WO2021124502A1

WO2021124502A1 - Inferring device, inferring method, and inferring program

Info

Publication number: WO2021124502A1
Application number: PCT/JP2019/049767
Authority: WO
Inventors: 淳平羽藤
Original assignee: 三菱電機株式会社
Priority date: 2019-12-19
Filing date: 2019-12-19
Publication date: 2021-06-24
Also published as: JPWO2021124502A1; JP7031079B2

Abstract

An inferring device (10) includes an information linking unit (14) which generates integrated information by linking, in a forward chain, information belonging to mutually different domains, from among dynamically varying external information, using knowledge information provided from a knowledge base (11a), and inference rules provided from a rule database (12a), and an information analyzing unit (15), wherein: the information belonging to mutually different domains is information that can be represented as a directed graph having nodes and edges; the integrated information is an integrated graph generated by combining directed graphs, which constitute the information belonging to different domains; and the information analyzing unit (15) calculates the importance of the nodes, which are constituent elements of the integrated graph, using the probability of reaching each node when a random walk on the integrated graph has converged to a steady state, or using the PageRank algorithm, and relationship data indicating the degree of relationship between the nodes, provided from a relationship database (19a).

Description

Inference device, inference method, and inference program

The present invention relates to an inference device, an inference method, and an inference program.

With the progress of sensor technology and information processing technology, there is an increasing number of forms in which machines grasp human activities and surrounding situations, infer support suitable for the activities and surrounding situations, and perform support based on the inference results. However, it is not possible to accurately grasp human activities and surrounding conditions only from the sensor data acquired by the sensor. For this reason, knowledge that acts as a bridge between sensor data and inference results is important. When constructing such knowledge, a knowledge graph (Knowledge Graph) is often used. The knowledge graph is represented by a triple, which is a set of three items, <subject, predicate, and object>. In the knowledge graph, the relationship between the subject, which is a node, and the object is expressed by a predicate. That is, in the knowledge graph, the relationship between nodes is represented by an edge. The "triple" is also called a "triplet".

Patent Document 1 constructs a knowledge graph by integrating a plurality of knowledge such as weather, human preference, and human behavior, and uses the user's speech, data obtained from a sensor, and the knowledge graph. It discloses an inference method for inferring the intent of. In this inference method, the arrival probability of each node is calculated by randomly walking on the knowledge graph with the node corresponding to the obtained data as the start node, and this is set as the importance of each node. After that, a query indicating the conditions to be searched is generated, and the nodes searched by the query are extracted in descending order of importance and output as an inference result.

Japanese Patent No. 6567218

However, the triple in the inference method of Patent Document 1 indicates "the existence of the relationship between the nodes", but does not indicate the "degree of the relationship between the nodes". Therefore, when there are multiple relationships between nodes that have the same graph structure characteristics of the relationships between the nodes (at this time, the arrival probabilities also match), the relationships between the nodes can be given superiority or inferiority. However, there is a problem that the inference result obtained by the query may not be appropriate.

The present invention has been made to solve the above problems, and an object of the present invention is to provide an inference device, an inference method, and an inference program capable of improving the accuracy of inference.

The inference device according to one aspect of the present invention uses knowledge information provided from a knowledge base including information about humans and inference rules provided from a rule database, and among dynamically changing external information. It has an information coupling unit that generates integrated information by combining information belonging to different domains in a forward chain, and an information analysis unit, and the information belonging to the different domains is as a directed graph having a node and an edge. The integrated information is information that can be expressed, and the integrated information is an integrated graph generated by combining the directed graphs that belong to the different domains, and the information analysis unit is a node that is a component of the integrated graph. A node that is a component of the integrated graph provided by the relationship database and an algorithm of the probability or page rank of reaching the node when the importance is randomly walked on the integrated graph and converges to a steady state. It is characterized in that it is calculated using the relationship data indicating the degree of the relationship between the nodes.

The inference method according to another aspect of the present invention is an inference method executed by a computer, in which the computer includes information about a human being, knowledge information provided by a knowledge base, and inference rules provided by a rule database. And, in the step of generating integrated information by combining information belonging to different domains among dynamically changing external information in a forward chain, the information belonging to the different domains is the step. Information that can be expressed as a directed graph having nodes and edges, and the integrated information is an integrated graph generated by combining the directed graphs that belong to the different domains, and the computer integrates the integrated graph. The importance of a node, which is a component of the graph, is provided from the relationship database with an algorithm of probability or page rank of reaching the node when it randomly walks on the integrated graph and converges to a steady state. It is characterized by having a step of calculating using relationship data indicating the degree of relationship between nodes, which is a component of the integrated graph.

According to the present invention, the accuracy of inference can be improved.

It is a functional block diagram which shows schematic structure of the inference apparatus which concerns on embodiment of this invention. It is a figure which shows the example of the hardware composition of the inference apparatus which concerns on this embodiment. It is a figure which shows another example of the hardware composition of the inference apparatus which concerns on this embodiment. It is a flowchart which shows the operation of the inference apparatus which concerns on this embodiment. It is a figure which shows the example of the triple using the predicate "rdfs: subClassOf" defined in RDFS as Table 1. FIG. It is a figure which shows the example of RDF representation using a directed graph. It is a figure which shows the example of the knowledge base in this embodiment as Table 2. FIG. 3 is a diagram showing an example of data generated by the dynamic information acquisition unit of the inference device according to the present embodiment as Table 3. (A) and (b) are diagrams showing how the graph connection portion of the inference device according to the present embodiment creates a directed graph by using a forward chaining rule which is an inference rule stored in the rule DB. It is a figure which shows the example of the graph obtained by sequentially applying the forward chaining rule which is an inference rule to the graph which expresses the knowledge base shown in FIG. 7 and FIG. (A) to (c) are diagrams showing deductive reasoning, inductive reasoning, and hypothetical reasoning as a graph. It is a figure which shows the example of the dialogue by the inference device which concerns on this embodiment as Table 4. FIG. 5 is a diagram showing an example of data acquired by the graph search unit of the inference device according to the present embodiment by searching the graph as Table 5. It is a figure which shows the example of the integrated graph (integrated graph corresponding to FIG. 13) including the importance of the node searched by the graph search part of the inference apparatus which concerns on this embodiment, and the search result. It is a figure which shows an example of the knowledge graph in the comparative example. When the importance is obtained from the knowledge graph of FIG. 15 by the inference method of the comparative example and the reciprocal of the importance is the distance, the total distance of the paths from any of the starting nodes to any of the target nodes. It is a figure which shows the result to the top 3 of the route which becomes the shortest. It is a figure which shows an example of the knowledge graph in this embodiment. When the importance is obtained from the knowledge graph of FIG. 17 by the inference method according to the present embodiment and the reciprocal of the importance is the distance, the path from any of the starting nodes to any of the target nodes is used. , It is a figure which shows the result to the top 3 of the route which the total distance becomes the shortest. It is a figure which shows another example of the knowledge graph in the comparative example. When the importance is obtained from the knowledge graph of FIG. 19 by the inference method of the comparative example and the reciprocal of the importance is the distance, the total distance of the paths from the start node to any of the target nodes is the shortest. It is a figure which shows the path. It is a figure which shows another example of the knowledge graph in this embodiment. When the importance is obtained from the knowledge graph of FIG. 21 by the inference method according to the present embodiment and the reciprocal of the importance is the distance, the total distance of the paths from the start node to any of the target nodes. It is a figure which shows the path which becomes the shortest.

Hereinafter, the inference device, the inference method, and the inference program according to the embodiment of the present invention will be described with reference to the drawings. The following embodiments are merely examples, and various modifications can be made within the scope of the present invention.

<< 1 >> Configuration The inference device according to the present embodiment integrates information belonging to a plurality of domains different from each other by a method using a graph to create an integrated graph as integrated information. Further, the inference device according to the embodiment can be provided with a configuration for deriving the inference result from the integrated graph as integrated information. Here, the graph is a figure having vertices (that is, nodes) and edges (that is, edges) as constituent elements. The method for deriving the inference result is, for example, a graph analysis method that does not require training data. Graphs are mainly classified into directed graphs and undirected graphs. A directed graph is composed of vertices and edges with orientation (ie, arrows). An undirected graph is composed of vertices and edges that have no orientation. An undirected graph is equivalent to a bidirectional graph composed of vertices and edges with orientations in both directions. The directed graphs in the embodiment are shown, for example, in FIGS. 6, 9 (a) and 9 (b), which will be described later.

FIG. 1 is a functional block diagram schematically showing the configuration of the inference device 10 according to the embodiment of the present invention. The inference device 10 is an device capable of implementing the inference method according to the present embodiment. As shown in FIG. 1, the inference device 10 includes a graph coupling unit 14 as an information coupling unit. Further, the inference device 10 has a knowledge base unit 11 for storing the knowledge base 11a and a rule database unit (also referred to as a "rule DB unit") 12 for storing the rule database (also referred to as "rule DB") 12a. , Dynamic information acquisition unit 13, graph connection unit 14 as information connection unit, graph analysis unit 15 as information analysis unit, graph search unit 16 as information search unit, information output unit 17, relationship analysis. A unit 18 and a relationship database unit (also referred to as “relationship DB unit”) 19 for storing the relationship database (also referred to as “relationship DB”) 19a may be provided.

The knowledge base unit 11 is a storage device that stores a knowledge base 11a composed of knowledge information. The knowledge base unit 11 is a part of the inference device 10. However, the knowledge base unit 11 may be provided in an external device connected so as to be able to communicate with the inference device 10.

The rule DB unit 12 is a storage device that stores a plurality of rules (hereinafter, also referred to as "inference rules") for combining (that is, concatenating) a plurality of graphs. The rule DB unit 12 is a part of the inference device 10. However, the rule DB unit 12 may be provided in an external device connected so as to be able to communicate with the inference device 10.

The dynamic information acquisition unit 13 acquires external information that changes dynamically. The dynamic information acquisition unit 13 acquires information that changes dynamically from moment to moment, converts the acquired information into information in a predetermined format, and outputs the information. External information is obtained from an external device, an application executed by the external device, and the like. External information is acquired from, for example, an application 41, a user interface 42, a GPS (Global Positioning System) 43, a camera 44, the Internet 45, an external database (also referred to as "external DB") 46, a sensor 47, and the like. .. The application 41 is software that operates on a device such as a computer or a portable information terminal.

The graph connection unit 14 is based on the knowledge information provided by the knowledge base 11a, the inference rule provided by the rule DB 12a, and the external information acquired by the dynamic information acquisition unit 13, and is based on a graph (“meaningful”. By dynamically combining "graphs" or "data structures" or "knowledge graphs"), an integrated graph with integrated information is generated.

The relationship analysis unit 18 determines the relationship between the nodes that are the components of the graph based on at least one of the knowledge information held by the knowledge base 11a and the external information handled by the graph connection unit 14. To analyze. However, if it is not necessary to dynamically generate the relationships between the nodes that are the components of the integrated graph, or if the relationships between the nodes are input by user operations, the relationship analysis unit 18 is not provided. It is also possible.

The relationship DB unit 19 is a storage device that stores the relationship DB 19a, and stores data indicating the relationship between the nodes, which is the analysis result of the relationship analysis unit 18. However, the relationship DB unit 19 may be provided in an external device connected so as to be able to communicate with the inference device 10.

The graph analysis unit 15 analyzes the integrated graph generated by the graph connecting unit 14 and calculates the importance of the nodes that are the components of this integrated graph. The importance of each node is stored in the relationship DB 19a with respect to the probability of reaching each node, that is, the probability of reaching each node when, for example, an infinite random walk on the integrated graph is performed to reach a steady state. It is a value converted based on the value representing the relationship being made.

The graph search unit 16 searches for desired information from the integrated graph analyzed by the graph analysis unit 15 using the importance calculated by the graph analysis unit 15. The desired information is, for example, a specific information pattern in the integrated graph analyzed by the graph analysis unit 15.

The information output unit 17 outputs the desired information searched by the graph search unit 16 to the user interface 51.

The user interface 51 is, for example, a speaker as an audio output device for outputting audio, a display as a video output device for displaying video, and the like. The user interface 51 provides the user with desired information searched by the graph search unit 16 by voice, video, or the like.

FIG. 2 is a diagram showing an example of the hardware configuration of the inference device 10 according to the present embodiment. In the example of FIG. 2, the inference device 10 includes a processing circuit 61, a hard disk drive (HDD) 62 as a storage device, an input interface device 63, and an output interface device 64. The processing circuit 61 realizes, for example, the functions of the dynamic information acquisition unit 13, the graph connection unit 14, the relationship analysis unit 18, the graph analysis unit 15, the graph search unit 16, and the information output unit 17 shown in FIG. It is a semiconductor integrated circuit that can be used. The input interface device 63 is, for example, a circuit that communicates with an external device, software executed by the external device, and the like. The output interface device 64 is, for example, a circuit that communicates with an external device, software executed by the external device, and the like.

The processing circuit 61 is, for example, dedicated hardware. The processing circuit 61 is, for example, a single circuit, a composite circuit, a processor that executes a program, a processor that executes a parallel program, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or any of these. It may be a combination of the above.

FIG. 3 is a diagram showing another example of the hardware configuration of the inference device 10 according to the present embodiment. In FIG. 3, components that are the same as or correspond to the components shown in FIG. 2 are designated by the same reference numerals as those shown in FIG. In the example of FIG. 3, the inference device 10 includes a CPU (Central Processing Unit) 71 as a processor as an information processing unit, a memory 72 as a storage device for storing a program, an HDD 62 as a storage device, and an input interface device. 63 and an output interface device 64 are provided. The memory 72 can store the inference program according to the present embodiment. By executing the inference program stored in the memory 72, the CPU 71 includes the dynamic information acquisition unit 13, the graph connection unit 14, the relationship analysis unit 18, the graph analysis unit 15, and the graph search unit 16 shown in FIG. And the function of the information output unit 17 can be realized.

The CPU 71 may include a plurality of cache memories such as a primary cache memory and a secondary cache memory. Further, the program stored in the HDD 62, the memory 72, or the like causes the computer to execute the procedure or method of the process performed by the inference device 10. The memory 72 includes, for example, a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), an EEPROM (Electrically Memory), an EEPROM (Electrically Memory), or an EEPROM (Electrically Memory). It is a volatile semiconductor memory.

In FIGS. 2 and 3, the HDD 62 can be used for storing the learning data. The learning data may be stored in an optical storage device using an optical disk such as a DVD (Digital Versaille Disc), a semiconductor memory, an external storage device, or the like.

Note that the inference device 10 may realize a part of its function by dedicated hardware and another part of its function by software or firmware. For example, the function of the dynamic information acquisition unit 13 is realized by a processing circuit as dedicated hardware, and the functions of the configuration other than the dynamic information acquisition unit 13 are realized by the CPU, which is a processor that executes a program stored in the memory. You may.

<< 2 >> Operation << 2-1 >> Outline of operation FIG. 4 is a flowchart showing the operation of the inference device 10 according to the present embodiment. The operation of the inference device 10 includes advance preparation, which is the operation of Phase Ph1, and execution of inference, which is the operation of Phase Ph2.

Phase Ph1 includes preparation of the knowledge base 11a, which is the process of step S101, and preparation of the rule DB 12a, which is the process of step S102. Further, the phase Ph1 may include the processing of the relationship analysis unit 18 or a part of the processing, and the storage of the processing result in the relationship DB 19a. This process is shown as step S103 in FIG. All or part of the processing of Phase Ph1 may be performed by an device other than the inference device 10.

Phase Ph2 is the acquisition of dynamic information, which is the process of step S104, the combination of graphs, which is the process of step S105, the analysis of the integrated graph, which is the process of step S107, and the search of the integrated graph, which is the process of step S108. And the output of information which is the process of step S109. The graph combination, which is the process of step S104, is executed using the prepared knowledge base 11a and the rule DB 12a. The analysis of the graph, which is the process of step S107, is executed using the relationship DB 19a. Further, the analysis of the graph, which is the process of step S107, may include the process of the relationship analysis unit 18 or a part of the process, and the storage of the process result in the relationship DB 19a. This process is shown as step S106 in FIG.

<< 2-2 >> Step S101: Preparation of Knowledge Base 11a First, a knowledge base 11a for storing various knowledge information used for operating the inference device 10 is created. The knowledge base 11a is stored in the knowledge base unit 11.

The format of the information stored as the knowledge base 11a is, for example, RDF (Resource Description Framework). RDF is a standardized framework for expressing information (resources) on the Internet. In RDF, data is represented using triples consisting of a subject, a predicate, and an object.

Also, RDFS (RDF Schema) is a standardized vocabulary definition for expressing resources. FIG. 5 is a diagram showing an example of a triple using the predicate “rdfs: subClassOf” defined in RDFS as Table 1. Example Y1 shown in Table 1 uses the predicate "rdfs: subClassOf" to express that the ramen class (subject) is a food class (object) using one triple. Here, "ramen class" means a thing belonging to ramen. Also, "food class" means something that belongs to food.

FIG. 6 is a diagram showing an example Y1 of the RDF representation shown in FIG. 5 using a directed graph. As shown in FIG. 6, in the RDF representation, the subject and the object are represented as vertices, that is, nodes, and the predicate is represented as edges, that is, edges. In the RDF representation, a triple consisting of a subject, a predicate, and an object includes a node that is the subject, a node that is the object, and an arrow that is a predicate, that is, an arrow that is an edge from the subject to the object. Expressed as a directed graph.

In the present application, for example, the triple consisting of the subject A0, the predicate B0, and the object C0 is described using parentheses as follows.
<A0, B0, C0>

The knowledge information stored as the knowledge base 11a is, for example, device information, common sense (that is, common sense) information, general knowledge information, ontology of these information, and the like. The knowledge information stored as the knowledge base 11a can include two or more types of device information, common knowledge information, general knowledge information, ontology of these information, and the like.

"Device information" is information about the device. The device information includes, for example, device attributes, device names, device types, device setting items, numerical ranges of device setting items, device usage restrictions, device usage counts, device usage, and the like. The attributes of the device can include, for example, an air conditioner class, a fan class, a music player class, and the like.

"Common sense information" is information related to human common sense. Common sense information is, for example, human values, human thoughts, human behavior, human condition, and the like. Common sense information includes information about human condition and information about human behavior. Information about the human state includes information about the physical state of the person and information about the psychological state of the person. Information about the human condition can include, for example, information that the human feels hot, that the human feels cold, that the human is in trouble, that the human is laughing, and so on. Information about human behavior can include, for example, information such as driving, eating, sleeping, searching, going to a specific place, and the like.

An example F1 of common sense information of the knowledge base 11a is shown below.
F1 = <Eat hot, ten, cold food>
Example of common sense information F1 expresses the knowledge that "human beings eat cold food when it is hot" with one triple of <subject, predicate, object>. The "then" (predicate) in the triple of Example F1 is defined as meaning "when a human being is in the (subject) state, the next action to be taken is the (object) action." Therefore, the triple of Example F1 shows that "when a human being is in a state of" hot "(subject), the next action that a human takes is the action of" eating cold food "(object)." ..

Examples of common sense information of the knowledge base 11a, F2a and F2b, are shown below.
F2a = <Eat cold food, slot, cold food class>
F2b = <Eat cold food, subAct, eat>
Examples of common sense information F2a and F2b state that "the action of" eating cold food "is the target of" cold food class ", and the action as a superordinate concept is" eating ". Is expressed by two triples including <subject, predicate, object>. Example F2a, "slot" (predicate) is defined as meaning "the object targeted by the action of (subject) is the object of (object)." In Example F2b, "subAct" (predicate) is defined as meaning "the action of the superordinate concept of (subject) is the action of (object)."

Note that the common sense information used may be different information for each region, gender, and domain of the application to be inferred. That is, the common sense information of the knowledge base 11a used by the inference device 10 may be switched to any one of a plurality of types of common sense information according to the person or field targeted by the inference device 10.

"General knowledge information" is information about universal facts. "General knowledge information" can include information on engineering, physics, science, biology, medicine, food, geography, religion, economy, politics, history, celebrities, and so on.

An example G1 of general knowledge information of the knowledge base 11a is shown below.
G1 = <ice cream, rdfs: subClassOf, cold food class>
Example of general knowledge information "rdfs: subClassOf" (predicate) in G1 is defined as meaning "a (subject) thing is a subclass of a (object) thing." Example G1 expresses the knowledge that "ice cream (subject) is a subclass of cold food (object)" with one triple of <subject, predicate, object>.

An example G2 of general knowledge information of the knowledge base 11a is shown below.
G2 = <Strawberry ice cream, instanceOf, ice cream class>
Example of general knowledge information "instanceOf" (predicate) in G2 is defined as meaning "a thing of (subject) is an instance of (object)." Example G2 expresses the knowledge that "strawberry ice cream (subject) is an instance (actual state) of ice cream class (object)" with one triple of <subject, predicate, object>. There is.

The ontology is the vocabulary of the target domain, the relationship between the vocabulary and the vocabulary, the class, the attribute, the constraint condition, and the like. The ontology is, for example, the information listed below.
-The predicate "slot" means "the object targeted by the action of (subject) is (object)."
-The subject that the predicate "slot" can take is a vocabulary related to behavior.
-The object that the predicate "slot" can take is a vocabulary related to a class of objects.

FIG. 7 is a diagram showing an example of the knowledge base 11a in the present embodiment as Table 2.

<< 2-3 >> Step S102: Preparation of Rule DB 12a Next, one or more inference rules used to operate the inference device 10 are prepared as rule DB 12a. The rule DB 12a is stored in the rule DB unit 12. The prepared inference rules are to integrate multiple pieces of information into one piece of information by combining information belonging to each of multiple domains that are different from each other, and to generate a new hypothesis from external information obtained from sensors or the like. , Etc. are used.

In order to combine information belonging to each of a plurality of different domains, for example, information on human condition and information on human behavior may be used. For example, the following inference rule examples W1 and W2 may be stored in the rule DB 12a.

Example W1 of the inference rule is "If the temperature outside the vehicle is 30 degrees or higher, the condition outside the vehicle and the driver is hot." The example W1 of the inference rule has a role of combining the temperature information outside the vehicle acquired by the temperature sensor and the information on the human condition.

Example W2 of the inference rule is related to the action of "eating" when the current time is 7:00, 12:00, or 17:00. ". The example W2 of the inference rule has a role of combining the time information which is the current time acquired by the clock and the information about the human behavior.

In addition, the inference rule may generate a new hypothesis by using information input from the outside. For example, when the voice spoken by a human can be collected by the microphone, the following inference rule examples W3 to W5 may be saved as the rule DB 12a.

An example of an inference rule, W3, is "If a person says'I'm full', he ate something at a restaurant just before."
An example of an inference rule, W4, is "When a person says'I'm full', he ate something in the car just before."
An example of an inference rule, W5, is "When a person says'I'm full', he ate something in the house just before."

<< 2-4 >> Step S103: Offline relationship analysis The relationship analysis unit 18 performs the process of step S103 when analyzing the relationship between nodes offline. The following examples H1 and H2 are assumed as specific examples of performing offline relationship analysis.
Example H1 is a case where the node corresponding to the dynamic information acquired in step S103 already exists in the knowledge base 11a. Example H2 statically completes only the relationship analysis of the nodes existing in the knowledge base 11a even if the node corresponding to the dynamic information acquired in step S103 does not exist in the knowledge base 11a, and is separately dynamic. This is a case where only the difference is analyzed online when a node corresponding to the information is added.

The analysis of the relationship between nodes is specifically to analyze the degree of relationship between nodes (that is, the strength of the relationship). For example, in FIG. 7, the predicate shows the relationship between the subject and the object, but does not show how strong the relationship is. For example, if the subject is "hot," the predicate "then" indicates that the subject "hot" is related to the objects "lower the temperature," "open the window," "eat cold food," and so on. However, it does not indicate the strength of the relationship. Therefore, the relationship analysis unit 18 analyzes the degree of the relationship between each node, converts the result into data, and stores it in the relationship DB 19a. In this way, by analyzing the degree of the relationship between each node, converting it into data, and saving it, the strength of the relationship between the nodes can be determined. However, the relationship analysis unit 18 does not have to analyze only the relationship between the node that is the subject and the node that is the object in the triple, but at least directly and indirectly with the node that can be the observation node. Analyze the relationship with all or some of the nodes that have a relationship regardless.

As a measure of the degree of relationship, for example, a method in which a developer or user of an inference device manually defines the degree of relationship between nodes can be considered. In that case, a plurality of degrees of relationship may be specified from various viewpoints.

Further, the degree of the relationship between the nodes may be defined by using another knowledge such as a thesaurus in which the nodes appear, for example, based on the number of steps between the nodes.

In addition, the degree of relationship between nodes may be determined based on the behavior history or characteristics of the target person expressed by knowledge, preferences, and the like. In this case, the degree of relationship between the nodes may be adjusted according to the probability that the action and the condition when the action is performed occur at the same time.

Further, the degree of the relationship between the nodes is determined by using word2vec, which is a technique for vector-expressing words, or doc2vec, which is a technique for vector-expressing a document of an arbitrary length, for a document in which knowledge as a node appears. , The word to be knowledge may be expressed in a vector and determined by using a method capable of calculating the similarity between knowledge. In other words, the relationship analysis unit 18 can generate relationship data using one or more of thesaurus data, human behavior history, human characteristics, human preferences, word2vec, doc2vec, and the like. ..

In this embodiment, there are no restrictions on the specific measure of the degree of relationship. In the present embodiment, the relationship analysis unit 18 can adopt an arbitrary method that can express the degree of the relationship between the nodes in a form that can be handled arithmetically so that it can be handled in the subsequent processing. In general, the degree of relationship between nodes is expressed numerically, so it can be expressed by a square matrix with rows and columns of the number of nodes to be analyzed. At least this square matrix is stored in the relationship DB 19a. This square matrix is expressed as a relational matrix R. The element r _ij of the relation matrix R is a value representing the degree of the relationship between the _{node N i} and the node N _j.

<< 2-5 >> Step S104: Acquisition of dynamic information The dynamic information acquisition unit 13 acquires information that changes dynamically from moment to moment, and converts the acquired information into information in the same format as the knowledge base 11a. And output.

The timing at which the dynamic information acquisition unit 13 acquires information is, for example, at regular time intervals. The timing at which the dynamic information acquisition unit 13 acquires information may be another timing. The timing at which the dynamic information acquisition unit 13 acquires information may be, for example, immediately after the dynamic information acquisition unit 13 receives the information acquisition command input by the user. Further, the timing at which the dynamic information acquisition unit 13 acquires information may be a specific time. The specific time is, for example, when it is 12 o'clock in the daytime, when a vehicle enters a specific urban area, and so on. Further, the dynamic information acquisition unit 13 acquires information at a plurality of timings, such as when a specific time is reached at the position where the information is acquired, or when there is a specific event at the position where the information is acquired. May be satisfied at the same time.

The information acquired by the dynamic information acquisition unit 13 is, for example, sensor information, application information, user's personal information, user's command information, and other information. The information acquired by the dynamic information acquisition unit 13 is also referred to as external information.

The sensor information is, for example, user operation information or utterance information obtained by the user interface 42 such as a touch panel or a microphone. Further, the sensor information may be a sensing result for the outside world or a human being obtained by the camera 44 or various sensors. Various sensors are, for example, thermometers, hygrometers, and the like. Further, the sensor information may be position information obtained by GPS43 or the like.

Taking a car as an example, the sensor information is, for example, information obtained by a visible light camera or an infrared camera installed in the car. The sensor information may be, for example, information indicating the result of detecting the positions of the driver's face and eyes. Further, the camera 44 may be a plurality of cameras installed in the vehicle. Alternatively, the sensor information may be voice information indicating a voice uttered by the driver or a passenger, which is acquired by a microphone installed in the vehicle. Further, the sensor information may be information acquired by a biological sensor, a pressure-sensitive sensor, or the like. The biosensor can include a motion sensor, a blood pressure sensor, an electrocardiogram sensor, a heart rate sensor, and the like.

For example, when the dynamic information acquisition unit 13 acquires the information "the temperature outside the vehicle is 35 degrees" and "the humidity outside the vehicle is 90%" from the temperature sensor and the humidity sensor, it is dynamic. The information acquisition unit 13 converts the information into the following triple examples J1 and J2 composed of <subject, predicate, object>.
J1 = <condition outside the car, temperature, 35 degrees>
J2 = <condition outside the car, humidity, 90%>

Further, when the dynamic information acquisition unit 13 acquires the information "currently 12 o'clock" as the information of the current time obtained from the clock, the dynamic information acquisition unit 13 uses the information as the <subject. , Predicate, object> is converted to the following triple example J3.
J3 = <current time, is, 12:00>

The application information acquired by the dynamic information acquisition unit 13 is, for example, information acquired by an application operating in the user's environment, application usage status, application usage history, screen displayed by the application, and displayed by the application. Including choice display, etc.

The dynamic information acquisition unit 13 acquired information on a nearby store using the application of the car navigation system, and found a store named "Ice cream shop # 6" in the neighborhood. In this case, the dynamic information acquisition unit 13 converts the information into the following triple example J4 composed of <subject, predicate, object>.
J4 = <Nearby shop, has, ice cream shop # 6>

The user command information acquired by the dynamic information acquisition unit 13 is, for example, the content explicitly or implicitly instructed by the user to the inference system including the inference device 10. For example, in the case of "the user activates the air conditioner and lowers the temperature by operating a button", the dynamic information acquisition unit 13 uses the information as the following triple consisting of <subject, predicate, object>. Convert to Example J5.
J5 = <latest command, is, lower temperature>

The user's personal information acquired by the dynamic information acquisition unit 13 is the user's name, the user's family, the user's taste, and the like.

The dynamic information acquisition unit 13 provides, for example, the information "The driver's name is Tokyo Taro, and the driver likes ramen in general." Convert to triple examples J6 and J7.
J6 = <driver, name, Tokyo Taro>
J7 = <driver, like, ramen class>

The external information acquired by the dynamic information acquisition unit 13 includes, for example, weather, news, geographic information, etc. obtained from the Internet 45. Further, the external information acquired by the dynamic information acquisition unit 13 may include information stored in the external DB 46 and provided by a country, a local government, a company, an individual, or the like.

FIG. 8 is a diagram showing an example of data generated by the dynamic information acquisition unit 13 of the inference device 10 according to the present embodiment as Table 3. Examples in Table 3 include conditions outside the vehicle, car navigation information, current time, command information, driver information, and the like.

<< 2-6 >> Step S105: Graph Joining Next, the graph joining unit 14 applies an inference rule saved as a rule DB 12a to, for example, the knowledge base 11a and the output of the dynamic information acquisition unit 13. , Forward chaining is applied sequentially to combine a plurality of graphs to finally generate one integrated graph consisting of the combined graphs. The plurality of graphs constituting the integrated graph are, for example, directed graphs.

The application of the inference rule in the rule DB 12a can be expressed using a graph as follows.
"If a specific information pattern is found in the target graph, add / change / delete nodes or edges to the graph."

A specific example of graph combination will be described using the above-mentioned examples of inference rules W1 and W2. In the following, "? X" indicates an arbitrary variable.

As an example W1 of the inference rule, "If the temperature outside the vehicle is 30 degrees or higher, the condition outside the vehicle and the driver is hot." Is expressed using a graph as follows.
"If the information pattern of triple <outside condition, temperature,? X> is found in the target graph and the condition"? X ≥ 30 "is satisfied, the graph shows triple <outside condition, temperature, hot>. , Triple <driver, temperature, hot>, and newly added. "

9 (a) and 9 (b) are diagrams showing how the graph connecting portion 14 of the inference device 10 according to the present embodiment creates a directed graph using a forward chaining rule which is an inference rule stored in the rule DB 12a. Is.

As an example W2 of the inference rule, "when the current time is 7 o'clock, 12 o'clock or 17 o'clock, the current time is related to the action of" eating "" is expressed graphically as follows.
"A triple <current time, is,? X> information pattern is found in the target graph, and the condition"? X = 7 o'clock "or"? X = 12 o'clock "or"? X = 17:00 "is satisfied. If so, add a new triple <current time, relate, eat> to the graph. "

For example, when the inference rule examples W1 and W2 are sequentially applied to the graphs showing Table 2 shown in FIG. 7 and Table 3 shown in FIG. 8, the graph as shown in FIG. 10 is obtained. can get. The graph of FIG. 10 is an example.

As described above, the graph connecting unit 14 can generate one integrated graph as shown in FIG. 10 from the information obtained from the sensor or the information belonging to each of the plurality of domains different from each other. .. That is, the graph connecting unit 14 can generate integrated information by combining information belonging to each of a plurality of domains different from each other by sequentially applying the rules by a forward chain. In addition, the integrated information can be expressed as an integrated graph.

In particular, the graph connecting unit 14 uses information about humans in order to combine information belonging to each of a plurality of domains different from each other. Information about humans includes, for example, information about human conditions and information about human behavior. Information about the human condition is, for example, "hot", "cold", "hungry", and so on. Information about human behavior is, for example, "lowering the temperature", "raising the temperature", "eating cold food", and the like.

For example, by applying the example W1 of the inference rule, the information of the sensor domain "state outside the vehicle" and the information of the food domain "chilled Chinese" can be obtained, that is, the information belonging to each of a plurality of domains different from each other. Can be combined.

Specifically, by using triples K1 to K4 consisting of <subject, predicate, and object>, "information outside the vehicle" and "chilled Chinese" can be combined using a directed graph.
K1 = <Outside the car, status, hot>
K2 = <Eat hot, ten, cold food>
K3 = <Eat cold food, slots, cold food class>
K4 = <Chilled Chinese, instanceOf, cold food class>

<< 2-7 >> Step S106: Online Relationship Analysis The relationship analysis unit 18 performs the process of step S106 when analyzing the relationship between nodes online. As a specific case of performing online analysis, there is a case where the node added in step S105 includes a node that does not statically exist in the knowledge base prepared in step S101. In this case, since the necessary relationship analysis is not completed even if step S103 is executed, the relationship analysis unit 18 needs to execute step S106. For example, there is a "driver" as a node that does not exist in FIG. 9 (a) but exists in FIG. 9 (b). Since the knowledge graph at the time of step S103 is in the state of FIG. 9A, the relationship between the “driver” and the other nodes cannot be analyzed. Therefore, the relationship analysis unit 18 analyzes the relationship between the "driver" and the other nodes in step S106.

The analysis of the relationship between the nodes performed in step S106 has the same basic processing contents as the analysis of the relationship in step S103, but in step S103, at least the relationship between the nodes related to the nodes for which the analysis of the relationship has not been completed. Perform analysis.

<< 2-8 >> Step S107: Graph Analysis Next, the graph analysis unit 15 traces the nodes of the integrated graph (for example, FIG. 10) generated by the graph combination in the graph connection unit 14 through the edges. Then, probabilistic inference is executed, and the importance of each node is calculated as a result of the inference. The graph analysis unit 15 executes deductive inference (deduction), inductive inference (induction), and hypothetical inference (abduction) by tracing the nodes on the integrated graph. 11 (a) to 11 (c) are diagrams showing deductive reasoning, inductive reasoning, and hypothetical reasoning as a graph.

For example, as in the graph shown in FIG. 11A, when there are triples <A0, x, B0> and triples <B0, y, C0>, the nodes A0 to C0 are deduced by tracing the graph. Can be inferred.

Specifically, if there is a triple <Socrates, is, person> and a triple <People, is, die>, the node "Die" is inferred from the node "Socrates", and "Socrates dies" is obtained as a conclusion. be able to.

Further, for example, as shown in the graph shown in FIG. 11B, the triple <A0, x, B0>, the triple <A0, x, B2>, the triple <B0, y, C0>, and the triple <B2 , Y, C0> and triple <B3, y, C0>, trace from node A0 to node C0 via node B0 or B2, and finally from node C0 to node B3. Then, the node B3 can be inductively inferred from the node A0.

Specifically, triple <Taro-san, like, soy sauce ramen>, triple <Taro-san, like, Tonkotsu ramen>, triple <soy sauce ramen, rdfs: subClassOf, ramen class>, triple <Tonkotsu ramen, rdfs: Consider the case where subClassOf, ramen class> and triple <miso ramen, rdfs: subClassOf, ramen class> are included in the graph. In this case, the node "Taro-san" and the node "Miso Ramen" are not directly connected in the graph. However, the node "Taro-san" likes the node "Shoyu ramen" and the node "Tonkotsu ramen", so he likes "Ramen class" which means ramen in general. In that case, the node "Taro-san" also likes the node "Miso Ramen". That is, such inductive reasoning is possible.

Further, for example, as shown in the graph shown in FIG. 11C, when the event of triple <A0, x, B0> occurs, the triple <B0, y, C0> and the triple <B0, y, C2> By following the graph, the hypothesis of node C0 or node C2 can be inferred from node A0.

Specifically, the triple <Hanako-san, 12 o'clock action, ate lunch>, the triple <I ate lunch, the next action, take a nap>, and the triple <I ate lunch>. , Next action, take a walk>, there are two hypotheses: the hypothesis "Hanako's next action is to take a nap" and the hypothesis "Hanako's next action is to take a walk." (That is, the theory that you do not know which is correct) can be inferred. Such reasoning is also called hypothetical reasoning.

The graph generated by the graph connecting unit 14 (for example, FIG. 10) is a cycle graph in which each node is complicatedly connected. Therefore, in the integrated graph generated by the graph connecting unit 14 (for example, FIG. 10), the conclusion of inference can be specified as one by simply following the integrated graph as shown in FIGS. 11A to 11C. I can't. Therefore, for example, the probability of reaching each node, that is, the arrival probability when the integrated graph (for example, FIG. 10) calculated by the graph connecting unit 14 is randomly walked infinitely and becomes a steady state is calculated. Since this probability can be interpreted as the "probability that the node is the conclusion of inference", this probability is defined as the importance of each node. Random walking on the integrated graph is equivalent to probabilistically performing deductive reasoning, inductive reasoning, and hypothetical reasoning.

Next, the contents of steps S107a, S107b, and S107c, that is, step S107 shown in FIG. 4, which show a specific calculation method of the graph analysis unit 15, will be described.

<Step S107a: Graph transformation and transition matrix creation>
The graph analysis unit 15 appropriately converts the integrated graph input to the graph analysis unit 15 to create a transition matrix between nodes for a random walk.

First, a new graph L2 is created according to the type of verb of each triple included in the graph L1 input to the graph analysis unit 15. The graph L2 created by the graph analysis unit 15 is also referred to as an "inference graph". In creating the inference graph, for example, any of the following conversion pattern TP1, conversion pattern TP2, and conversion pattern TP3 is applied.

In the "conversion pattern TP1", the directed graph which is the input graph L1 is used as it is as the inference graph L2. In the "conversion pattern TP2", the inference graph L2 is created by converting the input graph L1 into an undirected graph, that is, a bidirectional graph. In the "conversion pattern TP3", the inference graph L2 is created by converting the input graph L1 into a directed graph in the opposite direction.

Next, an adjacency matrix of the inference graph L2 is created, and the transition matrix P between the nodes in which the total value of each column of this matrix is normalized by 1 is obtained. Assuming that the number of nodes is N (N is a positive integer), the transition matrix P between the nodes is a matrix having a size (N × N).

Here, the element P (i, j) of the transition matrix P indicates the transition probability from the j-th node to the i-th node in the random walk on the integrated graph. j and i are positive integers. For example, when a certain node A5 is the subject and the nodes connected to the node A5 are four nodes A6 to A9, the transition probability from the node A5 to each of the nodes A6 to A9 is 0, which is an equal value. Set to .25. However, the transition probabilities from node A5 to each of nodes A6 to A9 do not have to be equal values. The transition probabilities set may be different values from each other.

As shown in the conversion pattern TP2, when the inference graph L2 is created by converting the input graph L1 into an undirected graph, the following effects can be obtained.

For example, in the triple of "hot, then, eat cold", it can be inferred that "hot" means "eat cold", and conversely, "eat cold" reverses the graph. It can be inferred that it is "hot". Further, as in the example of inductive reasoning shown in FIG. 11B, the conclusion of the reasoning can be obtained by tracing the graph from node C0 to node B3, that is, in the reverse direction.

For example, in the triple <j, x, i>, the initial settings of the transition probabilities of the elements P (i, j) and the elements P (j, i) of the transition matrix P are as follows.
P (i, j) = α (0 <α ≦ 1)
P (j, i) = 0

However, when the graph L1 which is a directed graph is converted into the graph L2 which is an undirected graph according to the conversion pattern TP2, in the triple of the converted graph L2, the elements P (i, j) of the transition matrix P and The initial setting of the transition probability of the element P (j, i) is as follows, for example.
P (i, j) = P (j, i) = α (0 <α ≦ 1)
Alternatively, the transition probability is set as follows with a bias.
P (i, j) = α (0 <α ≦ 1)
P (j, i) = β (0 <β ≦ 1)
Here, α and β are not equal to each other. As a result, in the random walk process described later, it is possible to freely move back and forth between the node i and the node j.

Next, as shown in the conversion pattern TP3, when the inference graph L2 is created by converting the input graph L1 into a directed graph in the opposite direction, the following effects can be obtained.

For example, if there is a triple <Taro-san, like, ramen> and a triple <Taro-san, dislike, durian>, consider the task of inferring what Taro-san likes.

If the above triple <Taro-san, like, ramen> and the triple <Taro-san, dislike, durian> are used as they are for inference, "ramen" and "durian" will be used in the importance calculation step of step S105c described later. May be of similar importance. Therefore, in the verb "dislike", the directed graph is reversed. This makes it possible to reduce the probability of passing through the "durian" when randomly walking the inference graph in the importance calculation step described later. Therefore, it has the effect of lowering the importance of "Durian".

Next, as shown in the conversion pattern TP3, a method of converting the input graph L1 directed graph into a directed graph in the reverse direction will be described below.
In particular,
In the triple <j, x, i>, the initial setting of the transition probability is as follows.
P (i, j) = α (0 <α ≦ 1)
P (j, i) = 0
Further, when converted to a directed graph in the reverse direction, the initial setting of the transition probability is as follows.
P (i, j) = 0
P (j, i) = α (0 <α ≦ 1)

<Step S107b: Observation node setting>
The graph analysis unit 15 sets an important node as a starting point of inference, such as information acquired from the outside or information detected by a sensor, as an “observation node”.

Next, the graph analysis unit 15 sets the weight of the observation node to 0 or more, and creates an N-dimensional observation vector v with the other nodes as 0. The sum of the probability distributions of the observation vector v is 1. For example, if the observation nodes are "outside the vehicle", "nearby shop", "most recent command", and "driver", the elements of the observation vector v corresponding to those nodes are 0.25 (each). = 1/4) is set. The element of the observation vector v corresponding to the other nodes is set to 0.

By restarting the random walk from the observation node in the process described later, there is an effect that the node closer to the observation node becomes more important.

<Step S107c: Calculation of importance>
Next, the graph analysis unit 15 sets the transition matrix P and the observation vector v, that is,

Is used to obtain the importance of each node included in the inference graph (graph L2). That is, the graph analysis unit 15 determines the start node, which is the node as the start point of the random walk on the integrated graph, based on the probability of the observation vector v, and the relation matrix stored in the transition matrix P and the relation DB 19a. The probability of reaching each node is obtained by randomly walking on the inference graph by R, and this is set as the importance of each node.

Specific calculation examples are as follows, for example. First, the inference matrix M _k _{when one start node N k} is selected from the set of nodes that can be the start node, that is,

Is defined as the following equation (1). k is a positive integer.

Here, α indicates a predetermined constant satisfying 0 <α ≦ 1.

Shows a transition matrix considering the relationship between each N × N node and the start node N _k. The elements of a transition matrix P and _{p ij,} when the elements of the relationship matrix R was _{r ij,} the elements of transition matrix _{P k} _is expressed by p ij _{* r ik.}

_{Here, the sum of the elements pij} of the transition matrix P with respect to j (that is, the sum in the column) is expressed by the following equation (2).

Thus, is the sum for j of _p ij _{* r ik} is an element of the transition matrix _{P k}

Is not always 1, and is inappropriate as a matrix expressing the transition probability. Therefore, it is necessary to make adjustments when obtaining the relational matrix R so that the sum of the columns is 1. Alternatively, without adjustment for obtaining the relationship matrix R, the elements of row i and column j of the transition matrix P _k is taken as p _ijk, as shown in the following equation (3), an element p _ijk , Normalize each column so that the sum of the columns is 1.

Further, before the above processing, _{if the rik} becomes a negative value, the transition probability becomes an inappropriate value. In that case, _{adjustments such as setting the rik} to 0 may be performed. Further, _{the normalization of the element pijk} is not limited to that of the equation (3), and a method of normalizing such that the sum of the columns is 1 may be adopted.

_{Further, the element p ijk} of the transition matrix P _{k in} the i-th row and the j-column may be defined as in the following equation (4), and the influence of the _{r ik may be controlled by the value of γ.}
p _ij * (1-γ + r _ik * γ) (4)
Here, γ is a value in the range of 0 to 1. _{By defining the element pijk in} this way, for example, by setting γ = 0.5, it is possible to make adjustments such as suppressing the relationship with the _{start node Nk in half with respect to the transition probability.} ..

When the equation (4) _{is expressed as pijk} ′ and the adjustment corresponding to the equation (3) is performed, _{the element pijk} of the i-by-j column of _{the transition matrix P k} may be the following equation (5). ..

In equation (1)

Indicates an observation vector v of size N × 1.

In equation (1)

Indicates a vector having a size of 1 × N in which all elements are 1, and is also referred to as ^{a vector 1 T.}

The first term on the right side of equation (1) indicates that the integrated graph is randomly walked based on the transition matrix P. The second term on the right side of equation (1) indicates that the random walk is restarted from the observation node with a probability of (1-α) in the middle of the random walk.

Next, the N-dimensional importance vector for _{the starting node Nk} _{is x vk} , that is,

Then, the importance vector assuming the _{start node N k} is obtained by the following equation (6).

Equation (6) represents that the importance vector x _vk has converged by repeating the random walk. x _vk (i) indicates the importance of the i-th node assuming the _{start node N k.}

The power method can be used to calculate the _{importance vector x vk} in the equation (6). For example, in _{the calculation of the importance vector x vk} , a method similar to PageRank, which is an algorithm for determining the importance of a web page, can be used. Further, in _{the calculation of the importance vector x vk} , the same calculation method as the PageRank described in Non-Patent Document 1 can be used.

Further, from the equations (1) and (6), the cost function CF1 _k , that is, by the following equation (7), is used.

_{By calculating the importance vector x vk} having a sum of 1 so as to minimize the value of this cost function, the importance of the node assuming the start node N _{k may be calculated.}

In equation (7)

Indicates an N × N identity matrix, which is also referred to as an identity matrix E.

The importance of the node is calculated on the premise of all the start nodes by the same method as the method of calculating _{the importance x vk} of the node on the premise of the start node _{Nk described above.}

<< Clustering of start nodes >>
_{Since the importance x vk} of the node assuming the start node _Nk is calculated for all the start nodes, the number of calculations increases in proportion to the number of start nodes. Therefore, each start is performed before the importance is calculated. node N _i, if any rule for N _j is constant or can be considered a method to handle as a single start node V _N virtually. As a reference, it is conceivable to use the value of the _{relation rij.} Relationship to V _N, each starting node N _i, it is conceivable to use to create a composite value such as taking the average of the N _j each relationship.

If the node to be selected as the start node is determined in advance, a set of start nodes that can be used as a virtual start node is obtained in advance, and then the virtual start node is included in addition to each node. It is also possible to obtain the relational matrix in step S103 or step S107.

_{Next, the procedure for obtaining the} overall importance x _v from all the importance x vk will be described.

<< Calculation method of importance 1 >>
As one of the procedures for obtaining the overall importance x _v from all the importance x _vk, _{there is a method of defining the importance x v} (i) for each node i as shown in the following equation (8).

Severity obtained in this way is the sum of values obtained by multiplying the importance weight w _k and the node i on the premise of start node N _k for each starting node N _k. The weight w _k may be a value set in advance or may be a variable value according to the situation. Further, by setting all the _{weights w k} to 1, it is also possible to make the sum of only the importance of the node i on the premise of the _{start node N k having no weight.}

That is, when the graph analysis unit 15 determines the start node which is the node as the start point of the random walk on the integrated graph and calculates the importance of each node of the integrated graph for each start node, the graph analysis unit 15 determines each start node. The weight w _k can be set, and the importance of each node on the integrated graph can be calculated based on the _{weight w k for each start node.} Further, the graph analysis unit 15 determines a start node which is a node as a start point of a random walk on the integrated graph, and calculates the sum of the importance of each node on the integrated graph for each start node as the importance of each start node. When calculating as, the weight w _k can be set for each start node, and the importance of each start node can be calculated based on the _{weight w k for each start node.}

<< Calculation method of importance 2 >>
As another procedure for obtaining the overall importance x _v from all the importance x _vk, _{there is a method of defining the importance x v} (i) for each node i as shown in the following equation (9).

However, k is an element of a set of _{k in} which N k that can reach the target node exists.

Here, the target node is a node to be derived by inference, and is a node finally obtained in step S106 (search for a graph) described later, and specifically, corresponds to "? Action". It is one of a set of nodes. However, although the importance vector x _v is used in step S106, in the importance calculation method 2, when the target node is obtained, x _vk is used _{instead of x v.}

In the importance calculation method 2, in addition to the condition of the importance calculation method 1, a restriction regarding k is provided that "k is an element of a set of _{k in which N k that can reach the target node exists".} It is a thing. Here, there is a method of considering the reachability of the graph as the reachability of the graph. In addition, there is also a method of adding some criterion to the reachability of the graph, such as considering the reachability in a network in which the importance of the node i assuming the _{start node Nk is limited to the nodes equal to or higher than the threshold value.}

In the importance calculation method 2, when the starting node _Nk cannot reach the target node, it means that the route to the target node cannot be obtained, that is, it is not related. The importance of such a start node means not to consider. On the other hand, in the importance calculation method 1, there is a difference in considering the importance of irrelevant start nodes.

<< 2-9 >> Step S108: Graph Search The graph search unit 16 searches for a specific information pattern from the graph L2 by using the _{importance vector x v in order to extract desired information.} A specific operation example of the graph search unit 16 will be described.

FIG. 12 is a diagram showing dialogue examples D1 and D2 as Table 4. Dialogue Examples D1 and D2 are examples of intellectual dialogue by an artificial intelligence agent (also referred to as "AG") that makes humans (also referred to as "users") think "smart".

As in dialogue example D2, AG said, "There is a chilled Chinese restaurant nearby. In order to make a nifty utterance automatically, two elemental technologies of "information integration" and "inference" are required. "Information integration" is an elemental technology that integrates information belonging to each of a plurality of different domains. Information belonging to each of a plurality of different domains includes, for example, weather, human operation history, navigation (also simply referred to as "navigation") information, human preference, and knowledge of human behavior. Humans are, for example, drivers. Knowledge of human behavior is information about human behavior, such as "eating cold food when it is hot."

Here, in the graph search unit 16, the subject, predicate, and object are <? action, slot ,? slotClass> and <? slotValue, instanceOf ,? An information pattern satisfying slotClass> is extracted from the integrated graph. Here, the character string starting with "?" Is a variable, and an arbitrary character string is entered.

For example, in the examples of Tables 2 and 3, "satisfy triple <eat cold food, slot, cold food class> and triple <chilled Chinese, instanceOf, cold food class>" or "triple <lower temperature, slot, air conditioner". Class> and triple <air conditioner # 1, temperatureOf, air conditioner class> are satisfied "corresponds to the information pattern to be searched.

FIG. 13 is a diagram showing an example of data acquired by the graph search unit 16 of the inference device 10 according to the present embodiment by searching the integrated graph as Table 5. That is, FIG. 13 shows the results of sorting the graph search results by the importance of "? Action" and then by the importance of "? SlotValue".

FIG. 14 is a diagram showing an example of an integrated graph (integrated graph corresponding to FIG. 13) including the importance of the nodes searched by the graph search unit 16 of the inference device 10 according to the present embodiment and the search results. In FIG. 14, the values in parentheses indicate the importance. The hatched squares indicate observation nodes, and the hatched circles indicate nodes of high importance.

In this example, at the top,
{SlotValue:'Chilled Chinese', slotClass:'Cold Food Class', action:'Eat Cold Food'}
Is obtained. This provided information for the in-vehicle agent to recommend to the user, "Why don't you eat chilled Chinese food?"

In the example of FIG. 13, the top-level search results include {'slotClass':'cold food class','action':'eat cold food','slotValue':'chilled Chinese'}.
Is obtained. The reason is that the top search results are
"The driver likes ramen, that is, <driver, like, ramen class>", "There is a ramen shop Ra and an ice cream shop Ic in a nearby shop.", "The current time is 12 o'clock. And it is related to eating behavior. "And" The temperature inside the car is 35 degrees, so the condition is hot. Therefore, humans eat cold foods or lower the temperature. " This is because the inference of the stage can be obtained by following the integrated graph by a random walk. Also, the node "'eat cold food'or'chilledChinese'" became more important (ie, more likely to be the conclusion of reasoning).

<< 2-10 >> Step S109: Output of Information Finally, the information output unit 17 outputs information in an appropriate form to a user interface 51 such as a display or a speaker by using the output of the graph search unit 16. For example, in the dialogue examples D1 and D2 in Table 1, the information output unit 17 outputs a voice such as "There is a chilled Chinese restaurant nearby" by voice synthesis.

<< 3 >> Effect FIG. 15 is a diagram showing an example of a knowledge graph in a comparative example. In the knowledge graph of FIG. 15, the graph structure when the start node a11 is the start point and the graph structure when the start node a12 is the start point are the same. Therefore, when the inference method of the comparative example described in Patent Document 1 is used, the total value of the importance of the nodes on the route from any of the start nodes a11 and a12 to any of the target nodes c11, c12, and c13. Is the same value for all routes.

In FIG. 16, when the importance is obtained from the knowledge graph of FIG. 15 by the inference method of the comparative example and the reciprocal of the importance is the distance, the target nodes c11, c12, and c13 are taken from any of the start nodes a11 and a12. It is a figure which shows the result to the top 3 of the route which the total distance becomes the shortest among the routes to any of. In the inference method of the comparative example, as shown in FIG. 16, since the total distances in all of the cases C-11, C-12, and C-13 have the same value, which route is the most important route? It is indistinguishable.

On the other hand, the inference device 10 according to the present embodiment determines the importance using the relationship data indicating the degree of the relationship between the start node and each node. FIG. 17 is a diagram showing an example of a knowledge graph according to the present embodiment. In FIG. 17, an index indicating the degree of the relationship between the nodes is shown numerically (for example, “0.1”, “0.5”, “1.0”). The numerical value in the upper row in the circular region indicating the node indicates the degree of the relationship with the start node a21, and the numerical value in the lower row indicates the degree of the relationship with the start node a22. In FIG. 17, the sum of the numerical values indicating the degree of the relationship between each of the starting nodes a21 and a22 and each of the intermediate nodes b21, b22 and b23 is "1.1 (= 0.1 + 1.0)" and "1", respectively. It is "0.0 (= 0.5 + 0.5)" and "1.1 (= 1.0 + 0.1)". Further, in FIG. 17, the numerical value indicating the degree of the relationship between each of the start nodes a21 and a22 and each of the target nodes c21, c22 and c23 is “1.0”. Therefore, the importance of the route from the start node a21 or a22 to the target node c21 via the intermediate node b21 and the route from the start node a21 or a22 to the destination node c23 via the intermediate node b23 is determined by the start node a21 or It is higher than the importance of the route from a22 to the target node c22 via the intermediate node b22.

In FIG. 18, when the importance is obtained from the knowledge graph of FIG. 17 by the inference method according to the present embodiment and the reciprocal of the importance is taken as the distance, the target node c21 from any of the start nodes a21 and a22. It is a figure which shows the result to the top 3 of the route which the total distance becomes the shortest among the routes to any of c22 and c23. In this inference method, as shown in FIG. 18, the total distance in each of the cases C-21 and C-22 is the same value, and the total distance in the case C-23 is the total distance in the cases C-21 and C-22. Longer than the total distance. That is, in the case of FIG. 18, by using the degree of relationship, the importance can be calculated based on the degree of relationship.

FIG. 19 is a diagram showing another example of the knowledge graph in the comparative example. In the knowledge graph of FIG. 19, the intermediate node b31 is used to reach the target node c31 from the start node a31, and the intermediate node b32 is used to reach the target node c32 from the start node a31. Here, the intermediate node b31 has a route only from the start node a31, whereas the intermediate node b32 has a route from nodes a32 to a37 other than the start node a31. However, when all of these nodes a32 to a37 are routes from nodes that are not semantically related to or have a low degree of relationship with the start node a31, the comparative example described in Patent Document 1 In the inference method of, the intermediate node b32 having many inflow routes is calculated to be more important than the intermediate node b31.

FIG. 20 shows the path from the start node a31 to any of the target nodes c31 and c32 when the importance is obtained from the knowledge graph of FIG. 19 by the inference method of the comparative example and the reciprocal of the importance is the distance. It is a figure which shows the result of the total distance. In the inference method of the comparative example, as shown in FIG. 20, the total distance in the case C-31 passing through the intermediate node b32 having many inflow paths passes through the intermediate node b31 having few inflow paths. It is determined that the importance of the route of Case C-31 is higher than the importance of the route of Case C-32, which is shorter than the total distance in. However, it is not appropriate to determine that the route of Case C-31 is of high importance because there are many routes that flow in from the nodes a32 to a37 whose degree of relationship with the start node a31 is "0". ..

On the other hand, the inference device 10 according to the present embodiment determines the importance using the relationship data indicating the degree of the relationship between the start node and each node. FIG. 21 is a diagram showing another example of the knowledge graph in the present embodiment. In FIG. 21, an index indicating the degree of the relationship between the nodes is shown numerically (for example, “0.9”, “0.5”, “1.0” in the circle indicating the node). In FIG. 21, the numerical values indicating the degree of the relationship between the start node a41 and each of the intermediate nodes b41 and b42 are “0.9” and “0.5”, respectively. Further, in FIG. 21, the numerical value indicating the degree of the relationship between the start node a41 and each of the target nodes c41 and c42 is “1.0”, respectively. Therefore, the importance of the route from the start node a41 to the destination node c41 via the intermediate node b41 is higher than the importance of the route from the start node a41 to the destination node c42 via the intermediate node b42.

In FIG. 22, the importance is obtained from the knowledge graph of FIG. 21 by the inference method according to the present embodiment, and when the reciprocal of the importance is taken as the distance, from the start node a41 to any of the target nodes c41 and c42. It is a figure which shows the route which the total distance becomes the shortest among the routes of. In this inference method, as shown in FIG. 22, the total distance in case C-41 is shorter than the total distance in case C-42. That is, in the case of FIG. 22, the importance can be appropriately calculated based on the degree of relationship.

As described above, by using the inference device 10, the inference method, or the inference program according to the present embodiment, the degree of the relationship between the nodes is taken into consideration as compared with the case where the importance is calculated only by the graph structure. It is possible to make inferences and improve the accuracy of inferences.

10 inference device, 11 knowledge base, 11a knowledge base, 12 rule DB, 12a rule DB, 13 dynamic information acquisition, 14 graph connection (information connection), 15 graph analysis (information analysis), 16 Graph search unit (information retrieval unit), 17 information output unit, 18 relationship analysis unit, 19 relationship DB unit, 19a relationship DB, 41 application, 42 user interface, 43 GPS, 44 camera, 45 Internet, 46 external database, 47 sensors, 51 user interface.

Claims

Using knowledge information provided by the knowledge base, including information about humans, and inference rules provided by the rules database, information belonging to different domains of dynamically changing external information is forward-chained. An information coupling part that generates integrated information by combining,
It has an information analysis department
The information belonging to the domains different from each other is information that can be expressed as a directed graph having nodes and edges.
The integrated information is an integrated graph generated by combining the directed graphs which are information belonging to the different domains.
The information analysis unit relates the importance of the node, which is a component of the integrated graph, to the probability of reaching the node or the PageRank algorithm when the node randomly walks on the integrated graph and converges to a steady state. An inference device provided from a sex database, which is calculated by using relationship data indicating the degree of relationship between nodes, which is a component of the integrated graph.
It also has a relationship analysis unit that generates the relationship data.
The inference device according to claim 1, wherein the relationship analysis unit stores the relationship data in the relationship database.
When the relationship analysis unit calculates the importance of a node that is a component of the integrated graph, if the relationship data for the node for which the importance is calculated does not exist in the relationship database, the relationship The inference device according to claim 2, wherein data is generated and stored in the relationship database.
The claim is characterized in that the relationship analysis unit generates the relationship data using one or more of cissolus data, human behavior history, human characteristics, human preference, word2vec, and doc2vec. The inference device according to 2 or 3.
Any one of claims 1 to 4, wherein the information analysis unit creates a transition matrix between nodes of nodes that are components of the integrated graph, which is used in a random walk on the integrated graph. The inference device described in.
The relationship data is represented by a relationship matrix.
The inference device according to claim 5, wherein the information analysis unit adjusts the relational matrix so that the sum of each column of the elements of the transition matrix is 1.
The inference device according to any one of claims 1 to 6, wherein the information analysis unit adjusts the degree of influence of the relationship data when calculating the importance.
The information analysis unit determines a start node which is a node as a start point of a random walk on the integrated graph, and calculates the importance of each node of the integrated graph for each start node. The inference device according to any one of claims 1 to 7.
The eighth aspect of the present invention, wherein the information analysis unit sets a weight for each start node and calculates the importance of each node on the integrated graph based on the weight for each start node. Inference device.
The information analysis unit determines a start node which is a node as a start point of a random walk on the integrated graph, and sums the importance of each node on the integrated graph for each start node for each start node. The inference device according to any one of claims 1 to 7, wherein the inference device is calculated as an importance.
The inference device according to claim 10, wherein the information analysis unit sets a weight for each start node and calculates the importance of each start node based on the weight for each start node.
It also has an information retrieval unit for searching the nodes of the integrated graph.
The inference device according to any one of claims 1 to 11, wherein the information retrieval unit searches for the node based on the importance.
The inference according to any one of claims 1 to 12, further comprising a dynamic information acquisition unit that converts the external information into information in a predetermined format and provides the information coupling unit. apparatus.
The knowledge base section that stores the knowledge base and
The rule database section that stores the rule database and
The relationship database section that stores the relationship database and
The inference device according to any one of claims 1 to 13, further comprising.
An inference method performed by a computer
The computer belongs to different domains of dynamically changing external information, including information about humans, using knowledge information provided by the knowledge base and inference rules provided by the rules database. It is a step of generating integrated information by combining information in a forward chain, and the information belonging to the different domains is information that can be expressed as a directed graph having nodes and edges, and the integrated information is the different domains. A step that is an integrated graph generated by combining the directed graphs that are information belonging to
A relationship database with a probability or page rank algorithm for the computer to reach the node when it randomly walks on the integrated graph and converges to a steady state with respect to the importance of the node which is a component of the integrated graph. A method of inference characterized by having a step of calculating using relationship data indicating the degree of relationship between nodes, which is a component of the integrated graph, provided by.
The process of acquiring knowledge information provided by the knowledge base, including information about humans, and inference rules provided by the rule database.
It is a process of generating integrated information by combining information belonging to different domains among dynamically changing external information in a forward chain using the knowledge information and the inference rule, which are different from each other. The information belonging to the domain is information that can be expressed as a directed graph having a node and an edge, and the integrated information is a process that is an integrated graph generated by combining the directed graphs that are information belonging to the different domains.
The importance of a node, which is a component of the integrated graph, is provided from the relationship database with an algorithm of probability or page rank of reaching the node when randomly walking on the integrated graph and converging to a steady state. , An inference program characterized by causing a computer to perform a process of calculating using relationship data indicating the degree of a relationship between nodes, which is a component of the integrated graph.