Computer system and method for generating interpretation sentence

Abstract

To generate an interpretation sentence for an inference result of an inference model.SOLUTION: A computer system stores a large-scale language model that accepts an instruction sentence as input and outputs an interpretation sentence for interpreting a result of an inference, and text template information that stores template data in which a characteristic of a contribution value of a feature in a group having features is associated with a template of the instruction sentence, calculates the contribution value of each of a plurality of features, generates a plurality of groups each constituted of one or more features, acquires, for each of the plurality of groups, the template data by referring to the text template information based on the characteristic of the contribution value of the feature included in the group, generates, based on the template data and the feature included in the group, the instruction sentence to be input to the large-scale language model, and outputs the interpretation sentence obtained by inputting the instruction sentence into the large-scale language model.SELECTED DRAWING: Figure 2B

Description

[Technical field]

[0001] The present disclosure relates to a computer system and method for generating an explanation for an inference result of a machine learning model. [Background technology]

[0002] As a background technique of the present disclosure, for example, JP 2020-123164 A (Patent Document 1) is known. Patent Document 1 discloses a system that presents data useful for a user to interpret the basis of a prediction result output by an AI. For example, it is disclosed that "a computer system holds interpretation factor conversion information for managing interpretation factors that are determined by the values ​​of each of a plurality of feature quantities contained in input data including values ​​of a plurality of feature quantities and the contribution values ​​of the plurality of feature quantities contained in the input data and are used to interpret the basis of a prediction result for the input data, and when evaluation target data is input, calculates a prediction result for the evaluation target data, calculates the contribution value of each of the plurality of feature quantities contained in the evaluation target data, and identifies the corresponding interpretation factor by referring to the interpretation factor conversion information based on the values ​​and contribution values ​​of each of the plurality of feature quantities contained in the evaluation target data, and generates and outputs display information for presenting the identified interpretation factor." (See, for example, Abstract). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-123164 A Summary of the Invention [Problem to be solved by the invention]

[0004] In order to generate the interpretation factor conversion information of Patent Document 1, the knowledge of an expert is indispensable, and there is a problem in that it cannot be used when the expert cannot identify the interpretation factor.

[0005] The present invention provides a system and method for presenting business interpretations without requiring expert knowledge by using a language model. [Means for solving the problem]

[0006] A representative example of the invention disclosed in the present application is as follows: That is, a computer system holds a large-scale language model that receives input data including a plurality of feature quantities, is connected to an inference system that performs inference using an inference model, receives an instruction sentence as an input, and outputs an interpretation sentence for interpreting the result of the inference, and holds first text template information that stores first template data that associates a characteristic of a contribution value indicating a magnitude of a contribution of the feature quantity to the result of the inference in a group consisting of one or more feature quantities with a template of the instruction sentence, calculates the contribution value of each of the plurality of feature quantities using the input data, the result of the inference, and the inference model, generates a plurality of groups consisting of one or more feature quantities, acquires the first template data for each of the plurality of groups by referring to the first text template information based on the characteristic of the contribution value of the feature quantity included in the group, generates the instruction sentence to be input to the large-scale language model based on the acquired first template data and the feature quantity included in the group, and outputs the interpretation sentence obtained by inputting the instruction sentence to the large-scale language model. Effect of the Invention

[0007] According to one aspect of the present disclosure, a computer system can present an interpretation for an inference result of an inference model. Problems, configurations, and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 illustrates an example of a hardware configuration of a computer system according to a first embodiment. [Figure 2A] FIG. 2 is a diagram illustrating an example of a software configuration of the inference device of the first embodiment. [Figure 2B] FIG. 2 is a diagram illustrating an example of a software configuration of the interpretation sentence generation device of the first embodiment. [Diagram 3] FIG. 4 is a diagram illustrating an example of a data structure of input data according to the first embodiment. [Figure 4] FIG. 11 is a diagram illustrating an example of a data structure of contribution value information according to the first embodiment. [Diagram 5] FIG. 4 is a diagram illustrating an example of a data structure of feature amount information according to the first embodiment. [Figure 6] FIG. 4 is a diagram illustrating an example of a data structure of a feature amount list according to the first embodiment. [Figure 7] FIG. 11 is a diagram illustrating an example of a data structure of group information according to the first embodiment. [Figure 8A] FIG. 11 is a diagram illustrating an example of a data structure of first text template information according to the first embodiment. [Figure 8B] FIG. 11 is a diagram illustrating an example of a data structure of first text template information according to the first embodiment. [Figure 9] FIG. 11 is a diagram illustrating an example of a data structure of prompt information according to the first embodiment. [Figure 10] FIG. 13 is a diagram illustrating an example of a data structure of analysis information according to the first embodiment. [Figure 11] FIG. 11 is a diagram illustrating an example of a data structure of second text template information according to the first embodiment. [Figure 12] FIG. 4 is a diagram showing an example of a data structure of related document information in the first embodiment. [Figure 13] FIG. 11 is a diagram illustrating an example of a data structure of a table for managing interpretation sentences in the first embodiment. [Figure 14] 1 is a flowchart illustrating an overview of a process executed by an interpretation sentence generation device according to a first embodiment. [Figure 15] 11 is a flowchart illustrating an example of a feature selection process executed by the interpretation sentence generation device according to the first embodiment. [Figure 16] 11 is a flowchart illustrating an example of a prompt generation process executed by the interpretation sentence generation device according to the first embodiment. [Figure 17]11 is a flowchart illustrating an example of a relationship document generation process executed by the interpretation sentence generation device according to the first embodiment. [Figure 18] 11 is a flowchart illustrating an example of a filtering process executed by the interpretation sentence generation device according to the first embodiment. [Figure 19] FIG. 4 is a diagram illustrating an example of a screen displayed by a feature amount presenting unit according to the first embodiment. [Figure 20] FIG. 11 is a diagram illustrating an example of a screen displayed by a prompt presenting unit according to the first embodiment. [Figure 21] FIG. 13 is a diagram showing an example of a screen displayed by an analysis result presenting unit according to the first embodiment. [Figure 22] FIG. 13 is a diagram showing an example of a screen displayed by a result presentation unit according to the first embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, the embodiments of the present invention will be described in detail with reference to the drawings. When necessary for convenience, the description will be divided into a number of sections or embodiments, but unless otherwise specified, they are not unrelated to each other, and one is related to the other as a partial or complete modification, detail, supplementary explanation, etc. Furthermore, in the following, when the number of elements (including the number, numerical value, amount, range, etc.) is mentioned, it is not limited to the specific number, and may be more than or less than the specific number, unless otherwise specified or when it is clearly limited to a specific number in principle.

[0010] The system of an embodiment of this specification may be a physical computer system (one or more physical computers) or a system built on a group of computing resources (multiple computing resources) such as a cloud infrastructure. The computer system or the group of computing resources may include one or more interface devices (including, for example, a communication device and an input / output device), one or more storage devices (including, for example, a memory (main memory) and an auxiliary storage device), and one or more arithmetic devices.

[0011] When a function is realized by a computing device executing a program including instruction codes, the defined processing is performed using a storage device and / or an interface device, etc., as appropriate, and therefore the function may be considered to be at least a part of the computing device. Processing described with the function as the subject may be processing performed by the computing device or a system having the computing device. The program may be installed from a program source.

[0012] The program source may be, for example, a program distribution computer or a computer-readable storage medium (for example, a computer-readable non-transitory storage medium). The description of each function is an example, and multiple functions may be combined into one function, or one function may be divided into multiple functions. EXAMPLES

[0013] FIG. 1 is a diagram illustrating an example of a hardware configuration of a computer system according to a first embodiment.

[0014] The computer system shown in Fig. 1 includes an inference device 110, an interpretation sentence generation device 100, and a user terminal 120. The devices are connected to each other via a network 130. The type of network 130 is arbitrary, and may be, for example, a wide area network (WAN) or a local area network (LAN). The connection method of the network 130 may be either wired or wireless.

[0015] The inference device 110 performs various inferences using an inference model. In this embodiment, an inference model for predicting the probability of a fire occurring in a property will be described as an example.

[0016] The inference device 110 includes, as hardware components, a CPU 111, a memory 112, an auxiliary storage device 113, and a network interface 114. The hardware elements communicate with each other via an internal bus.

[0017] The CPU 111 executes a program stored in the memory 112. The memory 112 stores the program executed by the CPU 111 and information necessary for the program. The memory 112 also includes a work area that is temporarily used by the program.

[0018] The auxiliary storage device 113 permanently stores data. The auxiliary storage device 113 may be a storage medium such as a hard disk drive (HDD) or a solid state drive (SSD), or a non-volatile memory. The programs and information stored in the memory 112 may be stored in the auxiliary storage device 113. In this case, the CPU 111 reads out the programs and information from the auxiliary storage device 113, loads the programs and information into the memory 112, and executes the programs loaded into the memory 112. The network interface 114 connects to other devices via a network.

[0019] The interpretation sentence generating device 100 generates an interpretation sentence reflecting knowledge of the business to which the inference is applied from the inference result, and presents the interpretation sentence to the user who uses the inference device 110.

[0020] The interpretation sentence generating device 100 has, as hardware, a CPU 101, a memory 102, an auxiliary storage device 103, and a network interface 104. The hardware elements communicate with each other via an internal bus or the like. The CPU 101, the memory 102, the auxiliary storage device 103, and the network interface 104 are the same hardware elements as the CPU 111, the memory 112, the auxiliary storage device 113, and the network interface 114.

[0021] The user terminal 120 is a terminal used by a user. The user terminal 120 accepts a user input for generating an interpretation of an inference result, and presents the generated interpretation to the user. The user terminal 120 has, as hardware, a CPU 121, a memory 122, an auxiliary storage device 123, a network interface 124, an input device 125, and an output device 126. The hardware elements communicate via an internal bus.

[0022] The CPU 121 , the memory 122 , the auxiliary storage device 123 , and the network interface 124 are hardware elements similar to the CPU 111 , the memory 112 , the auxiliary storage device 113 , and the network interface 114 .

[0023] The input device 125 is a device for inputting data, etc., and includes a keyboard, a mouse, a touch panel, etc. The output device 126 is a device for outputting data, etc., and includes a display, a touch panel, etc.

[0024] In the above device, the CPU operates as a functional unit having a predetermined function by executing processing according to a program. In the following description, when processing is described with the program as the subject, it means that the CPU or the device in which the CPU is implemented is executing the program that realizes the functional unit.

[0025] In the example configuration of Figure 1, different computers execute the tasks of operational management, decision support, and user interface, respectively. In other examples, a combination of all or part of these tasks may be executed by a single computer. For example, the inference device 110 and the interpretation sentence generation device 100 may be realized as a virtual computer running on a single computer.

[0026] As discussed above, a computing system may be comprised of one or more computers including one or more computing devices and one or more storage devices including non-transitory storage media. Memory, secondary storage devices, or a combination thereof are storage devices. A CPU is an example of a computing device. A computing device may be comprised of a single processing unit or multiple processing units and may include single or multiple computing units, or multiple processing cores. A computing device may be implemented as one or more central processing units, microprocessors, microcomputers, microcontrollers, digital signal processors, state machines, logic circuits, graphic processing units, systems on a chip, and / or any device that manipulates signals based on control instructions.

[0027] FIG. 2A is a diagram illustrating an example of a software configuration of the inference device 110 according to the first embodiment.

[0028] The inference device 110 includes a prediction unit 201 that performs prediction using an inference model. The prediction unit 201 executes processing using the inference model on input data 202, and outputs a predicted value 203.

[0029] FIG. 2B is a diagram illustrating an example of a software configuration of the interpretation sentence generation device 100 according to the first embodiment.

[0030] The interpretation sentence generation device 100 includes an XAI (eXplainable Artificial Intelligence) unit 211, a selection unit 212, a feature presentation unit 213, a group generation unit 221, a prompt generation unit 222, a prompt presentation unit 223, an analysis unit 231, an analysis result presentation unit 232, a relationship document generation unit 233, an interpretation sentence generation unit 241, a filtering unit 242, and a result presentation unit 243.

[0031] The XAI unit 211 acquires the input data 202 and the predicted value 203 from the inference device 110, and calculates the contribution value of each feature of the input data 202 to the predicted value 203. The contribution value of each feature is output as contribution value information 214.

[0032] The XAI unit 211 calculates the contribution value using a known algorithm, such as Shapley Additive Explanation (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), Integrated gradient, or the like.

[0033] The selection unit 212 acquires contribution value information 214 from the XAI unit 211, and refers to feature amount information 215 to verbalize the features and select features to be used for generating an interpretation sentence based on the contribution value of each feature amount. The selected verbalized features are output as a feature amount list 216. In the following description, the verbalized features are referred to as verbalized features.

[0034] The feature presenting unit 213 displays the verbalized feature selected by the selection unit 212 on the screen of the user terminal 120.

[0035] The group generation unit 221 acquires the feature list 216 from the selection unit 212, and generates a plurality of groups having the verbalized features as elements. A group of verbalized features is composed of at least one verbalized feature. The generated groups are output as group information 224.

[0036] The prompt generation unit 222 acquires group information 224 from the group generation unit 221, and generates a prompt for each group by referring to first text template information 225. Here, the prompt is an instruction sentence to be input to a large-scale language model, which will be described later. The prompt generation unit 222 outputs prompt information 226 that stores the prompt for each group.

[0037] The prompt presentation unit 223 displays the prompt generated by the prompt generation unit 222 on the screen of the user terminal 120 .

[0038] The analysis unit 231 obtains the feature amount list 216 from the selection unit 212, and analyzes the relationship between the feature amount value and the contribution value.

[0039] In this embodiment, the analysis unit 231 uses the XAI unit 211 to generate, as an analysis result, a graph showing the relationship between the value of the selected feature amount and the contribution value to the predicted value (fire probability).

[0040] The analysis result presentation unit 232 displays the analysis results of the analysis unit 231 on the screen of the user terminal 120. The user checks the screen and, if necessary, inputs information regarding the relationship using the user terminal 120. The analysis unit 231 outputs the information input via the user terminal 120 and the analysis results as analysis information 234.

[0041] The relationship document generation unit 233 acquires the analysis information 234 from the analysis unit 231, and generates one or more relationship documents by verbalizing the relationships included in the analysis information 234 with reference to the second text template information 235. The relationship document generation unit 233 outputs relationship document information 236 that stores one or more relationship documents.

[0042] The interpretation sentence generation unit 241 acquires the prompt information 226 from the prompt generation unit 222, and generates an interpretation sentence by inputting the prompt to a large-scale language model. For example, the large-scale language model may be GPT-4 (Generative pre-trained transformer-4), PaLM (Pathways Language Model), or the like. Note that the present invention is not limited to the type of the large-scale language model.

[0043] The filtering unit 242 acquires the relationship document information 236 from the relationship document generating unit 233, and excludes from the interpretation sentence 244 to be output, an interpretation sentence 244 that includes a fact that contradicts the fact expressed by the relationship document.

[0044] The result presentation unit 243 displays on the screen the interpretation sentences 244 that have not been filtered out by the filtering unit 242 .

[0045] The data structure of the data and information handled by the computer system will be described below.

[0046] FIG. 3 is a diagram showing an example of the data structure of the input data 202 according to the first embodiment.

[0047] The input data 202 includes a plurality of entries each consisting of a feature ID 301 and a value 302. One entry corresponds to one feature.

[0048] Feature ID 301 is a field that stores the ID of a feature. Value 302 is a field that stores the value of a feature. In predicting the probability of fire, the input data 202 is, for example, property data, and includes the age of the building, construction method, etc. as features.

[0049] FIG. 4 is a diagram illustrating an example of a data structure of the contribution value information 214 according to the first embodiment.

[0050] The contribution value information 214 includes a plurality of entries each consisting of a feature amount ID 401, a value 402, and a contribution value 403. One entry corresponds to one feature amount.

[0051] The feature amount ID 401 and value 402 are the same fields as the feature amount ID 301 and value 302. The contribution value 403 is a field that stores the contribution value.

[0052] FIG. 5 is a diagram illustrating an example of a data structure of the feature amount information 215 according to the first embodiment.

[0053] The feature amount information 215 stores entries each consisting of a feature amount ID 501 , a feature amount name 502 , a value 503 , and a languageization template 504 .

[0054] Feature ID 501 is the same field as feature ID 301. Feature name 502 is a field that stores the name of the feature. Value 503 is a field that stores the value of the feature. One feature contains as many rows as the number of values ​​that the feature can take. Verbalization template 504 is a field that stores a template for verbalizing the value of the feature. The template is text that expresses the value of the feature in language, and is set by the user based on the value and meaning of the feature.

[0055] FIG. 6 is a diagram illustrating an example of a data structure of the feature amount list 216 according to the first embodiment.

[0056] The feature list 216 stores entries each consisting of a feature ID 601, a value 602, a feature name 603, a verbalized feature 604, and a contribution value 605. One entry corresponds to one feature.

[0057] The feature amount ID 601 and value 602 are the same fields as the feature amount ID 301 and value 302. The feature amount name 603 is the same field as the feature amount name 502. The verbalized feature amount 604 is a field that stores the verbalized feature amount. The contribution value 605 is the same field as the contribution value 403.

[0058] FIG. 7 is a diagram showing an example of the data structure of the group information 224 according to the first embodiment.

[0059] The group information 224 stores entries each composed of a group ID 701, a feature ID list 702, a first verbalized feature list 703, a second verbalized feature list 704, and a contribution value sum 705. One entry corresponds to one group.

[0060] Group ID 701 is a field that stores the ID of a group. Feature ID list 702 is a field that stores a list of IDs of features that make up a group. First verbalized feature list 703 stores verbalized features whose contribution values ​​are positive. Second verbalized feature list 704 stores verbalized features whose contribution values ​​are negative. Sum of contribution values ​​705 stores the sum of the contribution values ​​of the features that make up a group.

[0061] 8A and 8B are diagrams illustrating an example of a data structure of the first text template information 225 according to the first embodiment.

[0062] The first text template information 225 includes a template table 800 that stores templates corresponding to the preamble of a prompt, and a template table 810 that stores templates corresponding to the main sentence of a prompt.

[0063] The template table 800 stores entries each consisting of a contribution value total type 801 and a text template 802 .

[0064] A contribution value total type 801 is a field that stores the type of the total of the contribution values ​​of the group. A text template 802 is a field that stores a template of a preamble according to the type of the total of the contribution values ​​of the group.

[0065] The template table 810 stores entries each consisting of a contribution value sum 811 and a text template 812 .

[0066] The contribution value sum 811 is a field for storing the range of the sum of the contribution values ​​of the group. The text template 812 is a field for storing a template of the main text of the prompt according to the sum of the contribution values. The main text is set in advance by the user.

[0067] In this embodiment, the main sentence is a question sentence, but is not limited to this. For example, the main sentence may be, for example, "The reason why there is a high possibility of a fire is," and the interpretation generation model may generate the following document.

[0068] Note that the division of the range of the contribution values ​​shown in FIG. 8B is an example and is not limited to this.

[0069] FIG. 9 is a diagram illustrating an example of the data structure of the prompt information 226 according to the first embodiment.

[0070] The prompt information 226 stores entries each consisting of a group ID 901 and a prompt 902. One entry corresponds to one group.

[0071] The group ID 901 is the same field as the group ID 701. The prompt 902 is a field for storing a prompt generated from the preamble and the main sentence. The prompt becomes an input to the interpretation sentence generation unit 241.

[0072] FIG. 10 is a diagram illustrating an example of a data structure of the analysis information 234 according to the first embodiment.

[0073] The analysis information 234 stores entries each consisting of a feature amount ID 1001 , a feature amount name 1002 , a relationship 1003 , and a comment 1004 .

[0074] Feature ID 1001 is the same field as feature ID 301. Feature name 1002 is a field that stores the name of the feature. Relationship 1003 is a field that stores information that indicates the relationship between the feature value and the contribution value. Possible relationships include monotonically increasing, monotonically decreasing, and abruptly increasing. Remarks 1004 is a field that stores information related to the relationship. For example, the value of a feature whose contribution value increases abruptly is stored.

[0075] FIG. 11 is a diagram illustrating an example of a data structure of the second text template information 235 according to the first embodiment.

[0076] The second text template information 235 stores an entry configured from a relationship 1101 and a text template 1102 .

[0077] The relationship 1101 is a field for storing the relationship between the feature value and the contribution value. The text template 1102 is a field for storing a text template that verbalizes the relationship.

[0078] FIG. 12 is a diagram showing an example of the data structure of the relationship document information 236 in the first embodiment.

[0079] The related document information 236 stores entries each consisting of a feature amount ID 1201, a feature amount name 1202, and a related document 1203. One entry corresponds to one related document.

[0080] The feature amount ID 1201 and the feature amount name 1202 are the same fields as the feature amount ID 1001 and the feature amount name 1002. The relationship document 1203 is a field that stores a document that verbalizes the relationship between the feature amount value and the contribution value.

[0081] FIG. 13 is a diagram showing an example of a data structure of a table for managing the interpretation sentence 244 in the first embodiment.

[0082] The table 1300 stores entries each consisting of a group ID 1301, a prompt 1302, and an interpretation sentence 1303. One entry corresponds to one group.

[0083] The group ID 1301 and the prompt 1302 correspond to the group ID 901 and the prompt 902. The interpretation sentence 1303 is a field for storing the interpretation sentence 244 obtained by inputting the prompt.

[0084] Next, a description will be given of the process executed by the interpretation sentence generating device 100. The interpretation sentence generating device 100 generates a business interpretation of the predicted value 203 and presents it to the user.

[0085] 14 is a flowchart for explaining an outline of a process executed by the interpretation sentence generating device 100 according to the first embodiment. In this embodiment, an interpretation sentence for a prediction of a fire probability will be explained as an example.

[0086] The XAI unit 211 of the interpretation sentence generation device 100 uses the input data 202 and the predicted value 203 to calculate the contribution value of each feature of the input data 202 (step S101).

[0087] As a method for calculating the contribution value, for example, SHAP can be used. SHAP uses a baseline that serves as a reference for calculating the contribution value. The XAI unit 211 determines the contribution value of each feature amount of the property data based on the relative value of the feature amount with the baseline value as a reference. In this example, the XAI unit 211 calculates the contribution value of each feature amount of the property data (input data 202) input to calculate the fire probability.

[0088] Next, the interpretation sentence generation device 100 executes a feature selection process for selecting features to be used for generating a prompt (step S102).

[0089] Here, the feature selection process will be described. FIG. 15 is a flowchart illustrating an example of the feature selection process executed by the interpretation sentence generation device 100 according to the first embodiment.

[0090] The selection unit 212 acquires the contribution value information 214 (step S201).

[0091] Next, the selection unit 212 refers to the feature amount information 215 and verbalizes the feature amount value (step S202).

[0092] Specifically, for each feature, the selection unit 212 refers to an entry with a matching feature ID, and acquires the languageization template 504 of the row corresponding to the value 503. The selection unit 212 may use the languageization template as is, or may modify the template according to the value of the feature, as necessary.

[0093] Next, the selection unit 212 selects a feature amount based on the contribution value of each feature amount (step S203).

[0094] Specifically, the selection unit 212 selects a predetermined number of feature quantities in descending order of the absolute value of the contribution value. This makes it possible to efficiently select feature quantities that are considered important by the inference model. In order to present the selected feature quantities, the selection unit 212 generates ranking information that stores entries made up of the verbalized feature quantities and contribution values ​​of the selected feature quantities. The entries of the ranking information are sorted in descending order of the absolute value of the contribution value.

[0095] The feature amount presenting unit 213 presents the selected feature amount (step S204). That is, ranking information is presented.

[0096] FIG. 19 is a diagram illustrating an example of a screen displayed by the feature amount presenting unit 213 according to the first embodiment.

[0097] The feature amount presenting unit 213 displays a screen 1900 on the user terminal 120. The screen 1900 includes sections 1901 and 1903, ranking information 1902, and operation buttons 1904.

[0098] Section 1901 is a section that displays the predicted value 203. Section 1903 is a section that displays the number of interpretation sentence patterns generated from the selected feature amount. The number of interpretation sentence patterns corresponds to the number of groups.

[0099] Ranking information 1902 stores entries each including a check box for the user to select a feature to be used, a ranking, a verbalized feature, and a contribution value.

[0100] Typically, the selection unit 212 selects a feature to be used based on the magnitude of the absolute value of the contribution value, and adds a check to the check box of the selected feature. However, the user can also operate the check box to select a feature. By the user selecting a feature, information for generating the interpretation sentence that the user wants can be acquired. Also, when the number of interpretation sentence patterns displayed in section 1903 is large, the user can reduce the number of features to be used.

[0101] When an operation button 1904 is pressed, information on the feature amount selected by the check box is transmitted to the selection unit 212 .

[0102] The selection unit 212 generates and outputs the feature amount list 216 based on the operation on the screen 1900 (step S205).

[0103] This concludes the description of the feature selection process. Returning to the description of Fig. 14, the interpretation sentence generation device 100 executes the prompt generation process using the selected feature (step S103).

[0104] The prompt generation process will now be described with reference to the flowchart of FIG 16. The prompt generation process is executed by the interpretation sentence generation device 100 according to the first embodiment.

[0105] The group generating unit 221 acquires the feature amount list 216 (step S301).

[0106] Next, the group generation unit 221 generates groups by combining the features registered in the feature list 216 (step S302). The group generation unit 221 registers information about the generated groups in the group information 224. Here, the same number of groups as the number of possible combinations of features are generated. As described later, a prompt is generated for each group. Therefore, a variety of interpretation sentences 244 can be generated.

[0107] The prompt generating section 222 refers to the first text template information 225 and generates a prompt for each group (Step S303).

[0108] As shown in Figures 8A and 8B, the first text template information 225 manages the sum of the contribution values ​​of the groups (the characteristics of the contribution values) in association with the corresponding text. The prompt generation unit 222 refers to the template table 800 in Figure 8A and generates a preamble based on the sum of the contribution values ​​of each group. The prompt generation unit 222 also refers to the template table 810 in Figure 8B and generates a question based on the sum of the contribution values ​​of each group. The prompt generation unit 222 prompts and generates a document that combines the preamble and the question.

[0109] The first text template information 225 in the first embodiment is an example, and may include, for example, only one table or three or more tables. Note that a table in which the characteristics of the contribution values ​​other than the sum of the contribution values ​​are associated with the templates may also be used.

[0110] The first text template information 225 can generate different prompts for each group because the prompt changes depending on the features included in the group and the contribution values ​​of the features. For example, by using the template table 800, it is possible to include an adversative conjunction corresponding to the positive or negative contribution value in the preamble. This information represents the results of XAI, and the prompt can include information on the target inference model.

[0111] Next, the prompt presentation unit 223 presents the prompt generated by the prompt generation unit 222 (step S304).

[0112] FIG. 20 is a diagram showing an example of a screen displayed by the prompt presenting unit 223 according to the first embodiment.

[0113] The prompt presentation unit 223 displays a screen 2000 on the user terminal 120. The screen 2000 includes a table 2001 corresponding to the feature list 216, a table 2002 corresponding to the prompt information 226, and operation buttons 2003.

[0114] The table 2001 stores entries each including, for example, a feature name, a verbalized feature, and a contribution value. The entries may include a field for storing the value of the feature.

[0115] Table 2002 stores entries consisting of, for example, check boxes and prompts. The entries may also include a field for storing the sum of the contribution values ​​of each group, or the like.

[0116] The checkboxes are used to select the prompts to be input. By default, interpretations are generated for all prompts, but the user can also select the prompts to be used using the checkboxes. Note that the input format may be one in which the prompts can be modified.

[0117] When operation button 2003 is pressed, information about the prompt selected by the check box is sent to prompt generating unit 222 .

[0118] The prompt presentation unit 223 generates and outputs the prompt information 226 based on the operation on the screen 2000 (step S305).

[0119] This concludes the explanation of the prompt generation process. Returning to the explanation of FIG. 14, the interpretation sentence generation device 100 determines whether or not the feature to be analyzed is present among the selected feature quantities (step S104). For example, in predicting the probability of fire, the feature quantities to be analyzed are the building age, floor area, etc. The feature quantities to be analyzed are specified in advance by the user.

[0120] If the feature to be analyzed is not included in the selected features (S104: NO), the interpretation sentence generation device 100 acquires the prompt information 226 and generates the interpretation sentence 244 (step S105). Specifically, the interpretation sentence generation unit 241 generates the interpretation sentence 244 by inputting the prompt to the large-scale language model.

[0121] If the feature to be analyzed is included in the selected feature amounts (S104: YES), the interpretation sentence generating device 100 executes a relationship document generating process (step S106).

[0122] Here, the relationship document generation process will be described below. Fig. 17 is a flowchart illustrating an example of the relationship document generation process executed by the interpretation sentence generation device 100 according to the first embodiment.

[0123] The analysis unit 231 acquires the feature amount list 216 (step S401).

[0124] Next, the analysis unit 231 starts loop processing of the feature amounts (step S402). Specifically, the analysis unit 231 selects one feature amount from the feature amount list 216.

[0125] The analysis unit 231 determines whether the selected feature amount is the feature amount to be analyzed (step S403).

[0126] If the selected feature amount is not the feature amount to be analyzed (S403: NO), the analysis unit 231 proceeds to step S406.

[0127] If the selected feature is the feature to be analyzed (S403: YES), the analysis unit 231 analyzes the relationship between the feature value and the contribution value (step S404). Specifically, the analysis unit 231 uses XAI to analyze the change in the contribution value when the value of the target feature is changed, and plots the result on a graph.

[0128] Next, the analysis result presenting unit 232 presents the analysis result (step S405).

[0129] FIG. 21 is a diagram showing an example of a screen displayed by the analysis result presenting part 232 in the first embodiment.

[0130] The analysis result presenting unit 232 displays a screen 2100 on the user terminal 120. The screen 2100 includes a graph 2101, input sections 2102 and 2103, and operation buttons 2104 and 2105.

[0131] Graph 2101 is a graph showing the relationship between feature values ​​and contribution values. The horizontal axis shows feature values, and the vertical axis shows contribution values. Note that graph 2101 clearly shows feature values ​​in input data 202 to be explained. For example, in FIG. 21, they are shown by dotted lines.

[0132] The input sections 2102 and 2103 are sections where the user can view the graph and input relationships.

[0133] A global relationship is input in the input section 2102. The user selects the relevant relationship from options such as monotonically increasing, monotonically decreasing, none, etc. A local relationship is input in the input section 2103. When there are multiple local relationships, an input section 2103 is added by pressing an operation button 2104. The amount of change, etc. can also be added in the input section 2103.

[0134] When the operation button 2105 is pressed, the relationship information input in the input sections 2102 and 2103 is output to the analysis unit 231. At this time, the analysis unit 231 adds an entry to the analysis information 234.

[0135] After accepting the operation of the operation button 2105, the analysis unit 231 determines whether or not the processing has been completed for all the feature amounts (step S406).

[0136] If it is determined that the processing has not been completed for all feature amounts (S406: NO), the analysis unit 231 returns to step S402 and selects a new feature amount.

[0137] If it is determined that the processing has been completed for all the feature amounts (S406: YES), the analysis unit 231 outputs the analysis information 234 (step S407).

[0138] The relationship document generating unit 233 refers to the second text template information 235 and verbalizes the relationships included in the analysis information 234 to generate and output relationship document information 236 (step S408).

[0139] This concludes the description of the relationship document generation process. Returning to the description of Fig. 14, the interpretation sentence generation device 100 acquires the prompt information 226 and generates the interpretation sentence 244 (step S107). The process of step S107 is the same as the process of step S105.

[0140] The filtering unit 242 uses the relationship document to execute a filtering process on the interpretation sentence 244 (step S108).

[0141] Here, the filtering process will be described with reference to a flowchart of FIG 18. The filtering process is executed by the interpretation sentence generation device 100 according to the first embodiment.

[0142] The filtering unit 242 acquires the interpretation sentence 244 and the related document information 236 (step S501).

[0143] The filtering unit 242 starts loop processing of the interpretation sentences 244 (step S502). Specifically, the filtering unit 242 selects one interpretation sentence 244 from among the interpretation sentences 244.

[0144] Next, the filtering unit 242 starts a loop process of the relationship document (step S503). Specifically, the filtering unit 242 selects one relationship document from the relationship document information 236. Note that only relationship documents that indicate a global relationship may be selected.

[0145] The filtering unit 242 determines whether the selected interpretation sentence 244 includes the feature quantity that is the target of the selected relationship document (step S504).

[0146] If the selected interpretation sentence does not include the feature quantity that is the target of the selected relationship document (S504: NO), the filtering unit 242 proceeds to step S507.

[0147] If the selected interpretation sentence contains the feature quantity targeted by the selected related document (S504: YES), the filtering unit 242 judges whether the interpretation sentence 244 contains a fact that contradicts the related document (step S505). For example, by performing entailment recognition (Recognizing Textual Entailment: RTE) using BERT, which is one of the language models, it is possible to judge entailment or contradiction between two documents.

[0148] If it is determined that the interpretation sentence 244 does not include any facts that contradict the relationship document (S505: NO), the filtering unit 242 proceeds to step S507.

[0149] If it is determined that the interpretation sentence 244 includes a fact that contradicts the relationship document (S505: YES), the filtering unit 242 excludes the interpretation sentence 244 from the output target (step S506). After that, the filtering unit 242 ends the loop of the relationship document and proceeds to step S508.

[0150] The relationship document represents the relationship between the value of the selected feature that is important to the inference model of the explanation target and the contribution value. Since the behavior of the inference result of the target feature is not input to the interpretation generation model, there is a possibility that an interpretation sentence 244 containing facts that contradict the relationship document will be output. Such an interpretation sentence 244 is not suitable as an interpretation sentence 244 for the inference result of the inference model, so it is excluded in advance.

[0151] In step S507, the filtering unit 242 determines whether or not the processing has been completed for all the related documents (step S507).

[0152] If it is determined that the processing has not been completed for all relationship documents (S507: NO), the filtering unit 242 returns to step S503 and selects a new relationship document.

[0153] If it is determined that the processing has been completed for all the relationship documents (S507: YES), the filtering unit 242 proceeds to step S508.

[0154] In step S508, the filtering unit 242 determines whether or not the processing has been completed for all the interpretation sentences 244 (step S508).

[0155] If it is determined that the processing has not been completed for all interpretation sentences 244 (S508: NO), the filtering unit 242 returns to step S502 and selects a new interpretation sentence 244.

[0156] If it is determined that the processing has been completed for all interpretation sentences 244, the filtering unit 242 ends the filtering processing.

[0157] This concludes the description of the filtering process. Returning to the description of Fig. 14, the result presenting unit 243 acquires and outputs the interpretation sentence 244 (step S109).

[0158] FIG. 22 is a diagram showing an example of a screen 2200 displayed by the result presenting unit 243 of the first embodiment.

[0159] The result presenting unit 243 displays a screen 2200 on the user terminal 120. The screen 2200 displays a prompt and an interpretation sentence.

[0160] As described above, according to this embodiment, the interpretation sentence generation device 100 can generate and present an interpretation sentence that reflects business knowledge without accepting input of specialized knowledge from a user.

[0161] The present invention is not limited to the above-mentioned embodiment, but includes various modified examples. For example, the above-mentioned embodiment describes the configuration in detail to easily explain the present invention, and the present invention is not necessarily limited to the configuration including all the described configurations. Also, it is possible to add, delete, or replace a part of the configuration of each embodiment with another configuration.

[0162] In addition, each of the above configurations, functions, processing units, processing means, etc. may be realized in part or in whole by hardware, for example, by designing them as integrated circuits. The present invention can also be realized by software program code that realizes the functions of the embodiments. In this case, a storage medium on which the program code is recorded is provided to a computer, and a processor included in the computer reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the functions of the above-mentioned embodiments, and the program code itself and the storage medium storing it constitute the present invention. Examples of storage media for supplying such program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, SSDs (Solid State Drives), optical disks, magneto-optical disks, CD-Rs, magnetic tapes, non-volatile memory cards, ROMs, etc.

[0163] Furthermore, the program code for realizing the functions described in this embodiment can be implemented in a wide range of program or script languages, such as assembler, C / C++, perl, Shell, PHP, Python, Java (registered trademark), etc.

[0164] Furthermore, the program code of the software that realizes the functions of the embodiments may be distributed over a network and stored in a storage means such as a computer's hard disk or memory, or in a storage medium such as a CD-RW or CD-R, and the processor of the computer may read out and execute the program code stored in the storage means or storage medium.

[0165] In the above-mentioned embodiment, the control lines and information lines are shown as those considered necessary for the explanation, and not all the control lines and information lines are shown in the product. All the components may be connected to each other. [Explanation of symbols]

[0166] 100 Interpretation Generation Device 110 Reasoning device 120 User terminals 130 Network 101, 111, 121 CPUs 102, 112, 122 Memory 103, 113, 123 Auxiliary storage device 104, 114, 124 Network Interfaces 125 Input Devices 126 Output Device 201 Prediction Department 202 Input Data 203 Prediction 211 XAI Department 212 Selection section 213 Feature presentation unit 214 Contribution Value Information 215 Feature Information 216 Feature List 221 Group Generation Unit 222 Prompt Generation Unit 223 Prompt presentation section 224 Group Information 225 First Text Template Information 226 Prompt Information 231 Analysis Department 232 Analysis result presentation section 233 Relationship Document Generation Unit 234 Analysis information 235 Second Text Template Information 236 Related Document Information 241 Interpretation Generation Unit 242 Filtering section 243 Results presentation section 244 Interpretation 1900, 2000, 2100, 2200 screen

Claims

1. 1. A computer system comprising: The input data includes a plurality of features, and the input data is connected to an inference system that performs inference using an inference model. a large-scale language model that receives a directive as an input and outputs an interpretation for interpreting the result of the inference; and first text template information that stores first template data that associates a template of the directive with a characteristic of a contribution value that indicates a degree of contribution of a feature amount to the result of the inference in a group consisting of one or more feature amounts, Calculating the contribution value of each of the plurality of feature quantities using the input data, the result of the inference, and the inference model; generating a plurality of groups each composed of one or more of the feature quantities; acquire the first template data for each of the plurality of groups by referring to the first text template information based on a characteristic of the contribution value of the feature amount included in the group; generating the instruction sentence to be input to the large-scale language model based on the acquired first template data and the feature amount included in the group; a computer system that outputs the interpretation sentence obtained by inputting the instruction sentence into the large-scale language model;

2. 2. The computer system of claim 1, holding second text template information storing second template data in which a relationship between the feature value and the contribution value is associated with a second template for verbalizing the relationship; Analyzing the relationship between the feature quantities for which the relationship needs to be analyzed; Identifying the second template data by referring to the second text template information based on a result of the analysis; generating a relationship document that verbalizes the relationship based on the identified second template data and the feature amount that is required to analyze the relationship; performing language processing to determine whether the interpretation includes facts that are inconsistent with the relationship corresponding to the relationship document; A computer system comprising: a computer that excludes from the interpretation sentences to be output, the interpretation sentences including facts that contradict the relationships shown in the relationship document.

3. 3. The computer system of claim 2, selecting a plurality of target features based on the magnitude of the contribution value of the plurality of feature values; A computer system that generates the groups having the plurality of target feature quantities as elements.

4. 4. The computer system of claim 3, an interface for modifying the target feature; an interface for selecting the instruction sentences to be input to the large-scale language model; and an interface for inputting or modifying the relationship.

5. A method for generating an interpretation sentence for interpreting an inference result of an inference model, the method being executed by a computer system, comprising: The computer system comprises: The input data includes a plurality of features, and the input data is connected to an inference system that performs inference using an inference model. a large-scale language model that receives a directive as an input and outputs an interpretation for interpreting the result of the inference; and first text template information that stores first template data that associates a template of the directive with a characteristic of a contribution value that indicates a degree of contribution of a feature amount to the result of the inference in a group consisting of one or more feature amounts, The method for generating an interpretation sentence comprises the steps of: a first step in which the computer system calculates the contribution value of each of the plurality of feature quantities by using the input data, the result of the inference, and the inference model; a second step of the computer system generating a plurality of groups each composed of one or more of the feature quantities; a third step in which the computer system acquires, for each of the plurality of groups, the first template data by referring to the first text template information based on a characteristic of the contribution value of the feature included in the group; a fourth step in which the computer system generates the instruction sentence to be input to the large-scale language model based on the acquired first template data and the feature amount included in the group; and a fifth step in which the computer system outputs the interpreted sentence obtained by inputting the instruction sentence into the large-scale language model.

6. A method for generating an interpretation sentence according to claim 5, comprising the steps of: the computer system holds second text template information storing second template data in which a relationship between the feature value and the contribution value is associated with a second template for verbalizing the relationship; The method for generating an interpretation sentence comprises the steps of: a sixth step of the computer system analyzing the relationship for the feature quantities for which the relationship needs to be analyzed; a seventh step of the computer system identifying the second template data by referring to the second text template information based on a result of the analysis; an eighth step of generating a relationship document in which the relationship is verbalized based on the specified second template data and the feature amount required to analyze the relationship by the computer system; Including, The fifth step includes: the computer system performing language processing to determine whether the interpretation sentence includes facts that contradict the relationship corresponding to the relationship document; A method for generating an interpretation, characterized in that the computer system includes a step of excluding from the interpretation sentences to be output, the interpretation sentences that contain facts that contradict the relationship indicated in the relationship document.

7. A method for generating an interpretation sentence according to claim 6, comprising the steps of: The second step includes: the computer system selecting a plurality of target feature quantities based on the magnitudes of the contribution values ​​of the plurality of feature quantities; a step of generating the group having the plurality of target features as elements by the computer system.

8. A method for generating an interpretation sentence according to claim 7, comprising the steps of: the second step includes a step of the computer system providing an interface for modifying the target feature quantity; the fifth step includes a step of the computer system providing an interface for selecting the instruction sentence to be input to the large-scale language model; The method for generating an interpretation sentence, wherein the eighth step includes a step in which the computer system provides an interface for inputting or modifying the relationship.

Images