table shown in FIG. 15 with the dimension matching table shown in FIG. As a result, as shown in FIG. 20, it is possible to generate a multidimensional matching table in which the qualification importance is changed to “0.1” and the experience years importance is changed to “0.8”.

Next, the matching unit 40 that has received the multidimensional matching table from the domain knowledge manager 31 calculates the degree of similarity with all target data based on the multidimensional matching table (step S402).

At this time, the matching unit 40 calculates the degree of similarity by weighting with the “importance” described in the matching table using the multidimensional matching table.

Then, the matching unit 40 rearranges the target data in descending order of similarity, and uses this as a matching result.

For example, in the example of “Chan” above, the matching unit 40 receives the multidimensional matching table shown in FIG. 17 and is similar to the target data of “Company A” and “Company B” in FIG. A case where the degree is calculated will be described.

The conditions for “Company A” are “Experienced industry = cook”, “Eligible qualification = Cooker license”, and “Year of experience = 1 year or more”, so “Chan” matches all conditions. Therefore, the matching unit 40 calculates the similarity with “Company A” as “(1 × 0.5) + (1 × 1.0) + (1 × 0.5) = 2.0” (FIG. 21 (B)).

Similarly, the conditions for “Company B” are “Experienced industry = cook”, “Holding qualification = Nutrition license”, and “Experience years = 5 years or more”. Matches Accordingly, the matching unit 40 calculates the similarity with “Company B” as “(1 × 0.5) + (0 × 1.0) + (1 × 0.5) = 1.0” (FIG. 21 (B)).

Therefore, the matching unit 40 outputs matching results in the order of “Company A” and “Company B” as shown in FIG.

Further, for example, in the example of “Δ-kun” described above, the matching unit 40 receives the multi-dimensional matching table shown in FIG. 19 and compares it with the target data of “Company C” and “Company D” in FIG. A case where the similarity is calculated will be described.

Since the conditions of “Company C” are “Experience type industry = SE”, “Eligible qualification = Beginner Sysad”, “Number of years of experience = 1 year or more”, “△ kun” matches all conditions. Therefore, the matching unit 40 calculates the similarity with “Company A” as “(1 × 0.5) + (1 × 0.1) + (1 × 0.8) = 1.4” (FIG. 22 (B)).

Similarly, the conditions of “Company D” are “Experienced industry = cock”, “Retention qualification = Advanced sysad”, and “Experience years = 5 years or more”. Therefore, “△ kun” is a condition other than “Experience years”. Matches Accordingly, the matching unit 40 calculates the similarity with “Company D” as “(1 × 0.5) + (1 × 0.1) + (0 × 0.8) = 0.6” (FIG. 22 (B)).

Therefore, the matching unit 40 outputs matching results in the order of “Company C” and “Company D” as shown in FIG.

(A-3-5) Multidimensional intention analysis processing (domain knowledge dictionary dynamic switching processing)
Next, multidimensional intention analysis processing (domain knowledge dictionary dynamic switching processing) will be described in detail with reference to the drawings.

FIG. 23 is a flowchart showing processing of the intention analysis execution unit 21 related to multidimensional intention analysis processing (domain knowledge dictionary dynamic switching processing). FIG. 24 is a flowchart showing processing of the domain knowledge dictionary manager 23 related to multidimensional intention analysis processing (domain knowledge dictionary dynamic switching processing).

First, in FIG. 23, the intention analysis unit 20 receives a user utterance from the dialogue control unit 11 (step S501).

In the intention analysis unit 20, if there is attached information, the attached information is set as user data in the extended user data 53 (step S502).

When the user utterance sentence is composed of a plurality of sentences, the intention analysis execution unit 21 divides the sentence into one sentence unit (step S503), and the morpheme analysis unit 211 performs morpheme analysis on the user utterance sentence (step S504). The unit 212 performs syntax analysis (step S505).

Furthermore, the intention analysis execution unit 21 performs intention analysis with reference to the intention analysis dictionary 22. At this time, the intention analysis execution unit 21 gives the user data to the domain knowledge dictionary manager 23 (step S506).

In FIG. 24, when user data is received from the intention analysis execution unit 21 (step S601), the user domain knowledge dictionary manager 23 matches the dimensional condition with the feature extraction unit 231 collating the user data with the dimensional condition of each dimension. A value is extracted (step S602).

The condition rearrangement unit 232 rearranges the dimensional domain knowledge dictionaries in priority order according to a certain standard with respect to the dimension condition extracted by the feature extraction unit 311 (step S603), and the dictionary setting unit 233 has a predetermined priority. The dimension-specific domain knowledge dictionary is set in the intention analysis dictionary 22 so as to be looked up by rank (step S604).

In step S507 of FIG. 23, the intention analysis unit 20 performs a dictionary lookup from the domain knowledge dictionary according to dimensions in the order set by the domain knowledge dictionary manager 23, and the intention according to the information obtained from the dialogue with the user. Is generated (step S507).

For example, with reference to FIG. 25 to FIG. 29, explanation will be given by way of example of intention analysis in the case where the user utterance about the question about the current address for “Chan” is “Mita is”.

First, before starting the system, a domain knowledge dictionary according to dimensions as shown in FIGS. 28A to 28C is automatically generated according to the class definition of the “address” class.

For example, FIG. 28A is a domain knowledge dictionary classified by dimension in which the “address” class of the “/ human / personality / working address = Osaka” dimension is defined, and the upper hierarchy is the “address” class. The “Osaka” class is under the “address” class, the “Takatsuki City” class, the “Ibaraki City” class, and so on are under the “Osaka” class. Also, “Takatsuki City” class, “Ibaraki City” class, etc. are also shown together.

Similarly, FIG. 28B is a domain knowledge dictionary defining an “address” class of the dimension “/ human / personality / work address = Kansai”, and FIG. 28C shows “/ human / personality”. This is a domain knowledge dictionary that defines an “address” class of the dimension “/ work address = Tokyo”.

When the user data shown in FIG. 26 is given to the domain knowledge dictionary manager 23, in the domain knowledge dictionary manager 32, the feature extraction unit 231 extracts features from the user data and extracts the property “workplace” -class “Osaka”. To do.

The condition rearrangement unit 232 selects a domain knowledge dictionary classified by dimension that satisfies the property “workplace” -class “Osaka” extracted by the feature extraction unit 231 and rearranges the domain knowledge dictionary classified by dimension according to a predetermined priority. I do. By rearranging the dimensional dictionaries, the dictionary can be looked up from the high-priority dimensional domain knowledge dictionary.

At this time, the condition rearranging unit 232 sets the domain knowledge dictionary according to dimension of FIG. 28 (A) in which the feature extracting unit 231 meets the condition of the property “place of work” −class “Osaka” as the highest priority, The domain knowledge dictionary of the “Kansai” class, which is a higher class of the “Osaka” class, is rearranged as the next highest priority.

As a result, the condition rearrangement unit 232, as shown in FIG. 27, from the extraction condition of the feature extraction unit 231, a domain knowledge dictionary of “/ person / personality / work address = Osaka” dimension in FIG. Rearrangement is performed in the order of the domain knowledge dictionary in the dimension “/ human / personality / working address = Kansai” in FIG. At this time, the domain knowledge dictionary of the dimension “/ human / personality / working address = Kanto” in FIG. 28C is not used because it does not meet the user data condition of FIG.

And, for example, the intention analysis unit 20 receives the user utterance sentence “I am Mita” for the system utterance sentence “Where are you?”

In the intention analysis execution unit 21, the morpheme analysis unit 211 performs morpheme analysis on the user utterance sentence “It is Mita”, obtains the morpheme analysis result “Mita: noun: auxiliary verb”, and the syntax analysis unit 212 performs the syntax analysis. To obtain a syntax analysis result as shown in FIG.

Then, as shown in FIG. 27, the intention analysis execution unit 21 analyzes using the domain knowledge dictionary of the dimension “/ human / personality / work address = Osaka” in FIG. Analysis is performed using the domain knowledge dictionary of the dimension “/ human / personality / working address = Kansai” in FIG.

As a result, using the domain knowledge dictionary of “/ Human / Personality / Work address = Kansai” dimension in FIG. 28 (B), the dictionary address “address-Mita city” corresponding to the parsing result “address-Mita” is searched. Then, an intention analysis result that the current address of the user is “Mita City” in Hyogo Prefecture is obtained and output (see FIG. 29). In this way, by using the conditions already given by the user, the ambiguity between Mita City in Hyogo Prefecture and Mita in Tokyo in the word Mita can be resolved.

(A-3-6) Multidimensional intention analysis processing (dynamic switching processing based on analysis target sentence information)
Next, multidimensional intention analysis processing (dynamic switching processing based on analysis target sentence information) will be described with reference to the drawings.

The analysis target sentence dynamic switching process is a multidimensional intention analysis process when constructing extension target data from target data.

The dynamic switching process of the sentence to be analyzed can be performed using the flowcharts shown in FIGS. 23 and 24, and will be described here with reference to FIGS. In FIG. 23 and FIG. 24, “user data” is described as “target data”.

30A shows the target data given to the intention analysis unit 20. FIG. In the intention analysis unit 20, additional information “analysis target: target data” is attached to the target data. Accordingly, the intention analysis execution unit 21 adds “analysis target: target data” to the target data (steps S5001 and S502).

Since the target data is composed of a plurality of sentences, the intention analysis execution unit 21 divides the target data into one sentence unit (step S503), performs morphological analysis on each sentence of the target data (step S504), and constructs the syntax. Analysis is performed (step S505).

For example, for the sentence “Network design experienced person preferential treatment” which is a part of the target data, a syntax analysis result as shown in FIG. That is, a parse analysis result as shown in FIG. 30B is obtained from a morphological analysis result of “experienced treatment: noun design: noun network: noun”.

The intention analysis execution unit 21 gives the user data to the domain knowledge dictionary manager 23 (step S506).

In the domain knowledge dictionary manager 23 that has received the user data from the intention analysis execution unit 21, the feature extraction unit 231 compares the user data with all the dimension conditions registered in the dimension-specific matching table, and sets each dimension condition. A suitable dimension condition is extracted (steps S601 and S602).

The condition rearrangement unit 232 rearranges the dimensional domain knowledge dictionaries into priorities according to a certain standard with respect to the dimension conditions extracted by the feature extraction unit 231 (step S603), and the dictionary setting unit 233 has a predetermined priority. The dimension-specific domain knowledge dictionary is set in the intention analysis dictionary 22 so as to be looked up by rank (step S604).

For example, when the domain knowledge DB of the domain knowledge dictionary manager 23 has a multidimensional syntax dictionary as shown in FIGS. 32A and 32B, the domain knowledge dictionary manager 23 Based on the extraction conditions, the “/ analysis target = target data” dimension syntax dictionary in FIG. 32A and the base dimension syntax dictionary in FIG.

As a result, as shown in FIG. 31, the domain knowledge dictionary manager 23 enables the dictionary of the “/ analysis target = target data” dimension syntax dictionary of FIG. Set the base dimension syntax dictionary at a low position.

Then, the intention analysis unit 20 draws a dictionary from the “/ analysis target = target data” dimension syntax dictionary in FIG. 32A and outputs an intention analysis result “experience job type = network designer” ( (See FIG. 33).

(A-3-7) Dialog Control Process Linked with Matching Result Next, the dialog control process linked with the matching result will be described in detail with reference to the drawings.

FIG. 34 is a flowchart showing a process in the dialog control unit 11 in the dialog control process linked with the matching result. FIG. 35 is a flowchart showing the matching process in the dialogue control process linked to the matching result.

The domain knowledge manager 31 analyzes the intention of the target data with respect to the target data acquired from the data providing server 80 and stores it in the extended target data 54 on the premise of performing the dialog control process in conjunction with the search process.

34, the dialogue control unit 11 receives an intention analysis result from the intention analysis unit 20 (step S703), stores it as user data in the extended user data 53 (step S702), and starts a matching process (step S703).

For example, the intention analysis result for the user's utterance sentence “Java (registered trademark) program has been working for about five years” through dialogue with the user “Chan” is “/ human / work experience / content = Java (registration)”. Trademark), / human / work experience / years = 5 years ”.

Also, FIG. 36 assumes that “○ -chan” is user data, and the dialogue control unit 11 stores the user data shown in FIG. 36 in the extended user data 53.

The matching unit 40 extracts all the expansion target data that completely matches the user data stored in the expansion target data 54 (step S801).

For example, FIG. 37 shows an example of the expansion target data that perfectly matches the user data. FIG. 37A shows the recruitment data (target data) of “Company A” and the property “Company name” -class “ “Company A”, property “language” -class “Java (registered trademark)”, property “years of experience” -class “3 years or more”, and so on. FIG. 37B shows recruitment data (target data) of “Company B”, property “Company name” -class “Company B”, property “Language used” —class “Java (registered trademark)”, ...

The matching unit 40 requests the domain knowledge manager 31 to create a matching table (step S802).

At this time, the domain knowledge manager 31 generates a multi-dimensional matching table by the method similar to the processing described in (A-3-4), and gives this to the matching unit 40.

Here, FIG. 38 is a multidimensional matching table given to the matching unit.

The matching unit 40 calculates the similarity with the user data for all target data based on the multidimensional matching table (step S803), and rearranges the target data in descending order of similarity (step S804).

At this time, the matching unit 40 checks whether there is an ambiguous matching with respect to the target data having the highest similarity (step S805).

Here, ambiguous matching refers to matching when there is a condition value in the target data, but no condition value exists in the user data.

FIG. 39 is a diagram for explaining a case where the user data of “Chan” in FIG. 36 is matched with the target data of “Company A” in FIG. 37A based on the multi-dimensional matchon table in FIG. It is.

In FIG. 39, there are “job type—programmer”, “use language—Java (registered trademark)”, and “years of experience—more than 3 years” as conditions for the target data of “Company A”. In addition, “programmer”, “Java (registered trademark)”, and “5 years” corresponding to the above condition values also exist in the user data of “Chan”. Therefore, it is determined that there is no ambiguous matching in the user data of “Chan”.

When the matching process is completed, the process proceeds to step S704 in FIG. 34. However, since the next question sentence is not set, the processes in steps S705 to S708 are performed. Steps S705 to S708 correspond to steps S304 to S307 in FIG.

On the other hand, in the case of the user “△ -kun”, the intention analysis result for the user utterance sentence “There is Java (registered trademark) experience” is “/ human / work experience / content = Java (registered trademark)”. The user data for “Δ-kun” is as shown in FIG.

Then, based on the “/ person / current job = programmer” dimension matchon table of FIG. 38, the user data of “Δ-kun” in FIG. 40 and the target data of “Company A” in FIG. FIG. 41 shows a case of matching.

As shown in FIG. 41, as described above, values for the target data of “Company A” include “job type—programmer”, “use language—Java (registered trademark)”, and “year of experience—more than 3 years”. .

However, in “△ kun” user data, there are “Occupation-Programmer” and “Language-Java (registered trademark)” corresponding to the value of the above condition, but “Experience years—over 3 years” There is no corresponding condition value.

Therefore, the matching unit 40 sets the deep question about the property “year of experience” and gives it to the domain knowledge manager 31 (step S806).

Then, the domain knowledge manager 31 sets the property “experience years” as the next question sentence, and the basic question sentence “Please tell me the concrete years of experience” of the property “experience years” to the dialogue control unit 11. Give the next question sentence.

In this case, the process proceeds to step S704, and since the following question message is set in the dialog control unit 11, the basic question message "Please tell me the specific years of experience." Output to the side.

In this way, in the dialogue process with “△ -kun”, it is possible to ask the question “year of experience”. For example, when an answer “5 years” is obtained, the target data of “Company A” “△ -kun” can search for a company that suits him more quickly. In addition, when an answer is obtained that the years of experience are “one year”, a company that is useless even if it is considered by itself can be excluded from the candidate earlier.

Also, if it is not “Company A” but “Company B”, the Fukahori question will not be asked because it is not based on years of experience. Thus, the question can be freely controlled according to the state of the target data that is the search result.

(A-4) Effects of the First Embodiment As described above, according to the first embodiment, matching is performed by using a multidimensional ontology to perform matching related to information retrieval according to the user's intention and situation. Since the importance of can be dynamically changed, more accurate matching can be performed.

In addition, according to the first embodiment, by applying multi-dimensional intention analysis, the dictionary to be applied can be dynamically switched according to the user's intention and situation. Can be selected. In addition, since an accurate selection criterion can be selected, a search target related to a question or matching to the user can be also intended by the user.

Furthermore, according to the first embodiment, the dictionary to be applied can be dynamically switched according to the information related to the target data, so that a selection criterion for accurate information search can be selected.

In addition, according to the first embodiment, by performing dialogue control in conjunction with the matching result, information necessary for matching with the target data can be presented to the user as the next question sentence. A precise matching process can be provided.

(B) Other Embodiments (B-1) Multidimensionalization of intention analysis means that an analysis engine (a dictionary used at the time of analysis) is dynamically used according to a value acquired from a user. The mechanism to use the analysis engine dynamically can be realized by applying the mechanism to use the ontology dynamically through the domain knowledge manager.

(B-2) By making the intention analysis multidimensional, the following can be achieved.

(1) If a person who wants Kansai says “Mita is good”, he can return “Mita City”. Also, those who want Kanto can return “Mita in Tokyo”.

(2) By using a colloquial analysis engine for those who input in spoken language, and Kansai dialect engineer for those who use Kansai dialect, the respective analysis accuracy is improved.

(3) A person who wants to be a producer and a person who wants to be a programmer can understand that the meaning of “create a program” is different.

(4) A person who wants a full-time employee can make an analysis with the annual income (or monthly income) as a default, and a part-time person with an hourly wage as the default salary.

(5) When analyzing job data and user utterances, the analysis accuracy is improved by using different analysis engines. (6) Job information is usually tagged with information. By using different dictionaries for each tag value, it is possible to analyze with high accuracy. For example, in the case of a job title tag, analysis is performed using a dictionary that presupposes the input of “job title”. In the case of the “senior PR” tag, analysis is performed using a dictionary based on the input of the contents described by the employee.

(B-3) Multi-dimensional search means that search conditions are dynamically used according to a value acquired from a user.

For example, even if there is no intention analysis (for example, even if there is no need for intention analysis because the answer from the user is determined like a questionnaire), the same processing is of course possible.

In addition, the presentation request is not necessarily after the search, and may be searched at any time.

Furthermore, in the first embodiment, the matching method is a simple example, but various methods can be widely applied as the matching method. For example, perfect matching may be used, or matching may be arranged in order as in the first embodiment.

Also, it is possible to apply a number of matching tables, display all the matching tables in advance, and display only the results of conditions specified by the user.

(B-4) Interaction between dialogue and search means that the content of the dialogue depends on the search result, such as “search based on the value acquired from the user and ask the user for further necessary information based on the search result”. It means a mechanism that changes dynamically.

In the linkage between dialogue and search, a multidimensional mechanism for ontology, search, and intention analysis may be introduced.

Claims (6)

1. Using the matching table having the correspondence between the value to be searched and the value of the target data corresponding to this value and the corresponding weight, the target data corresponding to the value to be searched is selected from the target data to be searched. In the search system to search,
   A target data storage unit for storing one or more target data to be searched;
   A dimension-specific matching table storage unit that stores one or a plurality of dimension-specific matching tables having different weights depending on the dimension of a combination of conditions;
   A basic matching table storage unit for storing a basic matching table in which the above weights are set when there is no condition;
   A priority storage unit by dimension for storing priority by dimension indicating the priority for each dimension;
   When the value to be searched is set as a condition, from the dimension-specific matching table storage unit,
   A dimension-specific matching table extraction unit for extracting the dimension-specific matching table that meets the above conditions;
   Of the extracted matching tables for each dimension, the contents of the matching table for each dimension are overwritten on the basic matching table in order from the lowest dimension priority according to the dimension priority in the dimension priority storage unit. A multi-dimensional matching table generation unit for generating a multi-dimensional matching table for searching for target data from input values;
   A search system comprising: a matching unit that searches for one or a plurality of target data that matches a value to be searched from the target data storage unit, using the multidimensional matching table generated by the multidimensional matching table generation unit.
2. Using the domain knowledge to be used for dialogue, a question for a user is defined, a question is asked, a user response to the question is received, and a dialogue control unit that performs dialogue progress according to the user answer;
   An intention analysis unit that receives the user response from the dialog control unit and performs an intention analysis based on the user response using an intention analysis dictionary that analyzes the intention of the sentence;
   The dimension-specific matching table extraction unit receives the intention analysis result of the user answer from the intention analysis unit, the intention analysis result is set as a condition, and one or more dimension-specific matching tables that meet the condition are set as the dimension. It is taken out from another matching table storage unit,
   The multidimensional matching table generation unit overwrites the basic matching table with the contents of the extracted matching table for each dimension, and dynamically changes the multidimensional matching according to the intention analysis result of the user answer. To generate a table,
   The matching unit uses the multi-dimensional matching table dynamically changed according to the intention analysis result of the user answer, and one or more matches the intention analysis result of the user answer from the target data storage unit To search the target data of
   The search system according to claim 1.
3. The intent analysis unit
   A dimensional dictionary storage unit for storing different dimensional dictionaries depending on the dimension of the combination of conditions;
   A basic dictionary storage unit for storing a basic dictionary when there is no condition;
   The background knowledge of the user response acquired up to the previous time is set as a condition, and a dimension-specific dictionary extraction unit that extracts one or a plurality of dimension-specific dictionaries that meet the condition from the dimension-specific dictionary storage unit,
   Among the extracted dimensional dictionaries, in descending order of application to the above conditions,
   A dictionary rearrangement unit for rearranging the above-mentioned dictionaries according to each dimension;
   A dictionary setting unit configured to set the dictionary according to each dimension sorted by the dictionary sorting unit as a dictionary for intention analysis,
   Using in order from the higher-dimensional dictionary applied to the conditions set by the dictionary setting unit in order, the intention analysis of the user response acquired this time is performed.
   The search system according to claim 2.
4. The intention analysis unit performs intention analysis also on the target data, and in this case, instead of setting the background knowledge of the user answer as a condition, the auxiliary information existing in the analysis target data is read in advance, and the auxiliary information The search system according to claim 2, wherein: is set as a condition.
5. When the matching unit performs matching using the multi-dimensional matching table, if there is a value that does not match the value to be searched among the values of the target data that match the value to be searched, this shortage The search system according to claim 2, wherein a predetermined value is given to the dialog control unit, and a deep question for obtaining a user answer to the missing value is presented.
6. Using a matching table having a correspondence relationship between the value to be searched and the value of the target data corresponding to this value and the corresponding weight, the target data corresponding to the value to be searched is selected from the target data to be searched. In a medium storing a search program to be searched,
   The search program is
   A target data storage unit for storing one or more target data to be searched;
   A dimension-specific matching table storage unit that stores one or a plurality of dimension-specific matching tables having different weights depending on the dimension of a combination of conditions;
   A basic matching table storage unit that stores a basic matching table in which the above weights are set when there is no condition;
   A priority storage unit by dimension for storing priority by dimension indicating the priority for each dimension, and
   On the computer,
   When a value to be searched is set as a condition, a dimension-specific matching table retrieval process for retrieving each dimension-specific matching table that meets the condition from the dimension-specific matching table storage unit,
   Of the extracted matching tables for each dimension, the contents of the matching table for each dimension are overwritten on the basic matching table in order from the lowest dimension priority according to the dimension priority in the dimension priority storage unit. Multi-dimensional matching table generation processing for generating a multi-dimensional matching table for searching for target data from input values,
   A search program storage medium for executing a matching process for searching for one or a plurality of target data that matches a value to be searched from the target data storage unit, using the multidimensional matching table generated by the multidimensional matching table generation unit.

Images