Inference device and inference method

Abstract

This inference device 1 comprises: a search unit 13 that searches for external information related to both a query pertaining to an operation of a communication network and preliminary knowledge information about the communication network; and an inference unit 15 that infers an output to the query and the external information from internal knowledge information.

Description

Inference device and inference method

[0001] The present disclosure relates to an inference device and an inference method.

[0002] There are large-scale language models that use internal knowledge information for learning and inference (Non-Patent Documents 1 and 2). To prevent hallucinations (plausible lies) from being included in the output from large-scale language models, there are techniques that utilize external information in addition to the internal knowledge information (Non-Patent Documents 3 and 4).

[0003] “Introducing ChatGPT”, OpenAI, [online], [Searched on February 28, 2020], <URL: https: / / openai.com / blog / chatgpt> “Introducing Llama 2”, Meta, [online], [Searched on February 28, 2020], <URL: https: / / ai.meta.com / llama / > Patrick Lewis, 11 others, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, arXiv:2005.11401v4 [cs.CL] 12 Apr 2021, [online], [Retrieved February 28, 2020], <URL: https: / / arxiv.org / abs / 2005.11401> Jingyu Wang, and 8 others, “Network Meets ChatGPT: Intent Autonomous Management, Control and Operation”, Journal of Communications and Information Networks, Vol.8, No.3, September 2023, p.239-p.255

[0004] When utilizing external information, external information related to the input query is searched for, and the searched external information is attached to the query and inferred from internal knowledge information using a large-scale language model. However, since the external information search is related only to the query, when applied to the operation of a communication network, there are cases where the necessary external information cannot be obtained sufficiently.

[0005] The present disclosure has been made in consideration of the above circumstances, and an object of the present disclosure is to provide a technology that can sufficiently acquire external information.

[0006] An inference device according to one aspect of the present disclosure includes a search unit that searches for external information related to both a query regarding the operation of a communication network and prior knowledge information about the communication network, and an inference unit that infers an output for the query and the external information from internal knowledge information.

[0007] An inference method according to one aspect of the present disclosure is an inference method performed by an inference device, which searches for external information related to both a query regarding the operation of a communication network and prior knowledge information about the communication network, and infers an output for the query and the external information from internal knowledge information.

[0008] According to the present disclosure, a technique that can sufficiently acquire external information can be provided.

[0009] FIG. 1 is a diagram showing a functional block configuration of an inference device. FIG. 2 is a diagram showing an example of prior knowledge information. FIG. 3 is a diagram showing a specific example 1 of prior knowledge information. FIG. 4 is a diagram showing a specific example 2 of prior knowledge information. FIG. 5 is a diagram showing an example of external information. FIG. 6 is a diagram showing a processing flow of an inference method. FIG. 7 is a diagram showing a hardware configuration of an inference device.

[0010] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the description of the drawings, the same parts are designated by the same reference numerals and the description thereof will be omitted.

[0011] [Summary of the Disclosure] In the operation of a communication network, data that can be easily handled mechanically (e.g., numerical values, alarms, logs) can be analyzed and utilized by the communication network system itself, whereas information written in natural language (e.g., free text) must be analyzed and utilized by a maintenance person.

[0012] In such cases, a large-scale language model (e.g., Large Language Model (LLM)) capable of analyzing and utilizing natural language can be utilized. Even when a large-scale language model is utilized in the operation of a communication network, there is information available as external knowledge (e.g., communication network design information, records of past troubleshooting, and manuals describing troubleshooting procedures), so the output for a query can be inferred by utilizing the external information.

[0013] However, when operating a communication network, it is necessary to search for external information taking into consideration not only the input query but also prerequisite prior knowledge (e.g., operational knowledge). For example, if a query such as "Device A, please tell me how to deal with alarm X," prior knowledge such as "the status of the redundant device of Device A" is also required.

[0014] Therefore, the present disclosure searches for external information related not only to the input query but also to prior knowledge information of the communication network. In this way, searching for external information related to both the query and prior knowledge information enables in-depth search of external information, making it possible to sufficiently obtain necessary external information.

[0015] [Configuration of Inference Apparatus] FIG. 1 is a diagram showing the functional block configuration of an inference apparatus 1 according to this embodiment.

[0016] The inference device 1 includes an input unit 11, an extraction unit 12, a search unit 13, a generation unit 14, an inference unit 15, an output unit 16, a prior knowledge information storage unit 17, an external information storage unit 18, and an internal knowledge information storage unit 19.

[0017] The input unit 11 has a function of receiving queries relating to the operation of the communication network to be managed from the operation terminal 2. The operation terminal 2 is a computer used by a maintenance person to operate the communication network.

[0018] The extraction unit 12 has a function of extracting search words from the received query.

[0019] The search unit 13 has a function of searching for external information related to both the query and the prior knowledge information of the communication network. For example, the search unit 13 includes an acquisition unit 131 that acquires the prior knowledge information of the communication network from the prior knowledge information storage unit 17, and a search unit 132 that searches the external information storage unit 18 for external information related to the extracted search word and the acquired prior knowledge information.

[0020] The generation unit 14 has a function of generating an inference query that is composed of a query and searched external information.

[0021] The inference unit 15 has a function of inferring an output for an inference query from pre-learned internal information stored in the internal knowledge information storage unit 19 .

[0022] The output unit 16 has a function of transmitting an output (inference result) in response to an inference query to the operation terminal 2 .

[0023] The prior knowledge information storage unit 17 has a function of storing prior knowledge information of a communication network.

[0024] The external information storage unit 18 has a function of storing external information.

[0025] The internal knowledge information storage unit 19 has a function of storing internal knowledge information that has been previously learned.

[0026] The external knowledge storage unit 18 may be provided in a device other than the inference device 1. The network knowledge information storage unit 17 and the internal knowledge information storage unit 19 may also be provided in a device other than the inference device 1.

[0027] [Example of Prior Knowledge Information] Prior knowledge information for a communication network is meta-knowledge information about external information. For example, it is how to link external information together and how to handle external information. Specifically, as shown in Figure 2, it can be expressed in sentences such as "The status of a group of redundantly configured devices must be checked simultaneously" and "Abnormalities in lower layers may affect upper layers."

[0028] (Specific Example 1 of Prior Knowledge Information) A specific example of prior knowledge information is configuration information of a communication network.

[0029] For example, a redundant configuration of devices, etc. If the communication network to be managed has a "combination of redundant devices" as shown in items 1 and 2 in Fig. 3(a), the "judgment conditions" and "deep search character strings" shown in items 1 and 2 in Fig. 3(c) are stored in the prior knowledge information storage unit 17 as prior knowledge information.

[0030] For example, it is a dependency relationship of devices, etc. When the communication network to be managed has a combination of a "dependency source" and a "dependency destination" as shown in Fig. 3(b), the "determination conditions" and "deep search character strings" described in items 3 and 4 of Fig. 3(c) are stored in the prior knowledge information storage unit 17 as prior knowledge information.

[0031] Figure 3(c) shows an image of a database of prior knowledge information related to the configuration information of a communication network. If the search data in the input query satisfies the "determination condition," the "deep search string" corresponding to the "determination condition" is used to search for external information.

[0032] 3(a), item 1 indicates that the three devices A, B, and C are in a redundant relationship, and when evaluating something for one of the devices, the information for the remaining devices must be evaluated at the same time. Therefore, as shown in item 1 in Fig. 3(c), a rule is defined that if a query includes one or more of the three devices A, B, and C, the search word in the query is replaced with "devices A, B, and C."

[0033] In the case of item 1 in Fig. 3(b), virtual server x exists and is dependent on the resources of physical server X. Therefore, as shown in item 3 in Fig. 3(c), a rule is defined that when evaluating something related to virtual server x, it is necessary to evaluate information related to physical server X as well. This rule states that when there is an abnormality in virtual server x, it is necessary to check whether there is an abnormality in the hardware of physical server X.

[0034] (Specific Example 2 of Prior Knowledge Information) A specific example of prior knowledge information is operation information of a communication network. Specifically, it is "common sense" and "standard tactics" of measures to be taken during operation and measures to be taken in the event of a failure that are considered necessary for performing network operations.

[0035] For example, if the state corresponds to a communication failure state, it is preferable to search for external information related to DNS. Therefore, as shown in item 1 of Figure 4, if the search data in the input query includes "Network communication available & HTTP communication unavailable", a deep search is performed using "DNS".

[0036] In addition to the "analogy of search character strings from communication status" shown in item 1 of FIG. 4, "analogy of search character strings from error codes" as shown in item 2-3 of FIG. 4, and "narrowing down causes using trends in events" as shown in item 4-6 of FIG. 4 may also be used.

[0037] [Examples of External Information] External information is, for example, various information stored on servers on an intranet or the Internet. There are information aggregation websites on the Internet, and external information is information written on those information aggregation websites.

[0038] For example, it is a manual that describes the design information of the communication network, the records of past troubleshooting, and the procedure for dealing with the problem. Specifically, it is something like "Device A has a redundant configuration with Device B", "Confirm alarm X of Device A, restart Device A, and complete recovery", and "Device B can be judged to be normal if there is a ping response" as shown in Figure 5.

[0039] [Operation of Inference Device] Fig. 6 is a diagram showing the processing flow of the inference method performed by the inference device 1. An example will be described in which the prior knowledge information is configuration information (redundant configuration) of a communication network.

[0040] Step S1: The input unit 11 receives a query regarding the operation of the communication network to be managed from the operation terminal 2. For example, the input unit 11 receives a query such as "Please tell me how to deal with alarm X on device A."

[0041] Step S2: The extraction unit 12 extracts a search word from the received query. For example, the extraction unit 12 extracts "device A, alarm X, action" from the query.

[0042] Step S3: The search unit 13 (acquisition unit 131) acquires prior knowledge information corresponding to the search word from the prior knowledge information storage unit 17, and adds the acquired prior knowledge information to the search word. For example, in the case of the above search word, since the "determination condition" of item No. 1 in Fig. 3(c) is satisfied, the search unit 13 acquires "device A," "device B," and "device C" from the "deep search character string" of the same item No. 1, and changes the above search word to "device A, device B, device C, alarm X, response."

[0043] Step S4: The search unit 13 (search unit 132) searches the external information storage unit 18 for external information related to the search word to which the prior knowledge information has been added. For example, in the case of the changed search word, since device B has been added, the search unit 13 acquires not only the external information of item No. 2 in Fig. 6 which includes "device A" and "alarm X" in the query, but also the external information of items No. 1 and No. 3 which include the added "device B."

[0044] In other words, the search unit 13 acquires external information such as "Device A has a redundant configuration with device B," "Alarm X of device A was confirmed, device A was restarted, and recovery was completed," and "Device B's normality can be confirmed by the presence of a ping response."

[0045] Step S5: The generation unit 14 generates an inference query consisting of the query received in step S1 and the external information searched for in step S4. For example, in the above example, the search unit 13 generates the inference queries such as "'Device A, please tell me how to deal with alarm X,' ", "Device A has a redundant configuration with device B,' ", "Confirm alarm X of device A, restart device A, and complete recovery,' and "Device B can be determined to be normal if there is a ping response."

[0046] Step S6: The inference unit 15 infers the output for the generated inference query from the pre-learned internal information stored in the internal knowledge information storage unit 19. For example, in the above example, the inference unit 15 outputs the inference result, "Alarm X of device A will be restored by restarting. Also, device A has a redundant configuration with device B. The normality of device B can be confirmed by the presence or absence of a ping response."

[0047] Step S7: The output unit 16 transmits an output (inference result) in response to the inference query to the operation terminal 2.

[0048] [Comparison with the Prior Art] In the prior art, external information related only to the query is searched for, and therefore the acquisition unit 131 and the prior knowledge information storage unit 17 shown in Fig. 1 do not exist. Since step S3 shown in Fig. 6 is not executed, the search words remain "device A, alarm X, countermeasure" extracted in step S2.

[0049] Therefore, in step S4, only "Alarm X of device A is confirmed, device A is restarted, and recovery is complete" is acquired as external information, and the other two pieces of external information are not acquired. In step S6, only "Alarm X of device A will be recovered by restarting" is output as the inference result.

[0050] On the other hand, in this embodiment, in addition to the inference result, a further inference result is output, which is "Device A has a redundant configuration with device B. The normality of device B can be confirmed by the presence or absence of a ping response." As the operational knowledge of a communication network is correct, "When restarting a device, it is necessary to check the status of the redundant pair," it can be said that including this further inference result is a more appropriate inference result.

[0051] [Effects] According to this embodiment, external information related to both a query regarding the operation of a communication network and prior knowledge information about the communication network is searched for, so that a technique capable of sufficiently acquiring external information can be provided.

[0052] [Others] The present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the present disclosure.

[0053] The inference device 1 of the present embodiment described above can be realized, for example, by using a general-purpose computer system including a CPU 901, a memory 902, a storage 903, a communication device 904, an input device 905, and an output device 906, as shown in Figure 7. The memory 902 and the storage 903 are storage devices. In this computer system, the CPU 901 executes a predetermined program loaded onto the memory 902, thereby realizing each function of the inference device 1.

[0054] The inference apparatus 1 may be implemented by a single computer. The inference apparatus 1 may be implemented by multiple computers. The inference apparatus 1 may be a virtual machine implemented on a computer.

[0055] The program for the inference device 1 can be stored in a computer-readable recording medium such as a HDD, SSD, USB memory, CD, or DVD. The computer-readable recording medium is, for example, a non-transitory recording medium. The program for the inference device 1 can also be distributed via a communication network.

[0056] REFERENCE SIGNS LIST 1 Inference device 11 Input unit 12 Extraction unit 13 Search unit 14 Generation unit 15 Inference unit 16 Output unit 17 Prior knowledge information storage unit 18 External information storage unit 19 Internal knowledge information storage unit 131 Acquisition unit 132 Search unit 2 Operation terminal 901 CPU 902 Memory 903 Storage 904 Communication device 905 Input device 906 Output device

Claims

1. An inference device comprising: a search unit that searches for external information related to both a query regarding the operation of a communication network and prior knowledge information of the communication network; and an inference unit that infers an output for the query and the external information from internal knowledge information.

2. The inference device according to claim 1, wherein the prior knowledge information is configuration information of the communication network or operation information of the communication network.

3. An inference method performed by an inference device, the inference method comprising: searching for external information related to both a query regarding the operation of a communication network and prior knowledge information of the communication network; and inferring an output for the query and the external information from internal knowledge information.

Images