Report generation method and program

Abstract

The present invention relates to a report generation method with training and inference capabilities, the report generation method involving: during training, creating a report generation model by re-training a pre-trained large language to generate a target report using a documented sub-intent report and knowledge; and during inference, documenting and combining knowledge and a sub-intent report created in an intended format, and generating a report from the combined sub-intent report and knowledge by using the re-trained report generation model.

Description

Report generation method and program

[0001] The present invention relates to a report generation method and program for generating a report in response to a user's intent.

[0002] In recent years, intent-based management, which autonomously controls networks and computing resources in response to user requests (intents), has been under consideration.

[0003] Intent-based management, in order to realize natural interaction with users, a form in which users specify requirements through a natural language interface such as a chatpot is being considered. However, in this case, it is necessary to provide feedback to the user / operator in a natural way as to whether the requirements specified by the user are met.

[0004] TM Forum Technical Report: Intent Common Model - Intent Reporting, TR290B, Team Approved Date: 15-Aug-2023

[0005] In light of this background, standardization organizations and the like are considering formats for reporting whether or not a higher-level Intent in Turtle or Json-LD satisfies user-specified requirements, as described in Non-Patent Document 1, for example.

[0006] However, there is a problem in that it is not possible to provide a natural response as to whether a service in which the user is interested is available for a user who is not interested in the configuration of lower-level services or resources.

[0007] The present invention has been made in consideration of the above circumstances, and aims to provide a report generation method and program that can report to a higher-level Intent via the resource layer, service layer, and business layer as to whether or not the lower-level Intent is satisfied.

[0008] To achieve the above object, a first aspect of the present invention is a report generation method having training and inference functions, in which during training, a pre-trained large-scale language is retrained to generate target reports using documented lower-level intent reports and knowledge, thereby creating a report generation model.

[0009] Furthermore, a second aspect of the present invention is a report generation method of the first aspect, in which, during inference, lower-level Intent reports and knowledge (turtle, Json-LD), etc. created in any format are documented and combined, and a report is generated from the combined lower-level Intent reports and knowledge using a retrained report generation model.

[0010] Furthermore, a third aspect of the present invention is a program for report generation having training and inference functions, which causes a processor to realize the following functions: during training, retraining a pre-trained large-scale language to generate a target report using documented lower-intent reports and knowledge, and creating a report generation model; and during inference, documenting and combining lower-intent reports and knowledge (turtle, Json-LD, etc.) created in any format, and generating a report from the combined lower-intent reports and knowledge using the retrained report generation model.

[0011] The report generation method and program of the present invention make it possible to generate reports for a variety of knowledge and report formats, or for other information, and therefore makes it possible to provide a natural response regarding the availability of a service in which the user is interested, even for users who are not interested in the configuration of lower-level services or resources.

[0012] FIG. 1 is a conceptual diagram showing an example of an architecture that realizes a report generation method according to this embodiment. FIG. 2 is a conceptual diagram showing an image of an Intent report realized by the architecture. FIG. 3 is a sequence diagram illustrating an example of a processing sequence by the architecture. FIG. 4 is a diagram showing a flow of creating a report generation model during training. FIG. 5 is a diagram showing a flow of creating a report generation model during training. FIG. 6 is a diagram showing a flow of report generation during inference. FIG. 7 is a diagram showing examples of lower-level Intent reports and knowledge, etc., which are various formats that are input during inference. FIG. 8 is a diagram showing a specific example of report generation during inference.

[0013] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The drawings are schematic or conceptual, and the relationship between the thickness and width of each part, the size ratio between parts, etc., are not necessarily the same as those in reality. Furthermore, even when the same part is shown, the dimensions and ratios may be different depending on the drawing. In this specification and each drawing, elements similar to those described above with reference to the previous drawings are designated by the same reference numerals, and detailed and redundant explanations will be omitted as appropriate.

[0014] The report generation method according to the present embodiment is a report generation method having training and inference functions. During training, a pre-trained large-scale language is retrained to generate a target report using documented lower-level intent reports and knowledge, thereby creating a report generation model. During inference, lower-level intent reports and knowledge (e.g., turtle, Json-LD) created in any format are documented and combined, and a report is generated from the combined lower-level intent reports and knowledge using the retrained report generation model.

[0015] FIG. 1 is a conceptual diagram showing an example of an architecture for realizing a report generation method according to this embodiment.

[0016] This architecture 10 includes, from the top of the figure, an Intent extraction unit 11, a business Intent analysis unit 12, a service Intent analysis unit 13, and a resource Intent analysis unit 14 arranged in series.

[0017] The Intent extraction unit 11 generates an Intent report for the user U.

[0018] The business intent analysis unit 12 reports the intent for the operator P.

[0019] The number of service intent analyzers 13 corresponds to the number of services, and they report service intents.

[0020] The resource intent analysis units 14 are provided in a number corresponding to the number of resources, and report resource intents.

[0021] FIG. 2 is a conceptual diagram showing an image of an Intent report realized by the architecture.

[0022] FIG. 3 is a sequence diagram illustrating a processing sequence according to the architecture.

[0023] The report generation method according to this embodiment can use a chatpot for interaction with a user U, as illustrated in Fig. 2, using the architecture 10 illustrated in Fig. 1. Note that using a chatpot for interaction with a user U is merely an example and is not limited thereto, and can also be realized using a tool other than a chatpot.

[0024] The example shown on the left side of Figure 2 and the example shown on the right side of Figure 2 are examples of chatpot screens displayed on the terminal of user U. Please note that the parts displayed as speech bubbles in the figures indicate conversations in natural language, and the parts displayed as cloud marks in the figures are added in this specification for explanatory purposes and are not actually displayed on the terminal.

[0025] In the report generation method according to this embodiment, when a user U specifies an Intent related to service usage in a natural way, such as a chatpot (S1), the Intent extraction unit 11, business Intent analysis unit 12, service Intent analysis unit 13, and resource Intent analysis unit 14 in the architecture 10 work in cooperation with each other in response to the Intent extraction unit 11, business Intent analysis unit 12, service Intent analysis unit 13, and resource Intent analysis unit 14, as shown in sequences A and B in FIG. 3, to present a friendly expression to the user U regarding whether or not the service can be provided (S2), thereby generating an Intent report.

[0026] Depending on the Intent, there may be cases where the Intent cannot be satisfied due to changes in the state of lower-level resources or services. Even in such cases, the report generation method according to the present embodiment generates a report by presenting a friendly representation to the user U using the architecture 10.

[0027] The example chatpot screen shown on the left side of Fig. 2 and the example chatpot screen shown on the right side of Fig. 2 both show examples of responses presented when an Intent cannot be satisfied. The left side of Fig. 2 shows an example where the user U specifies an Intent (S1) and receives a response regarding whether or not the service can be provided (S2) in a single step.

[0028] On the other hand, the example of the chatpot screen shown on the right side of Fig. 2 is an example in which the user U specifies an intent (S1a, S1b, ...) and replies (S2a, ... S2n) regarding whether or not the service can be provided are made multiple times. The intent specification S1a, reply S2a, and intent specification S1b on the screen on the right side of Fig. 2 correspond to S1a, S2a, and S1b shown in Fig. 3.

[0029] (Example) Learning and inference functions can be incorporated into the architecture 10. In the following, an example in which learning and inference functions are incorporated into the architecture 10 will be described.

[0030] 4 and 5 are diagrams showing the flow of creating a report generation model during training.

[0031] During training, a pre-trained large language model (LLM) is retrained to generate a target report E using documented lower-intent reports and knowledge C, as illustrated in FIG. 4, to create a report generation model D.

[0032] An example of the documented lower-level Intent report and knowledge C is as illustrated in FIG.

[0033] FIG. 6 is a diagram showing the flow of report generation during inference.

[0034] During inference, lower-level Intent reports and knowledge F (F1, F2, ..., Fn ...) created in any format are converted into natural language to document G (G1, G2, ..., Gn ...) and combined H, and a report K is generated from the combined lower-level Intent reports and knowledge I using a retrained report generation model D, like a pre-trained large-scale language. This makes it possible to generate reports for a variety of knowledge and report formats, or for other information.

[0035] FIG. 7 is a diagram illustrating various formats of lower-level Intent reports and knowledge that are input during inference.

[0036] In Figure 7, examples of lower-level Intent reports and knowledge F include Lower layer intent 1 (F1) using turtle, Lower layer intent 2 (F2) using Json-LD, Knowledge 1 (Fn) using knowledge graph, and metadata (Fx) using natural language.

[0037] FIG. 8 is a diagram showing a specific example of report generation during inference.

[0038] As shown in Figure 8, lower layer intent 1 (F1) based on turtle is converted into natural language (G1). Similarly, lower layer intent 2 (F2) based on Json-LD, knowledge 1 (Fn) based on knowledge graph, and metadata (Fx) based on natural language are converted into natural language (G2, Gn, Gx). Furthermore, these converted natural language expressions are combined (H). Then, report K is generated from the combined natural language expressions using report generation model D.

[0039] As described above, according to the report generation method of this embodiment, various reports, knowledge, external information, etc. are documented, and a report generation model generated by retraining from an LLM, for example, is used to determine whether or not a user's request is satisfied from report information, knowledge information, metadata, and external information in a variety of formats, and a response can be given to the user U or operator P in the form of natural language. This makes it possible to provide a natural response regarding whether or not a service in which the user is interested can be provided, for a user who is not interested in the configuration of lower-level services or resources.

[0040] The present invention is not limited to the above-described embodiments, and the components can be modified and embodied in practice without departing from the spirit of the invention. Furthermore, various inventions can be created by appropriately combining multiple components disclosed in the above-described embodiments. For example, some components may be omitted from all the components shown in the embodiments. Furthermore, components from different embodiments may be appropriately combined.

[0041] 10 Architecture 11 Intent Extraction Unit 12 Business Intent Analysis Unit 13 Service Intent Analysis Unit 14 Resource Intent Analysis Unit A Sequence B Sequence C Documented Lower Intent Report and Knowledge D Report Generation Model E Target Report F Combined Lower Intent Report and Knowledge H Combine K Report P Operator U User

Claims

1. A report generation method having training and inference capabilities, wherein during training, a pre-trained large-scale language is retrained to generate target reports using documented lower-level intent reports and knowledge to create a report generation model.

2. The report generation method of claim 1, wherein, during inference, lower-level intent reports and knowledge created in any format are documented and combined, and a report is generated from the combined lower-level intent reports and knowledge using the retrained report generation model.

3. A program for report generation having training and inference functions, which causes a processor to realize the following functions: (a) during training, retraining a pre-trained large-scale language to generate a target report using documented sub-intent reports and knowledge, and creating a report generation model; and (b) during inference, documenting and combining sub-intent reports and knowledge created in any format, and generating a report from the combined sub-intent reports and knowledge using the retrained report generation model.

Images