Work procedure generation device and program

Abstract

A work procedure generation device according to the present invention comprises: an input unit configured to input first text that indicates a request for generating a work procedure pertaining to a certain task; a search unit configured to acquire similar information that is similar to the first text from among information related to the certain task; an input text generation unit configured to generate second text for requesting generation of a response to the first text with reference to the similar information, and input the second text to a machine learning model that has been trained to generate text corresponding to input text; and an output unit configured to output the text generated by the machine learning model. Thus, the work procedure generation device assists preparation of a work procedure manual pertaining to the task.

Description

Work procedure generation device and program

[0001] The present invention relates to a work procedure generation device and a program.

[0002] The maintenance and operation of network infrastructure, which interconnects numerous devices, involves extremely complex procedures. Because networks are made up of a complex combination of multiple devices and it is often difficult to shut down the system, even simple tasks can easily involve a large number of steps.

[0003] For example, when updating a router on a network, an operator must perform three steps: reroute communications that go through that router to a detour router, update the equipment, and restore the original route. Even within these three major steps, there are many detailed tasks involved, such as verifying the accuracy of the routing settings and ensuring that the router is operating normally. Additionally, in the unlikely event of an abnormality, the router settings must be restored to their original state, requiring a rollback procedure.

[0004] As described above, the maintenance and operation of network infrastructure involves many work procedures, and creating work procedure manuals places a heavy burden on operators.

[0005] Another factor that makes creating work procedure manuals a burden is the large number of parameters involved in work procedures. In the case of network systems, parameters related to work include device names, IP addresses, port numbers, etc. These are parameters for each connected device, so the number is enormous. Even when performing the same work on multiple devices, operators must modify the parameters in the work procedure manual for each device they work on. This type of modification work is difficult for inexperienced operators, but on the other hand, human error, such as forgetting to replace IP addresses, is likely to occur, posing a risk of causing a major malfunction.

[0006] When performing unexpected maintenance such as responding to a breakdown, operators refer to a response manual to perform the maintenance. For breakdowns that occur infrequently, there is often no clear manual, and responses are often decided based on the action history, which is the history of past breakdown responses. In this case, the past action history is not kept in the format of a work procedure manual, so it can be difficult for inexperienced operators to obtain the procedure for responding to a breakdown. Furthermore, even for tasks other than responding to breakdowns, procedures often vary slightly from device to device. For this reason, procedures are often created by referring to examples of similar work procedures from the past. There is often no set format for creating work procedures, so operators must create work procedures that suit the situation by looking at similar procedures from the past.

[0007] To summarize, the complexity of the procedures themselves, the large number of parameters, and the unstructured format of action history make network maintenance and operation highly difficult. To address these challenges, there are a variety of business manual creation tools available on the market. For example, a tool targeted at creating business manuals (Non-Patent Document 1) records computer operations and generates work procedure manuals based on the records, enabling the semi-automatic creation of work procedure manuals.

[0008] "Business manual creation tool 'Dojo'," [online], Tenda Co., Ltd., [Retrieved November 21, 2023], Internet, <URL: https: / / tepss.com / >

[0009] However, in the operation procedures for IT systems, different parameters are used for each operation, and the man-hours required for the procedure vary depending on the number of devices involved.

[0010] Because business manual creation tools focus on creating generalized procedures, it is difficult to generate work procedure manuals that contain parameters and work procedures specific to a particular case, as is required in the operation of modern IT systems.

[0011] The present invention has been made in view of the above points, and has an object to support the creation of work procedure manuals related to business.

[0012] In order to solve the above problem, the work procedure generation device has an input unit configured to input first text indicating a request to generate a work procedure for a certain task, a search unit configured to acquire similar information that is similar to the first text from information related to the certain task, an input text generation unit configured to generate second text that requests that a response to the first text be generated by referring to the similar information, and to input the second text to a machine learning model trained to generate text corresponding to the input text, and an output unit configured to output the text generated by the machine learning model.

[0013] We can assist in creating work procedures for business operations.

[0014] FIG. 1 is a diagram illustrating an example of a hardware configuration of a work procedure generation device 10 according to an embodiment of the present invention. FIG. 2 is a diagram illustrating an example of a functional configuration of the work procedure generation device 10 according to an embodiment of the present invention. FIG. 3 is a flowchart for explaining an example of a processing procedure executed by the work procedure generation device 10. FIG. 4 is a diagram illustrating a first example of a work request. FIG. 5 is a diagram illustrating a first example of similarity information. FIG. 6 is a diagram illustrating a first example of a context-attached statement. FIG. 7 is a diagram illustrating an example of a format of a context-attached statement. FIG. 8 is a diagram illustrating a first example of a work procedure manual. FIG. 9 is a diagram illustrating an example of an input / output interface according to the present embodiment. FIG. 10 is a diagram illustrating a second example of a work request. FIG. 11 is a diagram illustrating a second example of similarity information. FIG. 12 is a diagram illustrating a second example of a context-attached statement. FIG. 13 is a diagram illustrating a second example of a work procedure manual.

[0015] An embodiment of the present invention will now be described with reference to the accompanying drawings. Fig. 1 is a diagram showing an example of the hardware configuration of a work procedure generation device 10 according to an embodiment of the present invention. The work procedure generation device 10 shown in Fig. 1 includes a drive device 100, an auxiliary storage device 102, a memory device 103, a processor 104, and an interface device 105, all of which are interconnected via a bus B.

[0016] The program that realizes the processing in the work procedure generation device 10 is provided by a recording medium 101 such as a CD-ROM. When the recording medium 101 storing the program is set in the drive device 100, the program is installed from the recording medium 101 to the auxiliary storage device 102 via the drive device 100. However, the program does not necessarily have to be installed from the recording medium 101, but may be downloaded from another computer via a network. The auxiliary storage device 102 stores the installed program as well as necessary files, data, etc.

[0017] When an instruction to start the program is received, the memory device 103 reads and stores the program from the auxiliary storage device 102. The processor 104 is a CPU or a GPU (Graphics Processing Unit), or a CPU and a GPU, and executes functions related to the work procedure generation device 10 in accordance with the program stored in the memory device 103. The interface device 105 is used as an interface for connecting to a network.

[0018] Fig. 2 is a diagram showing an example of the functional configuration of a work procedure generation device 10 according to an embodiment of the present invention. In Fig. 2, the work procedure generation device 10 includes an input unit 11, a search unit 12, an input text generation unit 13, a response unit 14, and an output unit 15. Each of these units is realized by a process executed by a processor 104 of one or more programs installed in the work procedure generation device 10. Furthermore, each of these units may function on a single computer (a physical or virtual computer) or may function in a distributed manner across multiple computers. In other words, the work procedure generation device 10 may be configured by multiple computers.

[0019] The input unit 11 accepts input of text indicating a request for the creation of a work procedure for a certain task (hereinafter simply referred to as a "work request") from a user. The user may be, for example, an operator who performs maintenance and operation work for a network infrastructure in which many devices are interconnected. A work procedure manual is a document that indicates the procedures for the tasks required for the task.

[0020] The search unit 12 acquires (extracts) information similar to the work request received by the input unit 11 (hereinafter referred to as "similar information") from a knowledge DB 20, which is a collection of specialized knowledge related to the user's work. Here, specialized knowledge is a general term for information related to the work procedure that the operator, who is the user in this embodiment, wants to execute. There may be one or more knowledge DBs 20. FIG. 2 shows a work procedure manual DB 21, a measure history DB 22, and a manual DB 23 as examples of the knowledge DB 20. Each of these databases can be realized using, for example, an auxiliary storage device 102 or an external system such as a storage device connectable to the work procedure generation device 10 via a network.

[0021] The work procedure manual DB 21 is a database that stores existing work procedure manuals. The action history DB 22 is a database that records the work content when responding to past failures. The manual DB 23 is a database that accumulates maintenance response manuals for each alert message of the operation system.

[0022] The input text generating unit 13 generates text requesting that a response to the work request be generated by referring to similar information, and inputs the generated text to the response unit 14 .

[0023] The response unit 14 is a machine learning model trained to generate text corresponding to input text. Specifically, the response unit 14 is an LLM such as chatGPT or Llama. The response unit 14 performs calculations based on learned parameters on the text input by the input text generation unit 13 to generate appropriate text (work procedure manual) as a response. Note that the LLM is a large language model. Using an LLM addresses the case-by-case atypical nature of work procedure manual generation. However, as long as the requirement that the input and output are text is met, a simple language model other than an LLM, such as Transformer or its advanced form, BERT, may also be used as the response unit 14.

[0024] The output unit 15 outputs the text (work procedure manual) generated by the response unit 14 .

[0025] The following describes the processing procedure executed by the work procedure generation device 10. FIG.

[0026] In step S101, the input unit 11 accepts an input of a work request from a user.

[0027] Figure 4 is a diagram showing a first example of a work request. Work request q1 shown in Figure 4 is text indicating the following contents: - A work procedure manual is requested to be created for adding two units of device A. - For the format of the work procedure manual, refer to the January 2022 addition procedure manual. - The names of the two devices to be added are a03 and a04, respectively. - The IP addresses of the two devices to be added are 192.168.1.103 and 192.168.1.104, respectively. - The IP addresses of the connection destinations of the two devices to be added are 192.168.1.1 and 192.168.1.2, respectively.

[0028] Here, the device name, IP address, etc. are examples of parameter values ​​related to the device that is the target of the business.

[0029] Next, based on the work request, the search unit 12 acquires documents with a score of 1 or more from the top of the document group stored in each knowledge DB 20 as similar information to the work request (S102).

[0030] In a typical document search, the score of each document is the degree of similarity between the query entered in the search request and the document. Similarly, the search unit 12 uses a work request as a query and acquires, as similar information, some documents that have a relatively high degree of similarity to the query.

[0031] 5 is a diagram showing a first example of similar information. Fig. 5 shows the file names and contents (description contents of the file) of documents (work procedure manuals) acquired as similar information. Fig. 5 shows an example in which a document d1 with the file name "Device A Addition_20220120.docx" and a document d2 with the file name "Device A 2 Units Added_20200310.docx" are acquired as similar information.

[0032] Both documents are work procedure manuals obtained from the work procedure manual DB21, but in this example, it is assumed that the similarity is relatively high because the word sequence in the file name is similar to the word sequence in the "January 2022 expansion work procedure manual" included in the input work request.

[0033] The similarity between a query and a document is generally calculated based on vectorization of the document (see "Reference 1" below), and this embodiment also assumes the use of a similar method.

[0034] In this vectorization technique, a large amount of document data is given, and parameters are pre-trained to convert all documents into vectors with the same number of dimensions based on the frequency and order of the words and characters they contain. At this time, the vectors of each document are trained so that the more similar the words and characters in the documents, the closer the vectors are to each other. Using the trained parameters, any word string (or word set) can be converted into a vector, so all documents in the knowledge DB 20 are vectorized. Query queries are also similarly considered to be word strings and converted into vectors.

[0035] In order to obtain documents that are as similar as possible to the query (work request), the search unit 12 selects documents whose vectors have a relatively large inner product with the vector of the query. The larger the inner product, the closer the frequency and appearance order of words and characters in the two documents.

[0036] Although we have described the vector-based method above, any method that can define the similarity between two texts is acceptable. Alternatively, we can simply create a frequency distribution of characters in a document and select documents that have a frequency distribution relatively close to the frequency distribution of the query.

[0037] As described above, the search unit 12 acquires one or more similar information items. That is, the similar information acquired by the search unit 12 may be one item (the document with the highest score) or multiple items (multiple documents with the highest scores).

[0038] For example, the search unit 12 may retrieve k documents (k is a setting value provided by the system administrator in this embodiment) in order of highest similarity, retrieve an indefinite number of documents that fall into the context described below, or retrieve documents from each of multiple knowledge DBs 20.

[0039] Alternatively, the input unit 11 may accept input of similar information from a user, rather than retrieving the similar information from the knowledge DB 20. In this case, the similar information (or the work request including the similar information) is input together with the work request to the search unit 12. The search unit 12 adds the similar information to the similar information acquired based on the similarity with the work request, and executes the processing described below.

[0040] Next, the input text generation unit 13 classifies the similar information with the highest score among the similar information acquired by the search unit 12 into the format of the work procedure manual, and the remaining similar information into reference information (S103). In the example of Figure 5, if the score of document d1 is higher than that of document d2, document d1 is classified as the format, and document d2 is classified as reference information. However, if only one piece of similar information is acquired, that similar information is classified as the format, and no similar information is classified as reference information.

[0041] The form serves as the base (template) for the work procedure manual to be created, and the output work procedure manual is generated by rewriting this text. Meanwhile, the reference information is used as supplementary information and as a guideline for adding or correcting information to the output. By using the reference information, it is possible to obtain information not included in the form, such as important points to note in previous procedures or the appropriate repetition order when performing multiple procedures, which is expected to improve the accuracy of the generated work procedure manual.

[0042] Next, the input text generating unit 13 generates text for inputting the work request, format, and reference information to the response unit 14 (S104).

[0043] As described above, the response unit 14 is an LLM. Since the LLM accepts text as input, the input text generation unit 13 must pass text containing both the work request and the format (and reference information, if necessary) to the response unit 14.

[0044] Therefore, the input text generation unit 13 generates text to input a sentence with context to the response unit 14. Context is one of the techniques for making the output of the LLM closer to what is desired, and by adding additional information called context to the query, it becomes possible to obtain an output that takes into account the information written in the context.

[0045] In this embodiment, the input text generation unit 13 generates text that combines the context and the question, using the format and reference information as context and the work request as a question, and then generates a context-attached sentence by adding the conditions for generating the work procedure manual to the text.

[0046] Figure 6 is a diagram showing a first example of a context-attached sentence. In Figure 6, the same parts as in Figure 4 or Figure 5 are given the same reference numerals. The context-attached sentence c1 shown in Figure 6 includes document d1 as a form, document d2 as reference information, and work request q1 as a question.

[0047] In the context-attached sentence c1, the description added as [condition] (hereinafter simply referred to as "generation condition") is a fixed phrase added to the context-attached sentence c1. The generation condition includes, for example, a description indicating the role of the form and the role of the reference information, as well as other descriptions indicating the generation condition for the work procedure manual (for example, an instruction to replace the value of a parameter included in the form, which is similar information, with the value included in the work request q1, which is a question). Like the reference information, the generation condition is information added for the purpose of improving the accuracy of the work procedure manual, and is not essential information.

[0048] There is no particular limitation on the format (structure) of the context-attached sentence, but the input text generating unit 13 may generate the context-attached sentence based on, for example, a predetermined format (template).

[0049] Fig. 7 is a diagram showing an example of the format of a context-attached sentence. Fig. 7 shows an example of the format of the context-attached sentence c1 shown in Fig. 6. In this case, the input text generation unit 13 can generate the context-attached sentence c1 by replacing {{style}}, {{reference information N}}, and {{work request}} in the format with similar information as the style, similar information as the Nth reference information, and a work request.

[0050] Next, the input text generating unit 13 inputs the generated context-attached sentence to the response unit 14 (S105).

[0051] The response unit 14 generates text (work procedure manual) as a response to the user in response to the input of the context-added sentence, and outputs the text to the output unit 15 as a response to the work request (S106).

[0052] The output unit 15 outputs the text (S107).

[0053] 8 is a diagram showing a first example of a work procedure manual. A work procedure manual a1 shown in FIG. 8 shows a work procedure corresponding to a work request q1.

[0054] 4 and 8 show the work request q1 and the work procedure manual a1 separately, but on the user interface, for example, input by the user and output from the work procedure generation device 10 may be performed via a single screen g1, as shown in Fig. 9. In Fig. 9, the same parts as in Fig. 4 or 8 are denoted by the same reference numerals.

[0055] In this embodiment, the use of LLM to generate responses makes it possible to accommodate the non-standard nature of work procedure manuals. While work procedure manuals are prone to change depending on procedure parameters, particularly the target model and version, and the number of devices, the use of LLM makes it possible to flexibly modify work procedure manuals depending on the target devices and the situation. For example, in the output example (work procedure manual a1) of Figure 8, the IP address and other device-specific information that varies depending on the device is changed to an appropriate value, and the number of procedures that change depending on the number of devices, etc., is also changed.

[0056] Furthermore, this embodiment overcomes the lack of specialized knowledge that is an issue with LLMs by providing a scheme for searching information from the knowledge DB 20. Because LLMs are general-purpose machine learning systems for text responses, they are unable to retain specialized knowledge, such as work procedures for specific systems in specific organizations. To address this issue, by combining them with a method for retrieving related documents from the knowledge DB 20, it becomes possible to apply them to tasks specific to each organization, such as generating work procedure manuals.

[0057] Furthermore, this embodiment employs a unique technique for generating context-attached sentences when passing information from the input text generation unit 13 to the response unit 14. Specifically, a "distinction (classification) between format and reference information" and generation conditions are added to the context-attached sentences. By clearly distinguishing between format and reference information, the user can freely specify the format of the work procedure manual that will be the basis for creation. On the other hand, reference information allows the addition of information that is not included in the format. The role of each piece of information is specified by conditions, which can improve the accuracy of the created work procedure manual.

[0058] Furthermore, this embodiment can be used for a variety of purposes by changing the knowledge DB 20 from which similar information is obtained and the method for obtaining similar information.

[0059] For example, if images (such as a configuration diagram of a target device) are stored in the knowledge DB 20, and a simple (e.g., handwritten) image (hereinafter referred to as an "input image") is included in a work request from a user, and the work request includes a request such as "I would like to modify the input image to make it easier to understand by referring to past images and include it in a procedure manual," it is necessary to search for similar images from the images. In this case, the search unit 12 may search for similar images based on a task generally called a similar image search.

[0060] Furthermore, while the above example illustrates the creation of work procedures for planned work under normal circumstances, work procedures for responding to malfunctions or abnormalities, which are non-routine tasks, can also be created by using the history of actions taken against similar malfunctions in the past.

[0061] In this case, for example, a work request q2 as shown in Fig. 10 is input in step S101. The work request q2 requests the generation of a work procedure manual for when an error "session request timeout" occurs.

[0062] The search unit 12 retrieves action histories similar to the work request q2 from the action history DB 22 (S102). Compared to work procedure manuals for routine tasks, action histories tend to be more atypical in their descriptions and the work content tends to vary depending on the event. Therefore, when creating a work procedure manual that describes a failure response procedure, it is difficult to generate an appropriate work procedure manual by simply referring to a single action history.

[0063] Therefore, the search unit 12 acquires, as similarity information, a plurality of (at least two) action histories from the action history DB 22 in descending order of similarity to the work request.

[0064] 11 is a diagram showing a second example of similar information, in which two action histories, action history d3 and action history d4, are acquired as similar information.

[0065] The input text generation unit 13 generates a context-attached sentence (S104) using all of the multiple action histories as reference information (in other words, without distinguishing between the multiple action histories as formats and reference information), and inputs the context-attached sentence to the response unit 14 (S105).

[0066] Figure 12 is a diagram showing a second example of a context-attached sentence. In Figure 12, the same parts as in Figure 10 or Figure 11 are assigned the same reference numerals. The context-attached sentence c2 shown in Figure 12 includes document d3 as reference information 1, document d4 as reference information 2, and work request q2 as a question. In the context-attached sentence c2, the creation condition added as [Condition] describes instructions to create a work procedure manual based on a combination of reference information.

[0067] The response unit 14 generates text (work procedure manual) based on such a context-attached sentence, with reference to a plurality of action histories (S106). The output unit 15 outputs the text (S107).

[0068] 13 is a diagram showing a second example of a work procedure manual. A work procedure manual a2 shown in FIG. 13 shows a work procedure corresponding to a work request q2.

[0069] As described above, this embodiment can assist in the creation of work procedure manuals for business operations. As a result, it is possible to efficiently create work procedure manuals suited to the situation, regardless of the operator's level of knowledge. For example, this can be expected to speed up work in the maintenance and operation of IT systems and reduce the frequency of work errors.

[0070] [Reference 1] Lewis, Patrick, et al. "Retrieval-augmented generation for knowledge-intensive nlp tasks." Advances in Neural Information Processing Systems 33 (2020): 9459-9474 The above describes in detail the embodiments of the present invention, but the present invention is not limited to such specific embodiments, and various modifications and variations are possible within the scope of the gist of the present invention described in the claims.

[0071] REFERENCE SIGNS LIST 10 Work procedure generation device 11 Input unit 12 Search unit 13 Input text generation unit 14 Response unit 15 Output unit 20 Knowledge DB 21 Work procedure manual DB 22 Action history DB 23 Manual DB 100 Drive device 101 Recording medium 102 Auxiliary storage device 103 Memory device 104 Processor 105 Interface device B Bus

Claims

1. A work procedure generation device comprising: an input unit configured to input first text indicating a request to generate a work procedure for a certain task; a search unit configured to retrieve similar information similar to the first text from information related to the certain task; an input text generation unit configured to generate second text requesting that a response to the first text be generated by referring to the similar information, and to input the second text to a machine learning model trained to generate text corresponding to the input text; and an output unit configured to output the text generated by the machine learning model.

2. The work procedure generation device according to claim 1, wherein the machine learning model is a large-scale language model.

3. The work procedure generation device according to claim 1, characterized in that the input text generation unit is configured to acquire a plurality of pieces of similar information, and generate the second text in a format based on one of the plurality of pieces of similar information, and request the generation of the response by referring to the other pieces of similar information.

4. The work procedure generation device described in claim 1, characterized in that the first text includes parameter values related to the equipment that is the target of the work, and the second text includes instructions to replace the parameter values included in the similar information with the values included in the first text.

5. A program that causes a computer to execute the following steps: an input procedure for inputting first text indicating a request for generating work procedures related to a certain task; a search procedure for retrieving similar information similar to the first text from information related to the certain task; an input text generation procedure for generating second text requesting that a response to the first text be generated by referring to the similar information, and inputting the second text into a machine learning model that has been trained to generate text corresponding to the input text; and an output procedure for outputting the text generated by the machine learning model.

Images