Causal inference device, causal inference method, and causal inference program

Abstract

To generate processes between a certain event and its result with high comprehensiveness on the basis of them.SOLUTION: A causal inference device includes a prompt generation unit and a prompt transmission unit. The prompt generation unit sets a part of a process that should be more comprehensive as another candidate generation part in a causal process that represents a causal flow from an event through one or more processes to an outcome to generate a prompt to cause a language model to generate another candidate that can be replaced by the other candidate generation part. The prompt transmission unit transmits the prompt to the language model.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present disclosure relates to a causal inference device, a causal inference method, and a causal inference program. [Background technology]

[0002] As society places more importance on well-being, companies have come to be evaluated based on their overall corporate activities, taking into account the impact that companies have on society and the environment from the perspective of social welfare. In particular, from the perspective of investors who invest in companies, ESG (Environment Social Governance) investment, which places emphasis on companies that address environmental, social, and corporate governance in addition to financial indicators, is attracting attention. Investors who make such ESG investments are called ESG investors. Companies often disclose ESG information for ESG investors. As ESG information, companies are required to report, in particular, on the positive or negative impact that their own business may have on society from an ESG perspective. A document disclosing ESG information is called an ESG disclosure document. Creating such an ESG disclosure document requires a significant amount of work.

[0003] Therefore, it is being considered to automatically create ESG disclosure documents using large-scale language models (Non-Patent Document 1), which have been developing rapidly in recent years. In particular, it is being considered to use large-scale language models to predict what positive or negative impact a company's business may have on society from an ESG perspective. Specifically, a system is being considered that receives text describing a company's business operations, inputs it into a large-scale language model, and outputs text describing how the business will affect ESG indicators. ESG indicators are indicators that measure the impact on society from an ESG perspective. Examples of ESG indicators include greenhouse gas emissions and their reduction.

[0004] Non-Patent Document 2 is a document disclosing an example of a means for generating the effect of things using a large-scale language model. Non-Patent Document 2 discloses a technology in which a text representing an event and a text representing its result are input to a large-scale language model, and a causal process from the event to the result is generated.

[0005] In general, the accuracy of output from a large-scale language model can be improved by devising the text to be input (hereinafter referred to as a prompt). Patent Document 1 is a document that discloses an example of a means for devising a prompt to be input to a large-scale language model. The technology disclosed in Patent Document 1 is a technology that uses a dialogue bot to devise and incorporate a user's reply into a prompt. Specifically, the technology disclosed searches the database of the persona for dialogue scenes related to the topic of the conversation from the user's reply, and incorporates the "topic" and "dialogue scene" into the prompt. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2022-180282 A [Non-patent literature]

[0007] [Non-Patent Document 1] Brown, Tom, et al. "Language models are few-shot learners." Advances in neural information processing systems 33 (2020): 1877-1901. [Non-Patent Document 2] Bhagavatula, Chandra, et al. "Abductive commonsense reasoning." arXiv preprint arXiv:1908.05739 (2019). Summary of the Invention [Problem to be solved by the invention]

[0008] Specifically, the technology disclosed in the above-mentioned non-patent document 2 inputs the text of an event and the text of a result into a large-scale language model, solves an optimization problem with likelihood as the objective function and latent variables as variables, and generates plausible text (causal process) that connects the causal relationship from the event to the result.

[0009] However, there is not necessarily only one causal process from an event to a result. For example, suppose that a railway company developed an indoor simulator for safe driving and trained its employees using the indoor simulator, which led to the result of a reduction in greenhouse gas emissions. In this case, a causal process is possible in which training using the simulator reduced railway accidents, which reduced the spillage of fuels such as coal carried by the accident vehicles, which avoided energy waste, and thus reduced greenhouse gas emissions. Another causal process is possible in which training using the simulator reduced railway accidents, which reduced the resources consumed to clean the accident vehicles, and thus reduced greenhouse gas emissions. Another causal process is possible in which training using the simulator reduced railway accidents, which reduced the number of emergency vehicles dispatched to evacuate passengers in the event of an accident, which reduced energy consumption, and thus reduced greenhouse gas emissions.

[0010] However, in Non-Patent Document 2, due to the structure of solving the optimization problem for the objective function, only one optimal causal process can be obtained, and it is not possible to cover multiple causal processes that may exist.

[0011] In view of the above-mentioned problems, one objective of the present disclosure is to provide a technology for comprehensively generating a process between an event and its result based on the event and its result. [Means for solving the problem]

[0012] A causal inference device according to one aspect of the present disclosure includes: a causal process that represents a flow of causal relationships from an event through one or more processes to a result; a prompt generation unit that generates a prompt for a language model to generate an alternative candidate that can be replaced by the alternative candidate generation point, the causal process representing a flow of causal relationships from an event through one or more processes to a result; and a prompt sending unit that sends the prompt to the language model. Effect of the Invention

[0013] The present disclosure allows for a comprehensive list of possible causal relationships between events and outcomes. [Brief description of the drawings]

[0014] [Figure 1] FIG. 1 illustrates a computer system for implementing an embodiment of the present disclosure. [Diagram 2] FIG. 2 is a diagram illustrating an example of a configuration of a causal inference system according to an embodiment of the present disclosure. [Diagram 3] FIG. 3 is a flowchart showing the flow of a causal inference method according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a diagram illustrating an improvement in coverage of causal processes according to an embodiment of the present disclosure. [Diagram 5] FIG. 5 is a diagram illustrating an improvement in coverage of causal processes according to an embodiment of the present disclosure. [Figure 6] FIG. 6 is a diagram illustrating an improvement in coverage of causal processes according to an embodiment of the present disclosure. [Figure 7] FIG. 7 is a diagram showing a specific example of a prompt that may be used in an embodiment of the present disclosure. [Figure 8] FIG. 8 illustrates an example of a prompt generated by an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Note that the present invention is not limited to this embodiment. In addition, in the description of the drawings, the same parts are indicated by the same reference numerals. Note that, although terms such as "first", "second", and "third" may be used to describe various elements or components in this disclosure, it will be understood that these elements or components should not be limited by these terms. These terms are used only to distinguish one element or component from another element or component. Thus, a first element or component discussed below can also be referred to as a second element or component without departing from the teaching of the inventive concept.

[0016] Next, referring to Fig. 1, a computer system 100 for implementing an embodiment of the present disclosure will be described. The mechanisms and devices of the various embodiments disclosed herein may be applied to any suitable computing system. The main components of the computer system 100 include one or more processors 102, memory 104, a terminal interface 112, a storage interface 113, an I / O (input / output) device interface 114, and a network interface 115. These components may be interconnected via a memory bus 106, an I / O bus 108, a bus interface unit 109, and an I / O bus interface unit 110.

[0017] Computer system 100 may include one or more general purpose programmable central processing units (CPUs) 102A and 102B, collectively referred to as processors 102. In some embodiments, computer system 100 may include multiple processors, while in other embodiments, computer system 100 may be a single CPU system. Each processor 102 executes instructions stored in memory 104 and may include an on-board cache.

[0018] In one embodiment, memory 104 may include random access semiconductor memory, storage devices, or storage media (either volatile or non-volatile) for storing data and programs. Memory 104 may store all or a portion of the programs, modules, and data structures that implement the functions described herein. For example, memory 104 may store a causal inference application 150. In one embodiment, causal inference application 150 may include instructions or descriptions that execute on processor 102 the functions described below.

[0019] In some embodiments, causal inference application 150 may be implemented in hardware via semiconductor devices, chips, logic gates, circuits, circuit cards, and / or other physical hardware devices instead of or in addition to a processor-based system. In some embodiments, causal inference application 150 may include data other than instructions or descriptions. In some embodiments, cameras, sensors, or other data input devices (not shown) may be provided to communicate directly with bus interface unit 109, processor 102, or other hardware of computer system 100.

[0020] Computer system 100 may include a bus interface unit 109 that provides communication between processor 102, memory 104, display system 124, and I / O bus interface unit 110. I / O bus interface unit 110 may couple to an I / O bus 108 for transferring data to and from various I / O units. I / O bus interface unit 110 may communicate via I / O bus 108 with a number of terminal interfaces 112, 113, 114, and 115, also known as I / O processors (IOPs) or I / O adapters (IOAs).

[0021] The display system 124 may include a display controller, a display memory, or both. The display controller can provide video, audio, or both data to the display device 126.

[0022] The display system 124 may be connected to a display device 126, such as a stand-alone display screen, a television, a tablet, or a handheld device.

[0023] The I / O interface unit provides the ability to communicate with various storage or I / O devices. For example, the terminal interface 112 may be attached to user I / O devices 116, such as user output devices, such as a video display, a television with speakers, and user input devices, such as a keyboard, a mouse, a keypad, a touchpad, a trackball, buttons, a light pen, or other pointing device. A user may use the user interface to enter input data or instructions to the user I / O devices 116 and the computer system 100, and receive output data from the computer system 100, by manipulating the user input devices. The user interface may be displayed on a display, played through speakers, or printed via a printer, for example, via the user I / O devices 116.

[0024] Storage interface 113 allows for attachment of one or more disk drives or direct access storage device 117 (usually a magnetic disk drive storage device, but may be an array of disk drives or other storage devices configured to appear as a single disk drive). In some embodiments, storage device 117 may be implemented as any secondary storage device. Contents of memory 104 may be stored in storage device 117 and retrieved from storage device 117 as needed. I / O device interface 114 may provide an interface to other I / O devices such as printers, fax machines, etc. Network interface 115 may provide a communications path to allow computer system 100 and other devices to communicate with each other. This communications path may be, for example, network 130.

[0025] In one embodiment, computer system 100 may be a device that receives requests from other computer systems (clients) without a direct user interface, such as a multi-user mainframe computer system, a single-user system, or a server computer. In another embodiment, computer system 100 may be implemented by virtual computer resources provided on a cloud. In another embodiment, computer system 100 may be a desktop computer, a portable computer, a laptop, a tablet computer, a pocket computer, a telephone, a smartphone, or any other suitable electronic device.

[0026] Next, a configuration of a causal inference system 200 according to an embodiment of the present disclosure will be described with reference to FIG.

[0027] 2 is a diagram illustrating an example of a configuration of a causal inference system 200 according to an embodiment of the present disclosure. As illustrated in FIG 2, the causal inference system 200 includes a prompt creation unit 210, a large-scale language model 220, a communication network 230, an initial causal process generation unit 240, and a communication network 250.

[0028] The prompt creation unit 210 includes a designation receiving unit 211 , a command text generation unit 212 , an example text generation unit 213 , a question text generation unit 214 , a prompt generation unit 215 , a prompt transmission unit 216 , and a presentation unit 217 .

[0029] The prompt creation unit 210 and the large-scale language model 220 are connected to each other via a communication network 230. As the large-scale language model 220, for example, GPT (registered trademark) 2, GPT3, GPT3.5, GPT4, BERT, or PaLM can be used.

[0030] The initial causal process generator 240 and the large-scale language model 220 are connected to each other via a communication network 250 .

[0031] Communications network 230 and communications network 250 may include, for example, a local area network (LAN), a wide area network (WAN), a satellite network, a cable network, a WiFi network, or any combination thereof.

[0032] The initial causal process generation unit 240 generates a "causal process" that represents a flow of causal relationships starting from an "event" through one or more "processes" to a "result". For example, the initial causal process generation unit 240 may be a unit that is given an "event" and a "result" and estimates a "process" that exists between the "event" and the "result" to complete the "causal process". The initial causal process generation unit 240 is a functional unit that generates a causal process before expanding the comprehensiveness. As an example, the initial causal process generation unit 240 generates one or more flows of a series of causal processes that lead from one event to one result through one or more processes. The causal process can be expressed, for example, in a form in which the existence and direction of causal relationships between the event, process, and result are linked by a symbol such as "→" (arrow). As an example, the initial causal process generation unit 240 may input a prompt including an event and a result to a large-scale language model generated in advance, and output text indicating a causal process including the event, process, and result. Hereinafter, text showing a causal process including an event, a process, and a result may be referred to as a "causal process text." FIG. 7 illustrates an example of a prompt input to the large-scale language model and a causal process text output from the large-scale language model. In the prompt in FIG. 7, a question for causing the large-scale language model to generate a causal process is Q2, and a causal process text shown by the causal process generated by the large-scale language model is A2. Note that the initial causal process generation unit 240 may prompt the user, via a GUI or the like, to input a causal process text including an event, a process, and a result, or a pair of a text corresponding to the prompt and a causal process text.

[0033] The designation receiving unit 211 is a functional unit that receives a user's designation of an alternative candidate generation location from within the causal process text generated by the initial causal process generating unit 240. An alternative candidate generation location is a location designated by the user as a process to be replaced with another process when it is desired to generate another possible causal process in which a part of a process included in a causal process is replaced with another process. For example, an implementation may be implemented in which an ID is assigned to each of the events, processes, and results linked by the symbol "→" in the causal process text, and the ID of the process designated by the user as the alternative candidate generation location is stored. The designation of a process may be received by, for example, clicking a mouse on the screen.

[0034] The instruction text generating unit 212 generates an instruction text that instructs the large-scale language model to replace an alternative candidate generation portion, which is at least one of the processes included in the given causal process, with one or more alternative candidates. The instruction text generating unit 212 is a functional unit that generates an incomplete causal process text in which an alternative candidate generation portion, which is received by the above-mentioned designation receiving unit 211 in the causal process text, is replaced with an arbitrary symbol such as "?". "?" (question mark) is an example of a symbol representing a space for inserting one or more alternative candidates. In the above implementation of assigning IDs, the process portion corresponding to the ID of the process specified by the user is deleted and a symbol such as "?" is overwritten in the same place to generate the ID. FIG. 8 shows an example of a prompt input to the large-scale language model and a causal process text output from the large-scale language model. The part of the last line starting with "CP's training facilities" in the prompt in FIG. 8 is an example of an instruction text generated by the instruction text generating unit 212. The instruction text shows the causal process and the alternative candidate generation portion in the causal process.

[0035] The example text generating unit 213 is a functional unit that creates an example part in the prompt to be given to the large-scale language model. The role of the example is to specify the description format of the causal process text generated by the large-scale language model. In other words, the example is inserted into the prompt to lead the output of the large-scale language model to the intended format. In the example of FIG. 8, Q1 and A1 shown as examples in the prompt are examples of examples generated by the example text generating unit 213. For example, as shown in the prompt of FIG. 8, it may be specified to fill in the ? part and list multiple possible processes in a format such as 1), 2), 3). The example text generating unit 213 may store examples created in advance by the user or the provider of this system in a text format in memory, and read and use the text when creating an example. In that case, the user or the like may create an example using the large-scale language model.

[0036] The question text generation unit 214 is a functional unit that generates a question asking about a causal process linking an event and a result. The question text generation unit 214 receives a text describing an event and a text describing a result, and substitutes them into a template of a question asking about a causal relationship that has been created in advance. For example, a question can be created by substituting the input event text into the {event} section and the result text into the {result} section for a template of a question created in advance, such as "As a result of {event}, {result} occurred. Please list multiple events that fit the ? part of the following sentence describing the causal relationship." In the example of FIG. 8, in the prompt Q2, all the sentences before the final sentence starting with "CP's training facilities" are examples of questions generated by the question text generation unit 214.

[0037] The prompt generation unit 215 is a functional unit that generates a text called a prompt by concatenating the causal process text with "?" generated by the instruction text generation unit 212, the example text generated by the example text generation unit 213, and the question text generated by the question text generation unit 214. The order of concatenation is thought to be the example text, the question text, and the causal process text with "?", but is not limited to this order. The prompt in FIG. 8 is an example of a prompt generated by the prompt generation unit 215.

[0038] Prompt sending unit 216 is a functional unit that sends the prompt generated by prompt generating unit 215 to the large-scale language model. The large-scale language model is usually stored on a server different from that of prompt creating unit 210, but may be stored on the same server.

[0039] The presentation unit 217 is a functional unit that presents to the user the text generated from the large-scale language model in response to the prompt transmitted by the prompt transmission unit 216. If the user determines that the causal process is not sufficiently comprehensive after the text is presented to the user, the specification reception unit 211 may receive further specifications of other candidate generation locations, and the generation of the causal process may be repeated.

[0040] The causal inference system 200 having the configuration described above can provide a means for comprehensively listing possible causal relationships between events and results by recursively generating them while involving the user.

[0041] Next, a causal inference method according to an embodiment of the present disclosure will be described with reference to FIG.

[0042] 3 is a flowchart showing the flow of a causal inference method 300 according to the first embodiment of the present disclosure. The causal inference method 300 shown in Fig. 3 is a method for increasing comprehensiveness by recursively generating possible causal relationships between events and results while involving a user, and is implemented by each device and functional unit of the causal inference system 200 shown in Fig. 2.

[0043] First, in step S305, initial causal process generation section 240 generates text in which a causal process between an event and a result is described using symbols such as "→" using a user input or a large-scale language model.

[0044] Next, in step S310, the presentation unit 217 presents the text generated in step S305 to the user and accepts the user's evaluation as to whether the generated process has sufficient comprehensiveness. If it is determined that there is sufficient comprehensiveness, the generated causal process at that point in time is presented as a product, and the system operation ends. If it is determined that there is insufficient comprehensiveness, the process proceeds to step S315.

[0045] Next, in step S315, the designation receiving unit 211 receives a designation by the user of an alternative candidate generation location.

[0046] Next, in step S320, the instruction text generating unit 212 generates an instruction text instructing to replace the alternative candidate generation portion with the alternative candidate.

[0047] Next, in step S325, the example text generation unit 213 generates an example text for specifying the description format of the causal process text for the large-scale language model.

[0048] Next, in step S330, the question text generator 214 generates a question text asking about the causal process linking the event and the result.

[0049] Next, in step S335, the prompt generation unit 215 concatenates the example text generated by the example text generation unit 213 described above, the question text generated by the question text generation unit 214 described above, and the other candidate generation location designation text generated by the instruction text generation unit 212 described above, to generate text that will be the prompt.

[0050] Next, in step S340, the prompt sending unit 216 inputs the prompt generated in step S335 into the large scale language model.

[0051] Next, returning to step S310, the presentation unit 217 presents the output of the large-scale language model to the user and accepts the user's evaluation as to whether the generated process has sufficient comprehensiveness. If it is determined that there is sufficient comprehensiveness, the generated causal process at that point in time is presented as a product, and the series of operations of the system is terminated. If it is determined that there is insufficient comprehensiveness, the process proceeds to step S315, and the above steps are repeated until the user evaluates that there is sufficient comprehensiveness.

[0052] According to the causal inference method 300 described above, possible causal relationships between events and results can be recursively generated while involving the user, thereby making it possible to exhaustively list them.

[0053] Next, an improvement in coverage of causal processes according to an embodiment of the present disclosure will be described with reference to FIGS.

[0054] 4 to 6 are diagrams showing an improvement in the comprehensiveness of causal processes according to an embodiment of the present disclosure. FIG. 4 shows an initial reference causal process generated by a large-scale language model or user input in step S305 of FIG. 3. Each process is represented by a capital alphabet such as "A" or "B." Here, A represents an event, and D represents a result. The other alphabets represent processes in between. A causal relationship between an event, a process, and a result is represented by "→". As in the example of FIG. 4, there may be multiple initial causal processes.

[0055] Fig. 5 shows a state in which the process "C" specified as the alternative candidate generation location is replaced with alternative candidates "C1, C2" by the processing of steps S315 to S340 in Fig. 3 from the state in Fig. 4. In this schematic diagram, there are two alternative candidates, but the number of alternative candidates is not particularly limited. In addition, the alternative candidate generation location is arbitrary and is not limited to the process "C".

[0056] Fig. 6 shows the state after the processing of steps S315 to S340 in Fig. 3 is performed again from the state in Fig. 5. In the example in Fig. 6, after other candidates "C1, C2" are generated for process "C", processing is performed to further increase the comprehensiveness of process "B" for the causal process passing through each of "C1" and "C2". Other candidates "B11, B12, B13" are generated for process "B" of the causal process "A → B → C1 → D". Other candidates "B21, B22, B23" are generated for process "B" of the causal process "A → B → C2 → D".

[0057] In this manner, the processes from S315 to S340 in FIG. 3 can be repeated until the user determines in S310 in FIG. 3 that sufficient comprehensiveness has been obtained.

[0058] Although the embodiment of the present invention has been described above, the present invention is not limited to the above-described embodiment, and various modifications are possible without departing from the gist of the present invention.

[0059] The above-described embodiment of the present invention includes the following features. However, the features included in the embodiment are not limited to the following features.

[0060] (Item 1) a prompt generating unit that generates a prompt for causing a language model to generate alternative candidates that can be replaced with the alternative candidate generation portion, the alternative candidate generation portion being a portion of a causal process that represents a flow of causal relations from an event through one or more processes to a result; and According to this, by replacing a process included in a causal process with another candidate, it becomes possible to generate processes between an event and a result with high comprehensiveness.

[0061] (Item 2) The causal inference device according to item 1 further includes an instruction text generation unit that generates instruction text indicating the causal process and the other candidate generation location in the causal process, and a question text generation unit that generates question text for asking the language model about other candidates that can be replaced with the other candidate generation location in the causal process, and the prompt generation unit generates the prompt so as to include the question text and the instruction text. This makes it possible to satisfactorily generate other candidates by the prompt including the question text and the instruction text.

[0062] (Item 3) The causal inference device according to item 2 further includes an example text generation unit that generates an example text by pairing an input text example to be input to the language model with an output text example in a desired output format, and the prompt generation unit generates the prompt by concatenating the question text, the instruction text, and the example text. This makes it possible to derive the output text of the language model into the desired output format by presenting an example that illustrates an output in the desired output format to the language model.

[0063] (Item 4) In the causal inference device described in item 3, the prompt generation unit generates the prompt by concatenating the example text, the question text, and the instruction text in that order. This makes it possible to correctly guide the output of the language model in line with the purpose of generating alternative candidates in a causal process by first providing an example to the language model, then asking a question, and then presenting a causal relationship to be queried and an alternative candidate generation location.

[0064] (Item 5) The causal inference device according to claim 2 further includes a presentation unit that presents a text of the causal process output from the language model to a user, and a designation receiving unit that receives from the user a designation of an alternative candidate generation location in the presented causal process, and the instruction text generation unit generates the instruction text from the causal process in accordance with the designation of the alternative candidate generation location. In this way, the causal process is presented to the user and the designation of the alternative candidate generation location is received, so that the user can easily designate the alternative candidate generation location.

[0065] (Item 6) In the causal inference device described in item 5, the designation receiving unit records identification information of a process designated at the alternative candidate generation location, and the instruction text generating unit identifies the alternative candidate generation location by referring to the identification information. This makes it easy to identify the alternative candidate generation location by identifying a process included in a causal process by the identification information.

[0066] (Item 7) The causal inference device according to item 3 further comprises a causal process generation unit that generates a causal process in a format in which a text is divided into units of an event, a process, and a result, and the causal relationship existing between the event, the process, and the result and its direction are expressed by a first symbol, the instruction text generation unit generates the instruction text by replacing a process to be an alternative candidate generation point among one or more processes divided by the first symbol in the causal process with a second symbol, and the question text generation unit generates a question text for replacing the location of the second symbol in the instruction text with an alternative candidate. According to this, by expressing the causal relationship constituting the causal process and the alternative candidate generation point in the causal process by symbols, it is possible to generate a prompt that is easy to understand.

[0067] (Item 8) In the causal inference device described in item 7, the example text generation unit generates an example text by pairing the input text example in which an alternative candidate generation portion in a causal process in which a causal relationship is expressed by the first symbol is replaced by the second symbol, and the output text example in which a plurality of processes in which the alternative candidate generation portion in the causal process can be replaced are listed in the output format. According to this, by expressing the example using symbols as well as the causal process to be improved in comprehensiveness, it is possible to generate a prompt that is easy to understand, including the example.

[0068] (Item 9) In the causal inference device according to item 7, the first symbol is an arrow and the second symbol is a question mark. This allows the arrow and the question mark to be used to express a prompt that is intuitively understandable to the user and the language model.

[0069] (Item 10) In the causal inference device described in item 1, the contents of events and results for which a comprehensive causal process is to be generated are stored in advance inside the system or are received from a user's input. This makes it possible to easily store or receive the targets for which a comprehensive causal process is to be generated by events and results.

[0070] (Item 11) The computer executes the following steps: in a causal process that represents the flow of causal relationships from an event through one or more processes to a result, a point in the process where comprehensiveness should be improved is designated as an alternative candidate generation point; a prompt is generated to cause a language model to generate an alternative candidate that can be replaced with the alternative candidate generation point; and the prompt is sent to the language model.

[0071] (Item 12) The causal inference program causes a computer to execute the following steps: designate a portion of a causal process that represents the flow of causal relationships from an event through one or more processes to a result as an alternative candidate generation portion, where comprehensiveness should be improved, generate a prompt to cause a language model to generate an alternative candidate that can be replaced with the alternative candidate generation portion, and send the prompt to the language model. [Explanation of symbols]

[0072] 210: prompt creation unit, 211: specification reception unit, 212: instruction text generation unit, 213: example text generation unit, 214: question text generation unit, 215: prompt generation unit, 216: prompt transmission unit, 217: presentation unit, 220: large-scale language model, 240: initial causal process generation unit

Claims

1. a prompt generating unit that generates a prompt for causing a language model to generate alternative candidates that can be replaced with the alternative candidate generation portion, the alternative candidate generation portion being a portion of a causal process that represents a flow of causal relations from an event through one or more processes to a result, and that is configured to generate an alternative candidate that can be replaced with the alternative candidate generation portion; a prompt sending unit for sending the prompt to the language model; A causal inference device comprising:

2. an instruction text generation unit that generates an instruction text indicating the causal process and the other candidate generation portion in the causal process; a question text generation unit that generates a question text for querying the language model about alternative candidates that can be replaced with the alternative candidate generation portion in the causal process; Further comprising: the prompt generation unit generates the prompt so as to include the question text and the instruction text. The causal inference device according to claim 1 .

3. An example text generating unit generates an example text by pairing an input text example to be input to the language model with an output text example in a desired output format, the prompt generation unit generates the prompt by concatenating the question text, the instruction text, and the example text. The causal inference device according to claim 2.

4. the prompt generation unit generates the prompt by concatenating the example text, the question text, and the instruction text in that order. The causal inference device according to claim 3.

5. a presentation unit that presents a text of the causal process output from the language model to a user; and a designation receiving unit that receives, from the user, a designation of an alternative candidate generation location in the presented causal process, the instruction text generation unit generates the instruction text from the causal process in accordance with the specification of the alternative candidate generation portion. The causal inference device according to claim 2.

6. The designation receiving unit records identification information of a process designated in the alternative candidate generation portion, The instruction text generation unit refers to the identification information to identify the alternative candidate generation portion. The causal inference device according to claim 5.

7. a causal process generation unit that divides the text into units of an event, a process, and a result, and generates a causal process in a format in which a causal relationship existing between the event, the process, and the result and a direction of the causal relationship are expressed by a first symbol; the instruction text generation unit generates the instruction text by replacing a process to be a candidate generation point among one or more processes separated by the first symbol in the causal process with a second symbol; the question text generation unit generates a question text for replacing the portion of the second symbol in the instruction text with another candidate; The causal inference device according to claim 3.

8. the example text generation unit generates an example text by pairing the input text example in which an alternative candidate generation portion in a causal process in which a causal relationship is expressed by the first symbol is replaced by the second symbol, and the output text example in which a plurality of processes in which the alternative candidate generation portion in the causal process can be replaced are listed in the output format. The causal inference device according to claim 7.

9. the first symbol is an arrow; the second symbol is a question mark; The causal inference device according to claim 7.

10. The contents of events and outcomes for which comprehensive causal processes are to be generated are either stored in the system beforehand or accepted from user input. The causal inference device according to claim 1 .

11. a causal process representing a flow of causal relations from an event through one or more processes to a result, the causal process representing a flow of causal relations from an event through one or more processes being a process in which comprehensiveness should be improved is set as an alternative candidate generation portion, and a prompt is generated for causing a language model to generate an alternative candidate that can be replaced with the alternative candidate generation portion; sending the prompt to the language model; A causal inference method implemented by a computer.

12. a causal process representing a flow of causal relations from an event through one or more processes to a result, the causal process representing a flow of causal relations from an event through one or more processes being a process in which comprehensiveness should be improved is set as an alternative candidate generation portion, and a prompt is generated for causing a language model to generate an alternative candidate that can be replaced with the alternative candidate generation portion; sending the prompt to the language model; A causal inference program that allows a computer to carry out this task.

Images