JP2016162276A - Explanation sentence creation device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To create an explanation sentence for making a human being understand determination process by a determination tree.SOLUTION: An explanation sentence creation device comprises: a determination tree acquisition unit for acquiring a determination tree comprising an inner node in which a determination condition of a feature amount and a branch destination node are associated, and a leaf node with which a regional class is associated; an explanation storage unit for storing an explanation term indicating the feature amount by a predetermined language, in association with the feature amount; and an explanation sentence creation unit for creating the explanation sentence of the determination tree on the basis of the explanation term associated with the feature amount used in the determination condition of the node.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an explanatory note generation device, an explanatory document creation method, and a program.

  Opportunities to use the judgment process based on machine learning technology by computer in real life are increasing. The determination process is optimized by a computer, and it is often difficult for a human to understand the determination process.

  A decision tree and a decision network (see, for example, Patent Document 1) are known as judgment processes that are relatively easy to understand by humans. A decision tree is a classification model of a tree structure consisting of an internal node associated with a feature value determination condition and a branch destination node, a leaf node associated with a belonging class, and a link connecting each node. . The decision network is a classification model having a graph structure that allows a plurality of data to be input. The decision network is a combination of a main decision tree that determines the class to which the target data to be classified belongs and a sub-decision tree that determines the criteria for determining the main decision tree. .

JP 2012-73750 A

However, although a decision tree is a relatively easy-to-understand decision process, it is still difficult to understand for users who are not familiar with information engineering.
An object of the present invention is to provide an explanatory note generation device, an explanatory document creation method, and a program for generating an explanatory note for allowing a human to understand a determination process based on a decision tree.

  According to the first aspect of the present invention, an explanatory note generating apparatus generates a decision tree having an internal node associated with a feature quantity determination condition and a branch destination node, and a leaf node associated with a belonging class. A decision tree acquisition unit to be acquired; an explanation storage unit for storing an explanatory word representing the feature amount in a predetermined language in association with the feature amount; and the explanatory word associated with the feature amount used in the determination condition of the node And an explanatory note generation unit for generating an explanatory note of the decision tree.

  According to the second aspect of the present invention, in the explanatory note generating apparatus according to the first aspect, the determination condition relates to a comparison between the feature amount and a predetermined threshold value, and the explanation storage unit , Storing a predicate representing a range defined by the threshold value in association with the threshold value, and the explanatory text generation unit uses the explanatory word associated with the feature quantity used for the determination condition of the node and the determination condition An explanatory text of the decision tree is generated based on the previous description word associated with the threshold value.

  According to the third aspect of the present invention, the explanatory note generating apparatus according to the first or second aspect further includes a node amount input unit that receives an input of the amount of a node to be expressed in the explanatory note, and the explanatory note generating unit The unit generates an explanatory note relating to the amount of nodes input to the node amount input unit.

  According to a fourth aspect of the present invention, in the explanatory note generating apparatus according to the third aspect, the explanatory note generating unit is configured such that the node amount input unit is in order from the smallest of the nodes of the decision tree. An explanatory note relating to the number of nodes input in the above is generated.

  According to a fifth aspect of the present invention, in the explanatory note generating apparatus according to the third aspect, the explanation storage unit stores a value indicating the easy-to-understand explanation word in association with the explanation word, and the explanation The sentence generation unit generates an explanatory sentence relating to the nodes of the amount input to the node amount input unit in descending order of the feature amount used for the determination condition among the nodes of the decision tree.

  According to the sixth aspect of the present invention, the explanatory note generating apparatus according to any one of the first to fifth aspects includes a target data input unit that receives input of target data to be classified by the decision tree; A classification unit that classifies the class to which the target data belongs based on the decision tree, and the description sentence generation unit includes an explanation sentence relating to a path that the classification unit has traced in the classification process of the decision tree. Generate.

  According to a seventh aspect of the present invention, in the explanatory note generating device according to any one of the first to fifth aspects, the explanatory note generating unit includes the decision tree among leaf nodes indicating one belonging class. An explanatory text relating to a node belonging to a route that reaches the largest number of arrivals of learning data used for learning is generated.

  According to an eighth aspect of the present invention, in the explanatory note generating device according to any one of the first to seventh aspects, the explanatory storage unit associates a plurality of feature amounts with each other and represents the combination And when the explanation sentence generation unit includes a node combination corresponding to the feature quantity combination among the nodes belonging to the path of the decision tree, the explanation word associated with the feature quantity combination. The description of the decision tree is generated based on

  According to the ninth aspect of the present invention, in the explanatory note generating device according to any one of the first to eighth aspects, the decision tree determines a affiliation class of target data to be classified by the decision tree. And a sub-decision tree that determines a range of the feature amount related to the determination condition of the main decision tree.

  According to a tenth aspect of the present invention, in the explanatory note generating apparatus according to the ninth aspect, the explanatory note generating unit generates an explanatory note of the main decision tree, and the nodes belong to the path of the main decision tree If the sub-decision tree includes those whose range of the feature amount is determined, an explanation of the sub-decision tree is further generated.

  According to an eleventh aspect of the present invention, in the explanatory document creating method, the explanatory note generating apparatus is configured such that the internal node associated with the feature amount determination condition and the branch destination node, and the leaf node associated with the belonging class. And a description storage unit that stores an explanatory word that represents the feature quantity in a predetermined language in association with the feature quantity, and is used as the determination condition of the node. The method includes a step of generating an explanatory text of the decision tree based on the explanatory word stored in association with a feature amount, and a step of creating a document including the explanatory text.

  According to the twelfth aspect of the present invention, a program causes a computer to execute a decision tree having an internal node associated with a feature value determination condition and a branch destination node, and a leaf node associated with a belonging class. A decision tree acquisition unit to acquire, an explanation storage unit for storing an explanatory word representing the feature quantity in a predetermined language in association with the feature quantity, based on the explanatory word associated with the feature quantity used for the determination condition of the node And functioning as an explanatory sentence generation unit for generating an explanatory sentence of the decision tree.

  According to at least one of the above aspects, the explanatory note generation device generates an explanatory sentence that expresses the determination condition of the node of the decision tree in a predetermined language. Thereby, the explanatory note generation device can make the user who can understand the language understand the determination process by the decision tree.

It is a schematic block diagram which shows the structure of the explanatory note production | generation apparatus which concerns on 1st Embodiment. It is a figure which shows the example of an explanatory word table. It is a figure which shows the example of a predicate table. It is a flowchart which shows operation | movement of the explanatory note generation apparatus which concerns on 1st Embodiment. It is a figure which shows an example of a decision tree. It is a schematic block diagram which shows the structure of the explanatory note production | generation apparatus which concerns on 2nd Embodiment. It is a flowchart which shows operation | movement of the explanatory note generation apparatus which concerns on 2nd Embodiment. It is a figure which shows the example of the explanatory word table which concerns on 3rd Embodiment. It is a figure which shows an example of a decision network. It is a schematic block diagram which shows the structure of the computer which concerns on at least 1 embodiment.

<< First Embodiment >>
Hereinafter, embodiments will be described in detail with reference to the drawings.
FIG. 1 is a schematic block diagram illustrating a configuration of an explanatory note generating apparatus according to the first embodiment.
The explanatory note generation device 100 receives an input of a decision tree and generates an explanatory note of a determination process based on the decision tree. Specifically, the explanatory note generation apparatus 100 generates, for each of a plurality of belonging classes classified by the decision tree, an explanatory note of a determination process for determining that the belonging class belongs. The decision tree is a classification model that classifies input data into one of a plurality of belonging classes. Specifically, the decision tree is a tree structure composed of internal nodes associated with feature quantity determination conditions and branch destination nodes, leaf nodes associated with belonging classes, and links connecting the nodes. Classification model.
The explanatory note generation apparatus 100 includes a decision tree acquisition unit 101, an explanation storage unit 102, a condition input unit 103, an explanatory sentence generation unit 104, and an output unit 105.

The decision tree acquisition unit 101 acquires a decision tree to be explained and metadata generated during learning of the decision tree. The metadata includes the passing rate or the passing number of learning data for each node of the decision tree.
The explanation storage unit 102 stores an explanation word table and a predicate table. In the explanatory word table, a feature quantity to be determined by the decision tree and a Japanese explanatory word representing the feature quantity are stored in association with each other.

FIG. 2 is a diagram illustrating an example of the explanatory word table.
According to the example shown in FIG. 2, the explanatory word table associates with the feature quantity “a value obtained by dividing the number of pixels of the hue around yellow (hue 30 degrees or more and less than 90 degrees) by the total number of pixels”. The explanatory word “is a sexy part” is stored. Also, in the explanatory word table, an explanatory word “a greenish part” is associated with the feature quantity “a value obtained by dividing the number of pixels of the hue around green (hue 90 degrees or more and less than 150 degrees) by the total number of pixels”. Is stored. In addition, the explanatory word table is associated with a feature amount “a value obtained by dividing the number of pixels in the hue around light blue (hue 150 degrees or more and less than 210 degrees) by the total number of pixels”, and an explanatory word “a light blue portion” Is stored. Also, in the explanatory word table, an explanatory word “a blueish part” is associated with a feature amount “a value obtained by dividing the number of pixels in the hue around blue (hue 210 degrees to less than 270 degrees) by the total number of pixels”. Is stored. Also, in the explanatory word table, an explanatory word “purple-like part” is associated with the feature quantity “a value obtained by dividing the number of pixels in the hue around purple (hue 270 degrees or more and less than 330 degrees) by the total number of pixels”. Is stored. In addition, the explanatory word table associates with the feature quantity “a value obtained by dividing the number of pixels of the hue around red (hue 330 degrees to less than 360 degrees and hue 0 degrees to less than 30 degrees) by the total number of pixels”. An explanatory word “is a reddish part” is stored. The explanatory word table stores an explanatory word “the number of colors” in association with the feature quantity “dispersion of hue”. The explanatory word table stores an explanatory word “brightness” in association with a feature quantity “average value of saturation”. The explanatory word table stores the explanatory word “brightness” in association with the feature value “average value of brightness”. The explanatory word table stores the explanatory word “vertical line” in association with the feature value “a value obtained by dividing the number of pixels having an edge of vertical line ± 20 degrees by the total number of pixels”. The explanatory word table stores an explanatory word “horizontal line” in association with a feature value “a value obtained by dividing the number of pixels having an edge of a horizontal line ± 20 degrees by the total number of pixels”. The explanatory word table stores an explanatory word “hatched” in association with a feature value “a value obtained by dividing the number of pixels having edges other than vertical and horizontal lines by the total number of pixels”.

In the predicate table, a Japanese predicate representing a range defined by the threshold value is stored in association with the threshold value of the feature value normalized to a value of 1 or less.
FIG. 3 is a diagram illustrating an example of a predicate table.
According to the example shown in FIG. 3, in the predicate table, a predicate “not so much” is stored as a predicate representing a range below the threshold in association with the threshold of “0.0 or more and less than 0.2”. A predicate “not so much” is stored as a predicate representing a range above the threshold. Also, in the predicate table, a predicate “less” is stored as a predicate representing a range less than the threshold in association with a threshold of “0.2 or more and less than 0.4”, and a predicate representing a range greater than or equal to the threshold. Stores the predicate “not less”. In addition, in the predicate table, a predicate “not to some extent” is stored as a predicate representing a range below the threshold in association with a threshold of “0.4 or more and less than 0.6”, and the range above the threshold Is stored as a predicate representing “a certain amount”. In addition, in the predicate table, a predicate “not too much” is stored as a predicate representing a range less than the threshold in association with the threshold of “0.6 or more and less than 0.8”, and a range greater than or equal to the threshold is stored. A predicate “many” is stored as a predicate to represent. In addition, in the predicate table, a predicate “not very much” is stored as a predicate representing a range less than the threshold in association with the threshold of “0.8 or more and 1.0 or less”, and a range greater than or equal to the threshold The predicate “very many” is stored as a predicate representing

  The condition input unit 103 accepts input of the level of detail of the explanatory text to be generated and the number of explanatory texts per affiliation class from the user. The level of detail of the explanatory text is a value indicating the level of detail of the determination process by the decision tree. Specifically, the level of detail represents the number of explanatory words and predicates included in the explanatory text. That is, the level of detail indicates the proportion of nodes that should be represented in the explanatory text among the nodes included in the route reaching the leaf node associated with the belonging class to be explained. The number of explanatory words and predicates included in the explanatory text increases as the level of detail increases. Therefore, as the level of detail increases, the accuracy of the explanatory text improves, but the ease of understanding the explanatory text decreases.

The description sentence generation unit 104 is classified according to the decision tree acquired by the decision tree acquisition unit 101 based on the level of detail and the number of sentences input to the condition input unit 103 and the information stored in the explanation storage unit 102. Generate a description for each class. Specifically, the explanatory note generation unit 104 determines by combining an explanatory word associated with a feature amount used as a determination condition for a classification node and a predicate associated with a range indicated by the determination condition. Generate a tree description.
The output unit 105 outputs data including the explanatory text generated by the explanatory text generation unit 104. The output unit 105 according to the present embodiment creates a document including an explanatory text by outputting document data including the explanatory text to a printer. Note that the output unit 105 according to another embodiment may output data including explanatory text to an external recording medium, a display, or another device.

Next, a description document creation method by the explanatory note generation apparatus 100 according to the present embodiment will be described.
FIG. 4 is a flowchart showing the operation of the explanatory note generating apparatus according to the first embodiment.
First, the decision tree acquisition unit 101 acquires a decision tree as a description generation target and metadata of the decision tree (step S1). Next, the condition input unit 103 receives an input of the level of detail of the explanatory text and the number of explanatory texts per affiliation class from the user (step S2).

  Next, the explanatory note generation unit 104 refers to the leaf node of the decision tree acquired by the decision tree acquisition unit 101, and identifies the belonging class classified by the decision tree (step S3). Next, the explanatory note generation unit 104 selects the identified belonging classes one by one, and executes the processes of steps S5 to S16 shown below for each selected belonging class (step S4).

  The description sentence generation unit 104 refers to the metadata of the decision tree acquired by the decision tree acquisition unit 101, and among the leaf nodes associated with the selected affiliation class, the leaf node with the highest learning data arrival rate or arrival number Is identified (step S5). When one or more explanatory texts of the class belonging to the class have already been generated, the explanatory text generation unit 104 has the highest learning data arrival rate or number of arrivals among the leaf nodes that are not used for generating the explanatory text. Identify high leaf nodes.

Next, the explanatory note generation unit 104 specifies a route that reaches the leaf node specified from the root node (step S6). Next, the explanatory note generation unit 104 specifies the number of internal nodes used for generation of the explanatory note based on the number of internal nodes included in the specified route and the level of detail of the input explanatory note (step S7). Specifically, the explanatory note generation unit 104 multiplies the number of internal nodes included in the route by a ratio representing the degree of detail, and rounds up the decimal point of the obtained value, thereby using the internal nodes used for generating the explanatory note Identify the number of
Next, the explanatory note generation unit 104 selects the number of internal nodes specified in step S7 from the internal nodes included in the route specified in step S6 in ascending order of depth, and for each selected node Then, the following processing of step S9 to step S13 is executed (step S8).

  The explanatory note generation unit 104 specifies a feature amount used for the determination condition associated with the selected internal node (step S9). The explanatory note generation unit 104 reads an explanatory word associated with the specified feature quantity from the explanatory word table stored in the explanatory storage unit 102 (step S10). Next, the explanatory note generation unit 104 specifies a threshold value used for the determination condition associated with the selected internal node (step S11). The description sentence generation unit 104 reads the predicate associated with the threshold value specified from the predicate table stored in the description storage unit 102 and the range belonging to the route specified in step S6 (whether it is equal to or greater than the threshold value) (step S12). Next, the explanatory note generation unit 104 generates an explanatory note of the internal node selected in step S8 by connecting the read explanatory words and predicates (step S13).

When the explanatory text generation unit 104 generates the explanatory texts of the number of internal nodes specified in step S7, the explanatory text generation unit 104 generates the explanatory text of the leaf node specified in step S5 by combining the generated explanatory texts (step S14). ).
The explanatory note generation unit 104 determines whether or not the number of explanatory sentences of the leaf node generated in step S14 has reached the number of sentences input in step S4 (step S15). If the number of explanatory texts has not reached the number of input texts (step S15: NO), the process returns to step S5 to generate an explanatory text for the leaf node with the next highest learning data arrival rate or number. When the number of explanatory texts reaches the number of input texts (step S15: YES), the explanatory text of the belonging class selected in step S4 is generated by combining the explanatory texts of the leaf nodes generated in step S14. (Step S16). At this time, the explanatory note generation unit 104 refers to the metadata of the decision tree acquired by the decision tree acquisition unit 101, and the explanatory note indicates the sum of the arrival rates of the learning data of each leaf node specified in step S5. The coverage rate may be added to the description. The number of sentences in the explanatory text is equal to the number of input texts, and the number of explanatory words and predicates included in the explanatory text depends on the input level of detail.

  When the explanatory note generation unit 104 generates explanatory notes of all belonging classes that can be classified by the decision tree, the output unit 105 generates document data including the explanatory note and outputs the document data to the printer (step) S17). Thereby, the explanatory note generation device 100 can create a document including the explanatory note of the decision tree.

Hereinafter, a specific example of the description sentence generation method according to the first embodiment will be described.
FIG. 5 is a diagram illustrating an example of a decision tree. The decision tree shown in FIG. 5 determines whether the input image data is image data showing a lotus or image data showing a sunflower. The decision tree is learned by using 50 images showing lotus and 50 images showing sunflower as learning data.
FIG. 5A is a diagram illustrating a connection relationship between nodes of a decision tree. FIG. 5B is a diagram illustrating the feature amount and threshold value of the determination condition associated with each internal node. FIG. 5C is a diagram showing the affiliation class associated with each leaf node and the number of learning data reaching the leaf node.

  In step S1, the decision tree acquisition unit 101 acquires the decision tree and metadata shown in FIG. Next, in step S2, the condition input unit 103 receives input of the degree of detail and the number of sentences. In this example, it is assumed that the level of detail is 50% and the number of sentences is 2. Next, the explanatory note generation unit 104 specifies a lotus class and a sunflower class as belonging classes that the decision tree can classify in step S3.

  In step S4, the explanatory note generation unit 104 first selects a lotus class. In step S5, the explanatory note generating unit 104 specifies the leaf node L1 (arrival number 25) having the highest arrival number among the leaf nodes L1, L2, L5, L6, and L7 of the lotus class. Next, the explanatory note generation unit 104 identifies a route that reaches the leaf node L1 in step S6. The internal node belonging to the path reaching the leaf node L1 is only the internal node M1 (root node). Therefore, the explanatory note generation unit 104 specifies that the number of internal nodes used for generation of the explanatory note is 1 in step S7.

  The explanatory note generation unit 104 outputs an explanatory word associated with the feature amount “value obtained by dividing the number of pixels of the hue around yellow (hue of 30 degrees or more and less than 90 degrees) by the total number of pixels” related to the determination condition of the internal node M1. , Read from the explanation storage unit 102. Referring to FIG. 2, the explanatory word associated with the feature amount “a value obtained by dividing the number of pixels in the hue around yellow (hue of 30 degrees or more and less than 90 degrees) by the total number of pixels” is “yellowish portion”. . The description sentence generation unit 104 reads the predicate associated with the range less than the threshold value 0.233 related to the determination condition of the internal node M1 from the description storage unit 102. Referring to FIG. 3, the number of predicates associated with the range less than the threshold value 0.233 is “few”. Thereby, the explanatory note generation unit 104 generates “there are few yellowish portions” as the explanatory note of the leaf node L1.

Next, the explanatory note generation unit 104 returns to step S5 and identifies the leaf node L5 (arrival number 10) having the next highest arrival number after the leaf node L1 among the leaf nodes L1, L2, L5, L6, and L7 of the lotus class. To do. Next, the explanatory note generation unit 104 specifies a route that reaches the leaf node L5 in step S6. Internal nodes belonging to the route reaching the leaf node L5 are internal nodes M1, M2, M4, and M5. Next, in step S7, the explanatory note generation unit 104 multiplies the number of internal nodes 4 belonging to the route by the detail level 50%, thereby specifying 2 as the number of internal nodes used for generating the explanatory text.
The explanatory note generation unit 104 generates explanatory notes of the two internal nodes M1 and M2 having the smallest depth belonging to the route reaching the leaf node L5. Thereby, the explanatory note generation unit 104 generates “not yellowish portions are not few and vividness is not so much” as the explanatory notes of the leaf node L5 in step S14.

As a result, in step S16, the explanatory note generation unit 104 determines that the description of the lotus class with a detail level of 50% and the number of sentences of 2 is “not a little yellowish part or little yellowish part, An explanation that generates “something that is not too much is considered a lotus class” is generated. At this time, referring to the metadata of the decision tree acquired by the decision tree acquisition unit 101, the sum of the arrival rates of the learning data of each leaf node specified in step S5 is used as the coverage rate of the explanation indicated by the explanation text of the lotus class. You may add to the said explanatory note. The number of learning nodes classified into the lotus class is 50, whereas the arrival number of the leaf node L1 is 25. Therefore, the arrival rate of the leaf node L1 is 50%. Since the arrival number of the leaf node L5 is 10, the arrival rate of the leaf node L5 is 20%. Therefore, the explanatory note generation unit 104 can calculate the coverage rate 70% of the explanatory sentence by adding these arrival rates.
By the same procedure, the explanatory note generation unit 104, in step S16, as the explanatory sentence of the sunflower class with a detail level of 50% and the number of sentences of 2, “the yellowish part is not small and the vividness is somewhat high, "The yellowish part is not small, the vividness is not so much, and the light blue part is not considered as a sunflower class".

  As described above, according to the present embodiment, the explanatory note generation apparatus 100 generates an explanatory note in which a determination condition of each belonging class that can be classified by the decision tree is expressed in a predetermined language. Thereby, the explanatory note generation device 100 can make a user who can understand the language understand the determination process based on the decision tree.

  In addition, although the explanatory note generation apparatus 100 which concerns on this embodiment produces | generates the explanatory note which concerns on all the affiliation classes, it is not restricted to this. For example, the explanatory note generation device 100 according to another embodiment may receive an input of a belonging class for which an explanatory note is to be generated from a user, and generate an explanatory note for the belonging class.

<< Second Embodiment >>
The second embodiment will be described in detail.
FIG. 6 is a schematic block diagram illustrating a configuration of the explanatory note generation apparatus according to the second embodiment.
The explanatory note generating apparatus 100 according to the second embodiment receives an input of target data to be classified from a user, and an explanatory text of a determination process until the target data is classified into a certain belonging class by a decision tree Is generated.
The explanatory note generation device 100 according to the second embodiment further includes a target data input unit 106 and a classification unit 107 in addition to the configuration of the first embodiment. The explanatory note generation apparatus 100 according to the second embodiment differs from the first embodiment in the operation of the explanatory note generation unit 104.

The target data input unit 106 receives input of target data to be classified by the decision tree acquired by the decision tree acquisition unit 101.
The classification unit 107 classifies the affiliation class of the input target data based on the decision tree acquired by the decision tree acquisition unit 101.
The explanatory note generation unit 104 generates an explanatory note relating to a route followed by the classification unit 107 during the classification process, out of the decision trees acquired by the decision tree acquisition unit 101.

Next, a description document creation method by the explanatory note generation apparatus 100 according to the present embodiment will be described.
FIG. 7 is a flowchart showing the operation of the explanatory note generating apparatus according to the second embodiment.
First, the decision tree acquisition unit 101 acquires a decision tree for which a description is to be generated and metadata of the decision tree (step S101). Next, the target data input unit 106 receives input of target data from the user (step S102). Next, the classification unit 107 classifies the class to which the input target data belongs based on the decision tree acquired by the decision tree acquisition unit (step S103). The classification unit 107 specifies the route from the root node to the leaf node traced in the classification process (step S104). Next, the condition input unit 103 receives an input of the detail level of the explanatory text from the user (step S105). Since the number of paths to be traced in the process of classifying the target data is 1, it is not necessary to input the number of documents.

Next, the explanatory note generation unit 104 specifies the number of internal nodes used for generation of the explanatory note based on the number of internal nodes included in the route specified in step S103 and the degree of detail of the input explanatory note. (Step S106). Specifically, the explanatory note generation unit 104 multiplies the number of internal nodes included in the route by a ratio representing the degree of detail, and rounds up the decimal point of the obtained value, thereby using the internal nodes used for generating the explanatory note Identify the number of
Next, the explanatory note generation unit 104 selects the number of internal nodes specified in step S106 from the internal nodes included in the route specified in step S103, in ascending order of depth, and for each selected node Then, the following processing of step S108 to step S112 is executed (step S107).

  The description sentence generation unit 104 specifies the feature amount used for the determination condition associated with the selected internal node (step S108). The explanatory note generation unit 104 reads an explanatory word associated with the identified feature quantity from the explanatory word table stored in the explanatory storage unit 102 (step S109). Next, the explanatory note generation unit 104 specifies a threshold value used for the determination condition associated with the selected internal node (step S110). The description sentence generation unit 104 reads out the predicate associated with the threshold value specified from the predicate table stored in the description storage unit 102 and the range belonging to the route specified in step S104 (step S111). Next, the explanatory note generation unit 104 generates an explanatory note of the internal node selected in step S8 by connecting the read explanatory words and predicates (step S112).

  When the explanatory text generation unit 104 generates the explanatory texts of the number of internal nodes specified in step S106, the explanatory text generation unit 104 generates the explanatory text of the route specified in step S104 by combining the generated explanatory texts (step S113). . Then, the output unit 105 generates document data including the explanatory text and outputs the document data to the printer (step S114). Thereby, the explanatory note generation device 100 can create a document including an explanatory note of the decision tree related to the classification of the target data.

  As described above, according to the present embodiment, the explanatory note generation apparatus 100 generates an explanatory note representing the reason why the input target data is classified into the belonging class indicated in the classification result. Thereby, the explanatory note generation apparatus 100 can make a user who can understand the language understand the determination process related to the classification of the target data by the decision tree.

<< Third Embodiment >>
Among the feature quantities to be determined by the decision tree, there are those that are easy to explain in a predetermined language and those that are difficult to explain in a predetermined language. For example, the feature quantities such as the hue dispersion, the average value of saturation, and the average value of brightness exemplified in FIG. 2 are relatively easy to explain in Japanese. On the other hand, for example, the feature values such as the average value of hue, the HOG feature value, and the SIFT feature value are difficult to explain in Japanese.
The explanatory note generation device 100 according to the third embodiment generates an explanatory note that suppresses expressions that are difficult for humans to understand.

  The explanatory note generating apparatus 100 according to the third embodiment is different from the first embodiment in the configuration of the explanatory word table stored in the explanatory storage unit 102. The explanatory note generating apparatus 100 according to the third embodiment is different from the first exemplary embodiment in the operation of the explanatory note generating unit 104.

The explanatory word table stored in the explanatory storage unit 102 stores, in association with the feature quantity, an explanatory word that represents the feature quantity and an explanatory aptitude value that indicates the ease of understanding the explanatory word. The explanatory suitability value indicates that the higher the value, the easier to understand the explanatory word. For example, the explanatory suitability value is set by an administrator who associates a feature quantity with an explanatory word.
FIG. 8 is a diagram illustrating an example of an explanatory word table according to the third embodiment.
According to the example shown in FIG. 2, the explanatory word table stores the explanatory word “the number of colors” and the explanatory aptitude value “0.7” in association with the feature quantity “hue dispersion”. The explanatory word table stores an explanatory word “brightness” and an explanatory suitability value “0.9” in association with the feature value “average value of saturation”. The explanatory word table stores an explanatory word “brightness” and an explanatory aptitude value “0.9” in association with the feature value “average value of brightness”. The explanatory word table stores an explanatory word “average hue value” and an explanatory suitability value “0.1” in association with the feature value “average hue value”. In addition, in the explanatory word table, an explanatory word “the frequency of change in brightness in the vertical direction” and an explanatory aptitude value “0.3” are associated with the feature value “average value of spatial frequency on the vertical axis”. Stored. The explanatory word table has an explanatory word “the frequency of change in brightness in the vertical direction” and an explanatory aptitude value “0.3” in association with the feature quantity “average value of spatial frequency on the horizontal axis”. Stored.
In this way, explanatory words that are easy to understand for human beings such as “brightness” and “brightness” are associated with high explanatory aptitude values, “average hue value”, “vertical brightness” An explanatory word that is difficult for humans to understand, such as “the frequency of change” is associated with a low explanatory aptitude value.

The description sentence generation unit 104 describes the number of nodes according to the level of detail input to the condition input unit 103 in descending order of the explanatory value of the feature amount used for the determination condition among the nodes of the decision tree. Generate a statement.
Thereby, the explanatory note generation device 100 can generate an explanatory note in which expressions difficult to understand for humans are suppressed.

<< Fourth Embodiment >>
Among the feature quantities to be determined by the decision tree, there are those that can explain a combination of a plurality of feature quantities with one explanatory word. For example, the R value of the RGB value can be described by an explanatory word “red component is”, the G value is “green component”, and the combination of the R value and the G value is an explanatory word “yellow component is”. Can also be explained.
The explanatory note generation device 100 according to the fourth embodiment generates an explanatory note that explains a combination of a plurality of feature amounts with one explanatory word.

The explanatory note generating apparatus 100 according to the fourth embodiment is different from the first embodiment in the configuration of the explanatory word table stored in the explanatory storage unit 102. The explanatory note generating apparatus 100 according to the fourth embodiment is different from the first embodiment in the operation of the explanation storage unit 102.
The explanatory word table stored in the explanatory storage unit 102 stores an explanatory word representing the feature quantity or the combination of the feature quantities in association with the feature quantity and the combination of the feature quantities.
When the node belonging to the path of the decision tree includes a node combination corresponding to the feature amount combination, the explanatory note generation unit 104 determines the decision tree based on the explanatory word associated with the feature amount combination. Generate descriptive text. Specifically, a combination including the feature amount specified in step S9 shown in FIG. 4 is stored in the explanatory word table, and an intermediate node related to another feature amount related to the combination is on the route specified in step S6. In the case where it exists, the explanatory note generating unit 104 reads the explanatory word associated with the combination in step S10.
In this way, the explanatory note generation device 100 generates the explanatory text to be explained that is explained by one explanatory word by combining a plurality of feature quantity combinations, so that the details of the explanatory text can be changed without changing the length of the explanatory text. The degree can be increased.

<< Fifth Embodiment >>
The explanatory note generation device 100 according to the fifth embodiment generates an explanatory note of the decision network according to Patent Document 1. The decision network is a DAG (Directed acyclic graph) classification model that allows a plurality of data to be input. The decision network is a combination of a main decision tree that determines the class to which the target data to be classified belongs and a sub-decision tree that determines the criteria for determining the main decision tree. . The main decision tree and the sub-decision tree share some nodes.

FIG. 9 is a diagram illustrating an example of a decision network.
The determination network shown in FIG. 9 determines whether the input image data is image data showing a leather product with a flaw or image data showing a leather product without a flaw.
FIG. 9A is a diagram illustrating a connection relationship of nodes in the decision network. FIG. 9B is a diagram illustrating the feature amount and threshold value of the determination condition associated with each internal node. FIG. 9C is a diagram showing the affiliation class associated with each leaf node and the number of learning data reaching the leaf node.

According to FIG. 9A, a link directed to the intermediate node M12 and a link directed to the leaf node L12 are connected to the intermediate node M11. The intermediate node M12 is connected to a link directed to the leaf node M11 and a link directed to the leaf node L12. A link directed to the intermediate node M11 and a link directed to the leaf node L12 are connected to the intermediate node M13. Target data that is image data to be classified is input to the intermediate node M11, and reference data that is image data in the vicinity of the target data is input to the intermediate node M13. The decision network includes a main decision tree including intermediate nodes M11 and M12 and leaf nodes L11 and L12 having an intermediate node M11 to which target data is input as a root node, and an intermediate node M11 having an intermediate node M13 as a root node. , M12 and M13 and a sub-decision tree consisting of leaf nodes L11 and L12. That is, the target data is determined by the main decision tree. The reference data is determined by the sub-decision tree.
Note that the main decision tree and the sub-decision tree include the intermediate nodes M11 and M12 in common. The range of the feature amount indicated by the intermediate node of the main decision tree varies depending on whether or not the reference data determination path includes the intermediate node. Specifically, when the intermediate node M11 is not included in the reference data determination path, the threshold value of the median range indicated by the intermediate node M11 of the main decision tree is 0.547. On the other hand, when the intermediate node M11 is included in the reference data determination path, the threshold value of the median range indicated by the intermediate node M11 of the main decision tree is 0.508 (the value in parentheses in FIG. 9B). Become.

The explanatory note generating apparatus 100 according to the fifth embodiment differs from the second exemplary embodiment in the operation of the explanatory note generating unit 104. The explanatory note generation unit 104 generates an explanatory note of the main decision tree based on the path of the main decision tree specified by the classification unit 107, as in the second embodiment. When the common intermediate node is included in the route of the main decision tree and the route of the sub decision tree specified by the classification unit 107, the explanation sentence generating device 100 further generates an explanation of the sub decision tree. The explanatory text of the sub-decision tree is an explanatory text about the route from the root node to the common intermediate node at the maximum.
That is, when the nodes belonging to the path of the main decision tree include nodes whose feature amount range is determined by the sub decision tree, the explanation sentence generation unit 104 further generates an explanation sentence of the sub decision tree. .
Thereby, the explanatory note generation device 100 can appropriately create an explanatory note of the decision network including a plurality of decision trees.

As described above, the embodiment has been described in detail with reference to the drawings. However, the specific configuration is not limited to that described above, and various design changes and the like can be made.
For example, in the above-described embodiment, the explanation sentence generating apparatus 100 has explained the case where the explanation text of the general decision tree and the decision network is created, but the present invention is not limited to this. For example, in another embodiment, the explanatory note generation device 100 may generate an explanatory note of a fuzzy decision tree.

  Further, in the above-described embodiment, the case where the explanatory note generation apparatus 100 generates a Japanese explanatory note has been described, but the present invention is not limited to this. For example, in another embodiment, the explanatory note generation device 100 may generate an explanatory note in English, Chinese, Russian, or another language.

  Further, in the above-described embodiment, the description sentence generating apparatus 100 has been described as receiving the details and the number of sentences. However, the present invention is not limited to this. For example, in another embodiment, the explanatory note generation apparatus 100 may create an explanatory note with a predetermined level of detail and the number of sentences. In another embodiment, the explanatory note generation apparatus 100 may accept an input of the number of nodes to be explained instead of the level of detail. In another embodiment, the explanatory note generation apparatus 100 may accept an input of a necessary coverage rate instead of the number of sentences. In this case, the explanatory note generating apparatus 100 repeatedly executes the loop of step S8 to step S13 until the input coverage rate is satisfied.

FIG. 10 is a schematic block diagram illustrating a configuration of a computer according to at least one embodiment.
The computer 901 includes a CPU 902, a main storage device 903, an auxiliary storage device 904, and an interface 905.
The above described explanatory note generation apparatus 100 is mounted on a computer 901. The operation of each processing unit described above is stored in the auxiliary storage device 904 in the form of a program. The CPU 902 reads the program from the auxiliary storage device 904, develops it in the main storage device 903, and executes the above processing according to the program. Further, the CPU 902 secures a storage area corresponding to each of the above-described storage units in the main storage device 903 according to the program.

  In at least one embodiment, the auxiliary storage device 904 is an example of a tangible medium that is not temporary. Other examples of the non-temporary tangible medium include a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, and a semiconductor memory connected via the interface 905. When this program is distributed to the computer 901 via a communication line, the computer 901 that has received the distribution may develop the program in the main storage device 903 and execute the above processing.

  The program may be for realizing a part of the functions described above. Further, the program may be a so-called difference file (difference program) that realizes the above-described function in combination with another program already stored in the auxiliary storage device 904.

DESCRIPTION OF SYMBOLS 100 Explanation sentence production | generation apparatus 101 Decision tree acquisition part 102 Explanation memory | storage part 103 Condition input part 104 Explanation sentence generation part 105 Output part 106 Target data input part 107 Classification part

Claims (12)

A decision tree acquisition unit for acquiring a decision tree having an internal node associated with a determination criterion for a feature quantity and a branch destination node, and a leaf node associated with a belonging class;
An explanation storage unit for storing an explanatory word representing the feature quantity in a predetermined language in association with the feature quantity;
An explanatory note generation device comprising: an explanatory note generation unit that generates an explanatory note of the decision tree based on the explanatory words associated with the feature amount used for the determination condition of the node. The determination condition relates to a comparison between the feature amount and a predetermined threshold value,
The explanation storage unit stores a predicate representing a range defined by the threshold in association with the threshold;
Based on the explanatory word associated with the feature amount used in the determination condition of the node and the previous descriptive word associated with the threshold used by the determination condition, the explanatory sentence generation unit describes the explanatory text of the decision tree. The explanatory note generation device according to claim 1. A node amount input unit that receives an input of a node amount to be represented in the explanatory text;
The explanatory note generation device according to claim 1, wherein the explanatory note generation unit generates an explanatory note relating to an amount of nodes input to the node amount input unit. The explanatory note generation unit according to claim 3, wherein the explanatory note generation unit generates an explanatory note relating to nodes of an amount input to the node amount input unit in order from a node having the smallest depth among the nodes of the decision tree. apparatus. The explanation storage unit stores a value indicating the easy-to-understand of the explanation word in association with the explanation word,
The description sentence generation unit generates an explanation sentence related to nodes of the amount input to the node amount input unit in descending order of the characteristic amount used for the determination condition among the nodes of the decision tree. 3. The explanatory note generating device according to 3. A target data input unit for receiving input of target data to be classified by the decision tree;
A classification unit that classifies a class to which the target data belongs based on the decision tree,
The explanatory note generation device according to any one of claims 1 to 5, wherein the explanatory note generation unit generates an explanatory note related to a path traced in the classification process by the classification unit in the decision tree. The explanation sentence generation unit generates an explanation sentence relating to a node belonging to a route that reaches the largest number of arrivals of learning data used for learning of the decision tree among leaf nodes indicating one belonging class. The explanatory note generation device according to any one of claims 1 to 5. The explanation storage unit stores an explanation word representing the combination in association with a combination of a plurality of feature amounts;
When the explanation sentence generation unit includes a combination of nodes corresponding to the combination of feature quantities among the nodes belonging to the path of the decision tree, based on the explanation word associated with the combination of feature quantities The explanatory note generation device according to any one of claims 1 to 7, wherein an explanatory note of a decision tree is generated. The decision tree includes: a main decision tree that determines a affiliation class of target data to be classified by the decision tree; and a sub-decision tree that determines a range of the feature amount related to the determination condition of the main decision tree. The explanatory note generation device according to any one of claims 1 to 8. When the explanatory note generation unit generates an explanatory note of the main decision tree, and nodes belonging to the path of the main decision tree include those whose range of the feature amount is determined by the sub-determination tree. Furthermore, the explanatory note production | generation apparatus of Claim 9 which produces | generates the explanatory note of the said sub decision tree. The explanatory note generating device acquires a decision tree having an internal node associated with a determination criterion for a feature amount and a branch destination node, and a leaf node associated with a belonging class;
The explanation storage unit that stores an explanatory word representing the feature quantity in a predetermined language in association with the feature quantity is stored in the explanation word stored in association with the feature quantity used for the determination condition of the node. Generating a description of the decision tree based on:
An explanatory document creating method comprising: creating a document including the explanatory text. Computer
A decision tree acquisition unit for acquiring a decision tree having an internal node associated with a determination criterion for a feature amount and a branch destination node, and a leaf node associated with a belonging class;
An explanation storage unit for storing an explanatory word representing the feature quantity in a predetermined language in association with the feature quantity;
The program for functioning as an explanatory note generation part which produces | generates the explanatory note of the said decision tree based on the said explanatory word linked | related with the feature-value used for the determination conditions of the said node.

Images