Context creation support system, context creation support method, and context creation support program

Abstract

To provide a context creation support system, a context creation support method, and a context creation support program for efficiently utilizing generative AI.SOLUTION: A support server 20 comprises a path information storage unit 23 that records paths for linking a plurality of tokens according to their path strengths, and a control unit 21 that is connected to an AI system 30. The control unit 21 performs a first process of generating each original token as a preceding token by vectorizing each element constituting a query; and a second process of acquiring subsequent tokens connected to each preceding token by the path according to the path strength using the path information storage unit 23. The control unit 21 repeats the second process to acquire a context consisting of the preceding token and the subsequent token, inputs the context to generative AI, acquires a response, and updates the path strength according to the evaluation result of the response.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a context creation support system, a context creation support method, and a context creation support program for using generative AI.

Background Art

[0002] Today, generative AI is being used in various fields. For example, a data generation method using a language model has been studied (for example, Patent Document 1). The data generation method using the language model described in this document constructs a prompt that becomes the input sentence of the language model using original data. Next, the prompt is input into the language model. Then, new data and label information for the new data are generated from the language model.

[0003] In addition, prompt engineering for designing and optimizing prompts for efficiently using a language model in order to solve a desired problem (task) has also attracted attention (for example, Non-Patent Document 1).

[0004] As such methods for utilizing generative AI, there are an embedding method and a fine-tuning method. In the embedding method, knowledge is embedded in tokens generated by text elements constituting a query to express the semantic relationship between the text elements. On the other hand, in the fine-tuning method, a pre-trained model is adjusted to be suitable for a specific task.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Non-Patent Documents

[0006]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0007] However, in the embedding method, the document is divided in advance based on the structure of the document. Then, nodes generated by dividing the document are prepared. And knowledge is extracted from the nodes using the query. In this case, since the nodes divided by the physical structure of the document are used, the knowledge extracted from the nodes does not always meet the user's needs.

Means for Solving the Problems

[0008] The context creation support system for solving the above problems includes a path information storage unit that records paths connecting a plurality of tokens according to path strength, and a control unit connected to the generative AI. And the control unit performs a first process of generating each original token obtained by vectorizing each element constituting the query as a preceding token, a second process of using the path information storage unit to obtain a succeeding token connected to each preceding token by a path according to the path strength, repeating the second process to obtain a context composed of the preceding token and the succeeding token, inputting the context into the generative AI to obtain a response, and updating the path strength according to the evaluation result of the response.

Effects of the Invention

[0009] According to the present invention, it is possible to generate a context for using a generative AI.

Brief Description of the Drawings

[0010]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Modes for Carrying Out the Invention

[0011] According to FIGS. 1 to 10, an embodiment in which a context creation support system, a context creation support method, and a context creation support program are embodied will be described. In this embodiment, it is assumed that a prompt based on a query is generated. Here, a query is a task that a user causes a generative AI to execute, and examples thereof include tasks such as generating text, code, images, music, and the like. A prompt is the content of specific instructions given to a generative AI to execute a query. As shown in FIG. 1, the context creation support system of this embodiment uses a user terminal 10, a support server 20, and an AI system 30 that are interconnected via a network.

[0012] (Hardware Configuration Example) FIG. 2 is a hardware configuration example of the information processing apparatus H10 that functions as the user terminal 10, the support server 20, the AI system 30, and the like.

[0013] The information processing apparatus H10 includes a communication device H11, an input device H12, a display device H13, a storage device H14, and a processor H15. Note that this hardware configuration is an example, and it may have other hardware.

[0014] The communication device H11 is an interface that establishes a communication path with other devices and performs data transmission and reception, and is, for example, a network interface or a wireless interface.

[0015] The input device H12 is a device that receives input from a user or the like, and is, for example, a mouse or a keyboard. The display device H13 is a display or a touch panel that displays various information.

[0016] The storage device H14 is a storage device that stores data and various programs for executing various functions of the user terminal 10, the support server 20, and the AI system 30. Examples of the storage device H14 include a ROM, a RAM, and a hard disk.

[0017] The processor H15 controls each process in the user terminal 10, the support server 20, and the AI system 30 using the programs and data stored in the storage device H14. Examples of the processor H15 include, for example, a CPU and an MPU. This processor H15 expands the program stored in the ROM or the like into the RAM and executes various processes corresponding to various processes. For example, when the application programs of the user terminal 10, the support server 20, and the AI system 30 are started, the processor H15 operates a process for executing each process described later.

[0018] Processor H15 is not limited to performing software processing for all processes it executes. For example, processor H15 may include a dedicated hardware circuit (e.g., an application-specific integrated circuit: ASIC) that performs hardware processing for at least a part of the processes it executes. That is, processor H15 may be configured, for example, as follows.

[0019] [1] One or more processors that operate according to a computer program (software) [2] One or more dedicated hardware circuits that execute at least a part of various processes [3] A combination thereof, including circuitry The processor includes a CPU and memories such as a RAM and a ROM, and the memories store program codes or instructions configured to cause the CPU to execute processes. The memory, that is, the computer-readable medium, includes any available medium that can be accessed by a general-purpose or dedicated computer.

[0020] (Function of each information processing device) Using FIG. 1, the functions of the user terminal 10, the support server 20, and the AI system 30 used for this machine learning will be described.

[0021] The user terminal 10 is a computer terminal used by a user who uses this system. The support server 20 is a computer system for using a generative AI. This support server 20 includes a control unit 21, a node information storage unit 22, a path information storage unit 23, and an evaluation information storage unit 24.

[0022] The control unit 21 performs processes (processes including a management stage, a prediction stage, an evaluation stage, etc.) described later. By executing a context creation support program for this purpose, the control unit 21 functions as a management unit 211, a prediction unit 212, an evaluation unit 213, etc.

[0023] The management unit 211 executes a process of acquiring various information from the user terminal 10. The prediction unit 212 generates a prompt to be input to the AI system 30, and uses the AI system 30 to execute a process of obtaining a response and generating an output. The evaluation unit 213 evaluates the response obtained from the AI system 30 and executes a process of learning a path.

[0024] In the node information storage unit 22, node management information for generating a context based on tokens is recorded. This node management information is registered when learning the relevance of tokens obtained by dividing a large number of documents and vectorizing words (elements). The node management information records a node ID and information regarding a token group.

[0025] The node ID is an identifier for specifying each node indicating knowledge obtained by grouping tokens of relevant words. The token group is a group of tokens of a plurality of relevant words included in each node. Here, a token group generated based on a group of words whose mutual relevance can be estimated is used. As shown in FIG. 6, the nodes N1, N2, N3,... Nm, Nn respectively include tokens T11, T21, T31... Tm1, Tn1.

[0026] In the path information storage unit 23 shown in FIG. 1, path management information regarding a path connecting tokens is recorded. The path management information is registered during path learning. The path management information records a path ID, a first token, a second token, and a weighting value.

[0027] The path ID is an identifier for specifying a path connecting tokens of relevant words (the first token, the second token). The first token and the second token are tokens of relevant words. The weighting value is a value indicating the path strength obtained by evaluating the strength of the path connecting the first token and the second token.

[0028] The evaluation information storage unit 24 stores evaluation management information obtained by evaluating the response acquired from the AI system 30. This evaluation management information is recorded when the evaluation information is acquired from the user terminal 10. In the evaluation information storage unit 24, evaluation results are recorded for query tokens, contexts, and response tokens.

[0029] A query token is a token of a word that constitutes a query input to the AI system 30. A context is the context used when this query is input to the AI system 30.

[0030] A response token is a token of a word that constitutes a response acquired from the AI system 30 by inputting a prompt generated based on a query to the AI system 30. The evaluation result is the evaluation result of the response acquired from the AI system 30.

[0031] The AI system 30 is a computer system that utilizes a generative AI technology for generating a response to a prompt. The AI system 30 is configured according to a calculation model and parameters that are elements for determining the structure of the generative AI.

[0032] (Path generation process) The path generation process will be described with reference to FIG. 3. First, the control unit 21 of the support server 20 executes a process of acquiring an existing question-and-answer set (step S10). Specifically, the management unit 211 of the control unit 21 acquires existing questions (Q) and answers (A) prepared in advance from the user terminal 10.

[0033] Next, the control unit 21 of the support server 20 executes a token generation process from the words of the existing question (step S11). Specifically, the prediction unit 212 of the control unit 21 generates tokens based on the words that constitute the acquired question. As shown in FIG. 6, tokens TQ11, TQ12, TQ13, etc. are generated from the words included in the question Q1.

[0034] Next, the control unit 21 of the support server 20 executes node identification processing using the token (step S12). Specifically, the prediction unit 212 of the control unit 21 identifies, in the node information storage unit 22, the nodes that include the tokens of the words constituting the question and the tokens estimated to have a high semantic similarity. Here, the semantic similarity is an index indicating the degree of semantic similarity between a plurality of tokens.

[0035] As shown in FIG. 6, the nodes N1, N3, Nn corresponding to the tokens TQ11, TQ12, TQ13 included in the question Q1 are identified. Next, the control unit 21 of the support server 20 executes context acquisition processing according to the node (step S13). Specifically, the prediction unit 212 of the control unit 21 acquires, as the context, the tokens included in the nodes recorded in the node information storage unit 22.

[0036] Next, the control unit 21 of the support server 20 executes prompt generation processing with the context added (step S14). Specifically, the prediction unit 212 of the control unit 21 generates a prompt with the acquired context added.

[0037] Next, the control unit 21 of the support server 20 executes input processing to the generation AI (step S15). Specifically, the prediction unit 212 of the control unit 21 inputs the generated prompt to the AI system 30.

[0038] Next, the control unit 21 of the support server 20 executes response acquisition processing from the generation AI (step S16). Specifically, the prediction unit 212 of the control unit 21 acquires the response corresponding to the prompt from the AI system 30.

[0039] Next, the control unit 21 of the support server 20 executes response evaluation processing based on the existing answer (step S17). Specifically, the evaluation unit 213 of the control unit 21 evaluates the response according to the degree of match between the response obtained from the AI system 30 and the existing answer. This degree of match is evaluated including not only the match of words but also the semantic match. When it is determined that the degree of match is higher than the reference value, the evaluation unit 213 determines that the context used for the prompt is correct. Then, the evaluation unit 213 records in the evaluation information storage unit 24 the evaluation management information obtained by evaluating the response obtained from the AI system 30.

[0040] Next, the control unit 21 of the support server 20 executes token path generation processing for the identified tokens (step S18). Specifically, when the evaluation unit 213 of the control unit 21 determines that the context is correct, it generates a path connecting the tokens of the plurality of words constituting the question and the tokens used in the context, and records it in the path information storage unit 23. As shown in FIG. 7, paths P1, P2, P3, etc. are generated between the tokens included in the nodes N1, N3, Nn and used in the context.

[0041] Next, the control unit 21 of the support server 20 executes path weighting processing according to the evaluation result (step S19). Specifically, the evaluation unit 213 of the control unit 21 calculates a weighting value using the evaluation management information recorded in the evaluation information storage unit 24. Then, the evaluation unit 213 records in the path information storage unit 23 the weighting value calculated based on the degree of match (evaluation result).

[0042] (Processing for new queries) The processing for new queries will be described with reference to FIG. 4. First, the control unit 21 of the support server 20 executes new query acquisition processing (step S20). Specifically, the management unit 211 of the control unit 21 acquires a new query from the user terminal 10. Here, it is assumed that a query including new words that have not existed in the past is acquired. Next, the control unit 21 of the support server 20 executes token generation processing from the words of the new query (step S21). Specifically, the prediction unit 212 of the control unit 21 generates tokens (new tokens) based on the words constituting the acquired new query. As shown in FIG. 8, tokens TQ21, TQ22, and TQ23 included in the new query Q2 are generated.

[0043] Next, the control unit 21 of the support server 20 executes association processing between the new tokens and the past tokens (step S22). Specifically, the prediction unit 212 of the control unit 21 identifies tokens with a high semantic similarity to the new tokens among the query tokens (past tokens) recorded in the evaluation information storage unit 24.

[0044] As shown in FIG. 8, tokens TQ31, TQ32, and TQ33 included in the past query Q3 are identified corresponding to the tokens TQ21, TQ22, and TQ23 included in the new query Q2.

[0045] Next, the control unit 21 of the support server 20 executes context acquisition processing according to the past tokens (step S23). Specifically, the prediction unit 212 of the control unit 21 acquires the path management information including the past tokens from the path information storage unit 23. Next, the prediction unit 212 acquires the subsequent tokens connected by paths whose weighting values exceed the reference value. Then, the prediction unit 212 identifies the acquired subsequent tokens as the context.

[0046] As shown in FIG. 8, tokens T11, T31, Tn1, etc. connected by paths corresponding to tokens TQ31, TQ32, TQ33, etc. are identified. Further, subsequent tokens connected to tokens T11, T31, Tn1 by paths are identified. Then, for the tokens included in the query, the subsequent tokens identified by the paths are used as the context.

[0047] Next, the control unit 21 of the support server 20 executes generation processing of a prompt with the context added, similar to step S14 (step S24). Next, the control unit 21 of the support server 20 executes an input process on the generation AI in the same manner as in step S15 (step S25). Next, the control unit 21 of the support server 20 executes a response acquisition process from the generation AI in the same manner as in step S16 (step S26).

[0048] Next, the control unit 21 of the support server 20 executes a response evaluation process (step S27). Specifically, the evaluation unit 213 of the control unit 21 outputs the response acquired from the AI system 30 to the user terminal 10. In this case, the user evaluates whether the response to the new query is appropriate on the user terminal 10. Next, the evaluation unit 213 acquires the evaluation result regarding the response from the user terminal 10. Then, the evaluation unit 213 records the evaluation management information obtained by evaluating the response acquired from the AI system 30 in the evaluation information storage unit 24.

[0049] Next, the control unit 21 of the support server 20 executes a path generation process for the identified token (step S28). Specifically, when the evaluation unit 213 of the control unit 21 determines that the context is correct, it generates a path that directly concatenates the tokens of the plurality of words constituting the new query and the tokens used in the context, and records it in the path information storage unit 23.

[0050] Next, the control unit 21 of the support server 20 executes a path weighting process according to the evaluation result (step S29). Specifically, the evaluation unit 213 of the control unit 21 calculates a weighting value using the evaluation management information recorded in the evaluation information storage unit 24. Then, the evaluation unit 213 updates the weighting value of the path of the token used in the prompt in the path information storage unit 23.

[0051] (Path update process) The path update process will be described with reference to FIG. 5. First, the control unit 21 of the support server 20 executes a new query acquisition process in the same manner as in step S20 (step S30).

[0052] Next, the control unit 21 of the support server 20 executes token generation processing from the words of the new query in the same manner as in step S21 (step S31). As shown in FIG. 9, a first process is performed to generate tokens TQ41, TQ42, TQ43 (original tokens) included in the new query Q4.

[0053] Next, the control unit 21 of the support server 20 executes path identification processing using the tokens (step S32). Specifically, the prediction unit 212 of the control unit 21 acquires all path management information including the new token from the path information storage unit 23.

[0054] As shown in FIG. 9, paths P11, P21, P31 related to the tokens TQ41, TQ42, TQ43 included in the new query Q4 are identified. Further, a path P22 is connected to the connection destination token of the path P21. In this way, subsequent tokens in which the paths are continuously connected are acquired.

[0055] As shown in FIG. 10, tokens TQ41, TQ42,... TQ4m included in the new query Q4 are identified as preceding tokens. A second process is performed to acquire tokens X11 to X1L, tokens X21 to X2L, and tokens Xm1 to XmL that are continuously connected from the preceding tokens as subsequent tokens. Then, for the tokens TQ41 to TQ4m included in the new query Q4, the tokens X11 to XmL identified by the path are used as context. In this case, the tokens X11 to XmL are generated as context by a matrix of m rows and L columns.

[0056] Next, the control unit 21 of the support server 20 executes processing to acquire path weighting (step S33). Specifically, the prediction unit 212 of the control unit 21 identifies paths in which the weighting values recorded in the acquired path management information are equal to or greater than the reference value. In this case, if the paths are connected by paths with low weighting values, the acquisition of subsequent tokens using the paths is terminated.

[0057] Next, the control unit 21 of the support server 20 executes a context acquisition process according to the path (step S34). Specifically, the prediction unit 212 of the control unit 21 uses the path information storage unit 23 to acquire tokens concatenated by the specified path. Then, the prediction unit 212 specifies the acquired tokens as the context.

[0058] Next, the control unit 21 of the support server 20 executes a process of generating a prompt with added context in the same manner as in step S14 (step S35). Next, the control unit 21 of the support server 20 executes an input process on the generation AI in the same manner as in step S15 (step S36).

[0059] Next, the control unit 21 of the support server 20 executes a process of acquiring a response from the generation AI in the same manner as in step S16 (step S37). Next, the control unit 21 of the support server 20 executes a response evaluation process in the same manner as in step S27 (step S38). Next, the control unit 21 of the support server 20 executes a path weighting process according to the evaluation result in the same manner as in step S29 (step S39).

[0060] (Operations of the Embodiment) A path concatenating tokens used for the context is generated according to the evaluation result of the response.

[0061] (Effects of the Embodiment) (1) In this embodiment, the control unit 21 of the support server 20 executes the acquisition process of the existing question-and-answer set (step S10) to the process of generating a prompt with added context (step S14). Thereby, a prompt with added context can be generated using the nodes.

[0062] (2) In this embodiment, the control unit 21 of the support server 20 executes response evaluation processing (step S17) and path generation processing for the identified token (step S18) based on the existing response. Thereby, it is possible to generate a path for the token used in the context not only within the node but also across nodes.

[0063] (3) In this embodiment, the control unit 21 of the support server 20 executes path weighting processing according to the evaluation result (step S19). Thereby, it is possible to weight the paths of tokens close to the existing response.

[0064] (4) In this embodiment, the control unit 21 of the support server 20 executes acquisition processing of a new query (step S20) to association processing of a new token and a past token (step S22). Thereby, it is possible to identify a token for acquiring a context by using the existing query.

[0065] (5) In this embodiment, the control unit 21 of the support server 20 executes response evaluation processing (step S27). Thereby, it is possible to evaluate the used path. (6) In this embodiment, the control unit 21 of the support server 20 executes path generation processing for the identified token (step S28) and path weighting processing according to the evaluation result (step S29). Thereby, it is possible to weight the paths of relevant tokens.

[0066] (7) In this embodiment, the control unit 21 of the support server 20 executes path identification processing using the token (step S32). Thereby, it is possible to identify a path for generating a context based on the query.

[0067] (8) In this embodiment, the control unit 21 of the support server 20 executes acquisition processing of path weighting (step S33). Thereby, it is possible to identify a token used in the context by the learned weighting.

[0068] (9) In this embodiment, the control unit 21 of the support server 20 executes context acquisition processing according to the path (step S34). Thereby, the context can be acquired based on the learned path without being limited to the range of nodes.

[0069] (10) In this embodiment, the control unit 21 of the support server 20 executes path weighting processing according to the evaluation result (step S39). Thereby, the path for concatenating tokens can be updated.

[0070] This embodiment can be implemented with the following modifications. This embodiment and the following modification examples can be implemented in combination with each other within a technically non - conflicting range. · In the above - described embodiment, the user terminal 10, the support server 20, and one AI system 30 are used. The hardware configuration is not limited thereto. For example, a plurality of AI systems 30 may be used. Also, the AI system 30 may be provided within the support server 20.

[0071] · In the above - described embodiment, the control unit 21 of the support server 20 executes path identification processing (step S32) and path weight acquisition processing (step S33) using tokens. In this case, when concatenated with a path having a low weight value, the acquisition of subsequent tokens using the path is terminated. The method of determining the length of the path to be used by terminating the acquisition of subsequent tokens is not limited to the case of using the weight value.

[0072] For example, an upper limit value of consecutive paths may be set, and the length of the path to be used may be adjusted according to this upper limit value. Alternatively, a function for calculating the similarity between a query token and each token obtained from the path from this token may be used. In this case, the length of the path is adjusted using the tokens when the calculated similarity is equal to or greater than a reference value. As a function for calculating the similarity, for example, a function that takes as variables the degree of coincidence (cosine similarity) of the directions of the vectors of two tokens and the weighting value of the path can be used.

[0073] Also, a statistical value (for example, an average value) obtained by integrating the similarities of each token sequentially obtained by the path may be sequentially calculated, and the length of the path may be adjusted using the tokens when this statistical value is equal to or greater than a reference value. Alternatively, a function for calculating the similarity between all tokens of the query and each token obtained from the path from a predetermined token may be used. In this case, the length of the path may be adjusted using the tokens when the statistical value (for example, the average value) of each calculated similarity is equal to or greater than a reference value.

[0074] · In the above embodiment, the control unit 21 of the support server 20 executes the context acquisition process (step S13) according to the node to the response evaluation process (step S17) by the existing response. Here, the tokens in the node may be sequentially specified and used as the context, or the tokens may be randomly specified and used as the context.

Explanation of Signs

[0075] 10... user terminal, 20... support server, 21... control unit, 211... management unit, 212... prediction unit, 213... evaluation unit, 22... node information storage unit, 23... path information storage unit, 24... evaluation information storage unit, 30... AI system.

Claims

1. A context creation support system comprising a path information storage unit that records paths connecting a plurality of tokens according to path strength, and a control unit connected to a generation AI, wherein the control unit, performs a first process of generating each original token obtained by vectorizing each element constituting a query as a preceding token, performs a second process of using the path information storage unit to obtain, according to path strength, succeeding tokens connected to the respective preceding tokens by paths, repeats the second process to obtain a context composed of the preceding tokens and the succeeding tokens, inputs the context into the generation AI to obtain a response, and updates the path strength according to the evaluation result of the response, characterized in that it is a context creation support system.

2. The context creation support system according to claim 1, characterized in that the control unit ends the repetition of the second process according to the path strength.

3. The control unit, identifies a node including each original token obtained by vectorizing each element constituting the query, obtains the context of the node, inputs the context into the generation AI to obtain the response, and generates a path to a token included in the context according to the evaluation result of the response included in the original token and the node, characterized in that it is the context creation support system according to claim 1 or 2.

4. A context creation support method for assisting in obtaining a response from a generation AI using a context creation support system comprising a path information storage unit that records paths connecting a plurality of tokens according to path strength, and a control unit connected to the generation AI, wherein the control unit, performs a first process of generating each original token obtained by vectorizing each element constituting a query as a preceding token, performs a second process of using the path information storage unit to obtain, according to path strength, succeeding tokens connected to the respective preceding tokens by paths, repeats the second process to obtain a context composed of the preceding tokens and the succeeding tokens, inputs the context into the generation AI to obtain a response, and updates the path strength according to the evaluation result of the response, characterized in that it is a context creation support method.

5. A context creation support program for assisting in obtaining a response from the generation AI, using a context creation support system including a path information storage unit that records a path for connecting a plurality of tokens according to path strength and a control unit connected to the generation AI, wherein the control unit performs a first process of generating each original token obtained by vectorizing each element constituting a query as a preceding token, performs a second process of obtaining, according to path strength, a succeeding token connected to each of the preceding tokens by a path using the path information storage unit, repeats the second process to obtain a context composed of the preceding tokens and the succeeding tokens, inputs the context to the generation AI to obtain a response, and functions as means for updating the path strength according to an evaluation result of the response. A context creation support program.

Images