Language processing system, language processing method, and language processing program

Abstract

An execution unit of the present invention provides a language model with a first query character string including the name of a function and instructing the model to respond according to the format of the function related to the name. The language model receives an input of a query character string and outputs a response character string to follow the query character string. The language model is adjusted so as to output, if a pre-defined function call condition is satisfied, a response character string in a predetermined format including arguments to be passed to the function. This response acquisition unit analyzes, according to the format, the first response character string output from the language model in response to the first query character string, and extracts values therefrom.

Description

Language processing system, language processing method, and language processing program

[0001] The present invention relates to a language processing system, a language processing method, and a language processing program.

[0002] General-purpose language models such as LLMs (Large Language Models) are known. When incorporating the output of a general-purpose language model into a system, there is a desire to control the output format of the language model for appropriate analysis. Non-Patent Document 1 discloses a technique for making the output in a desired format by re-training the language model.

[0003] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, “LoRA: Low-Rank Adaptation of Large Language Models”,ICLR 2022, January 29, 2022

[0004] However, when the language model is re-trained, while the compliance with the output format increases, there is a possibility that the generalization performance may decrease. An object of the present invention is to provide a language processing system, a language processing method, and a language processing program that can obtain a response in a predetermined format without re-training the language model.

[0005] According to a first aspect of the present invention, a language processing system includes an execution unit that provides a first query string including a name of the function to a language model that receives an input of a query string and outputs a response string following the query string, and is adjusted to output a response string in a predetermined format including an argument to be passed to the function when a calling condition of a predefined function is satisfied, and instructs an answer according to a format of the function related to the name; and a response acquisition unit that analyzes the first response string output from the language model according to the query string and extracts a value.

[0006] According to the above embodiment, the language processing system can obtain a response in a predetermined format without retraining the language model.

[0007] This is a diagram showing the configuration of the language processing system according to the first embodiment. This is a diagram showing an example of the configuration of the front server according to the first embodiment. This is a flowchart showing the operation of the language processing system in the first embodiment. This is a diagram showing an example of the configuration of the front server according to the second embodiment. This is a diagram showing an example of the configuration of the front server according to the third embodiment. This is a schematic block diagram showing the configuration of a computer according to at least one embodiment.

[0008] <First Embodiment> <Configuration of the Language Processing System> The embodiments will be described in detail below with reference to the drawings. Figure 1 is a diagram showing the configuration of the language processing system according to the first embodiment. The language processing system 1 comprises a language model server 10, a front server 20, and a user terminal 30.

[0009] The language model server 10 provides an application service that accepts a query string as input and returns a response string following the query string. The language model server 10 generates the response string using a language model such as an LLM. The language model may be composed of machine learning models such as a Generative Adversarial Network (GAN) or a Transformer. Furthermore, the language model is not limited to models computed by a computer, as in the machine learning models described above, but may also include human input and output. A Large-Scale Language Model (LLM) is an example of a language model trained to receive natural language text as input and generate natural language text (strings) following the input natural language text. The query string for the LLM is also called a "query string". In this embodiment, the "language model" includes not only probabilistic models such as LLMs (language models in the narrow sense) but also application services that add pre-processing or post-processing to the input and output of probabilistic models (language models in the broad sense).

[0010] The language model server 10 according to the first embodiment has a “Function Calling” function. “Function Calling” is a mechanism for the language model server 10 to obtain information from an external service in order to generate a response string using information obtained from the external service in addition to the query string. Specifically, when the language model server 10 needs to call a function to obtain information from an external service based on the query string, it outputs a response string for calling that function. The format of the response string for calling the function is predetermined. The user of the language model server 10 (for example, the front server 20) can obtain the desired information from the external service by parsing the response string to obtain the arguments necessary to call the external service and calling the external service. In order to call the function, the format must be strictly adhered to, so the language model server 10 is adjusted so that the specified format is maintained for the response string.

[0011] The front server 20 receives a query statement from the user terminal 30 and provides an application service that generates a response to the query statement. The response may be represented as text or as an image such as a graph. The front server 20 receives a query statement from the user terminal 30, generates a query string containing the query statement, and outputs it to the language model server 10. The front server 20 retrieves the response string generated based on the query string from the language model server 10. The front server 20 parses the response string obtained from the language model server 10 and returns the response to the user terminal 30.

[0012] The user terminal 30 accesses the front server 20 and displays the interface screen for the application services provided by the front server 20. The user terminal 30 sends the query text entered by the user to the front server 20 via the interface screen. The interface screen that the user terminal 30 receives from the front server 20 after sending the query text includes the response to the query text.

[0013] In the first embodiment, the front server 20 parses the response string obtained from the language model server 10 and generates an interface screen to be output to the user terminal 30. Therefore, the response string that the front server 20 receives from the language model server 10 must be provided in a parseable format such as JSON.

[0014] In the first embodiment of the language processing system 1, the language model server 10, the front server 20, and the user terminal 30 are provided as separate devices, but the language processing system 1 according to other embodiments is not limited to this. For example, the language processing system 1 according to other embodiments may have a configuration in which the front server 20 has a language model and does not have a language model server 10, or it may have a configuration in which the front server 20 is not provided and application service software is installed on the user terminal 30.

[0015] 《Front Server 20》 The front server 20 according to the first embodiment receives an agenda item from a user and provides a discussion application that enables role-playing of a discussion on the agenda item using a language model.

[0016] The front server 20 can obtain the content of a statement on an agenda item from the language model server 10 by providing the language model server 10 with a query string that includes the agenda item and the role. A role is a command given to the language model server 10 as input data, and is setting information that represents the persona (character, personality, personality, etc.), occupation, age, background, and other attributes of the person that the language model server 10 is to roleplay. Since the language model server 10 generates output data that follows the input data, by inputting input data including the role to the language model server 10, a response string that would be spoken by the person represented by the role can be obtained.

[0017] The language model server 10 may be given knowledge of the person to be role-played by RAG (Retrieval-Augmented Generation), fine-tuning (full fine-tuning, adapter tuning), etc. When using RAG, the data that can be referenced may differ for each role. For example, when generating output data in accordance with the role of a lawyer, the generation system references legal data via RAG. On the other hand, when generating output data in accordance with the role of a meteorologist, the generation system references meteorological data via RAG. When using adapter tuning, adapter models for each role are prepared in advance, and the output data obtained by inputting input data into the adapter corresponding to the specified role is input into the LLM.

[0018] In the first embodiment, the front server 20 simulates individuals (agents) with different personalities by assigning different roles to the language model server 10. The front server 20 can generate output data representing the second agent's statement after receiving the first agent's statement by inputting output data, which includes the first agent's statement and the second agent's role, into the language model server 10. The front server 20 can simulate a discussion between multiple agents by repeating this process. In this case, the front server 20 may assign all of the multiple roles to a single language model server 10, allowing a single language model server 10 to simulate all of the multiple agents. Alternatively, the front server 20 may distribute the multiple roles among several language model servers 10, sharing the task of simulating multiple agents among multiple language model servers 10. In this way, the front server 20 simulates a discussion between multiple agents with different settings, allowing users to obtain diverse opinions from the discussion results.

[0019] Figure 2 shows an example configuration of the front server 20 according to the first embodiment. The front server 20 includes an input unit 201, a query acquisition unit 202, a role storage unit 203, an execution unit 204, a response acquisition unit 205, and an output unit 206.

[0020] The input unit 201 receives input operations from the user terminal 30. The query acquisition unit 202 prompts the user to input a topic and determines the string entered by the user via the input unit 201 as the topic. The topic is an example of a query statement that represents a question to the language model server 10.

[0021] The role storage unit 203 stores role information related to the roles to be given to the language model server 10. The role storage unit 203 stores, for example, agent names, job titles, and personas in association with each other. The agent name is a string indicating the name of the agent to be simulated by the role. The job title and persona are strings indicating the job title and persona of the agent to be simulated by the language model server 10, respectively.

[0022] The execution unit 204 inputs a query string containing the agenda and roles to the language model server 10, thereby generating text (spoken text) that simulates a discussion between agents.

[0023] Below are examples of query strings to simulate a discussion. In the following example query strings, [] indicates a variable. "#Instructions You are [Agent Name]. Your occupation is [Job Title] and your persona is [Persona]. Please continue speaking in the following conversation. #Topic [Topic] #Previous conversation [Agent A1]: [Statement 1] [Agent A2]: [Statement 2] [Agent A3]: [Statement 3] [Agent A4]: [Statement 4] ... #Answer format When answering, please return the execution format of the function send_answer.<function calling> { name: send_answer, properties: { Statement: { type: string description: Your statement on the topic.} Confidence level: { type: float description: Your degree of confidence in your statement. Expressed as a value between 0 and 1.} argument: {Statement, Confidence level}}<function calling> "

[0024] Of the query strings above, "Instruction," "Topic," and "Previous Conversation" are examples of query statements representing questions to the language model server 10. "Previous Conversation" is a history of spoken sentences generated based on other agents and the roles associated with that agent. Note that when the execution unit 204 generates the first spoken sentence, "Previous Conversation" may be blank or a sentence indicating that there is no spoken sentence. The query strings above include a definition of a function (sent_answer) and a string that specifies the execution format of the function as the format of the response string. The function definition (definition string) is written in the format (JSON) specified by the language model server 10. In the example above, the function definition is expressed in JSON format, with the function name and the format of the arguments.

[0025] The response acquisition unit 205 acquires the response string generated by the language model server 10 in response to the query string generated by the execution unit 204. Because the query string specifies the execution format of a function as "function calling," the response string is expressed in the format of the specified function. An example of a response string is shown below.<function calling> { name: send_answer, argument: {"I think that on this topic, ...", 0.86}}<function calling> "

[0026] The response acquisition unit 205 parses the acquired response string according to the format specified in the query string and obtains a value. In the above example, the response acquisition unit obtains "I think that..." as the value of the statement from the response string and obtains 0.86 as the confidence value.

[0027] The output unit 206 generates a display screen representing the status of the discussion among multiple agents based on the values ​​obtained by the response acquisition unit 205, and outputs it to the user terminal 30. For example, the output unit 206 may generate a display screen that displays the statements of each agent in a font size corresponding to the confidence level value. Since the response string format allows for the separation of the statement content and the confidence level, the output unit 206 can handle the statement content and the confidence level separately. Also, since the confidence level type (variable type) in the response string format is a number (float), the confidence level value can be used in numerical calculations.

[0028] 《Operation of the Language Processing System 1》 Figure 3 is a flowchart illustrating the operation of the language processing system 1 in the first embodiment. When a user accesses the front server 20 via the user terminal 30, the query acquisition unit 202 generates an agenda input screen for the user terminal 30 to input an agenda item and outputs it to the user terminal 30 (step S1). When the input unit 201 receives a string representing an agenda item from the user, the query acquisition unit 202 determines the string entered by the user as the agenda item (step S2).

[0029] The execution unit 204 selects one role from among the roles stored in the role storage unit 203 that has not yet been selected for creating a query string to obtain spoken text (step S3). In other words, the execution unit 204 selects one agent that has not spoken. The execution unit 204 selects a role using a method such as a round-robin method or a method using the language model server 10. The query acquisition unit 202 generates a query text using the agenda determined in step S2, the role selected in step S3, and the previously generated spoken text (step S4). The execution unit 204 sends the query text generated in step S4, a query string including a function definition including the format of the response string, and a specification of the answer format specifying the name of the defined function to the language model server 10 (step S5). The function definition and answer format according to the first embodiment are predetermined by the creator of the discussion application.

[0030] The response acquisition unit 205 receives a response string from the language model server 10 (step S6). The response acquisition unit 205 obtains the uttered text and confidence value from the received response string according to the format included in the function definition (step S7).

[0031] The output unit 206 generates a discussion progress screen and outputs it to the user terminal 30 (step S8). On the discussion progress screen, the statements made by agents and icons representing the agents who made the statements are arranged in chronological order. The statements are displayed in a font size corresponding to the level of confidence. For example, if the level of confidence is less than 0.3, the statement is displayed in a small font size (e.g., 7 points); if the level of confidence is 0.3 or more but less than 0.7, it is displayed in a normal font size (e.g., 10 points); and if the level of confidence is 0.7 or more, it is displayed in a large font size (e.g., 12 points).

[0032] The execution unit 204 determines whether there are any roles that have not yet been selected for creating a query string to obtain the spoken text (step S9). If there are any roles that have not been selected (step S9: YES), the execution unit 204 returns to step S3 and continues generating the next spoken text. In the next step S4, the query acquisition unit 202 generates a query string using the spoken text obtained from the response string in step S7. On the other hand, if there are no roles that have not been selected (step S9: NO), the execution unit 204 finishes generating the spoken text for the target group.

[0033] 《Effects》 As described above, the language processing system 1 according to the first embodiment performs the following processing. The execution unit 204 provides the language model server 10 with a first query string that includes the name of a function and instructs it to provide a response in accordance with the format of the function related to that name. The language model server 10 is a language model that accepts the input of a query string and outputs a response string that follows the query string, and is configured to output a response string in a predetermined format that includes arguments to be passed to a predefined function when the conditions for calling that function are met. The response acquisition unit 205 parses the first response string output from the language model in accordance with the specified format and extracts values.

[0034] A language model with a "function calling" function is configured to reliably generate a response string that conforms to the specified function format when a function call is deemed necessary. This is because the "function calling" function cannot be realized unless the format specification is followed. Therefore, the language processing system 1 according to the first embodiment can obtain a response string written in the desired format by specifying the execution format of a function defined according to the "function calling" specification as the response format.

[0035] Furthermore, the language processing system 1 according to the first embodiment specifies the name of the function related to the output format in the query string. This prevents the language model server 10 from selecting a format unintended by the designer of the front server 20.

[0036] <Second Embodiment> The front server 20 according to the first embodiment obtains a response string in a format specified in advance by the designer by including the format in the query string. In contrast, the front server 20 according to the second embodiment adaptively generates a format and instructs the language model server 10 to output a response string according to the generated format.

[0037] Figure 4 shows an example configuration of the front server 20 according to the second embodiment. The front server 20 according to the second embodiment includes a determination unit 207 in addition to the configuration of the first embodiment. The determination unit 207 determines a format that enables the parsing of the response string of the language model server 10 based on the query text generated by the query acquisition unit 202, and generates a definition of a function to realize that format. The determination unit 207 may, for example, use the language model server 10 to create the definition of the function.

[0038] Below is an example of a prompt for generating function definitions. "# Instructions Generate a function schema for use when sending answers to the following instruction text. Please refer to the example function schema when creating it. # Instruction Text [Query Text] # Example Function Schema { name: send_choice_question_answer, description: “A function for sending answers to multiple-choice questions”, argument: {answer}} { name: send_choice_question_answer_with_confidence, description: “A function for sending answers including confidence levels to multiple-choice questions”, argument: {answer, confidence}} { name: send_debate_answer, description: “A function for sending debate results in debates”, argument: {conclusion, reason}}"

[0039] As a result, the language processing system 1 according to the second embodiment can automatically generate a suitable format for parsing response strings for any query text. This allows designers to design application services without having to consider detailed formats tailored to the content of the questions in advance.

[0040] <Third Embodiment> The language processing system 1 according to the third embodiment specifies the format of the response string using a set of functions prepared in advance for "function calling". Figure 5 is a diagram showing an example of the configuration of the front server 20 according to the third embodiment. In addition to the configuration of the second embodiment, the front server 20 according to the third embodiment further includes a function database 208. The function database 208 is a database in which a set of functions prepared in advance for "function calling" is recorded. The functions included in the set of functions may include processing that accesses an external service in accordance with the API (Application Programming Interface) of the external service.

[0041] In the third embodiment, the determination unit 207 searches for a suitable function from the function database 208's group of functions based on the query statement, to represent the format for obtaining a response string for the query statement. The determination unit 207 may, for example, project the query statement and the descriptions of each function stored in the function database 208 onto a vector space to convert them into text vectors, and then calculate the similarity between the text vector related to the query statement (query vector) and the text vectors related to the descriptions of each function (condition vectors) to determine a suitable function. The function descriptions are examples of the conditions for calling the function. In other words, the determination unit 207 can obtain the function associated with the description that has the highest similarity to the text vector related to the query statement. The similarity in this case corresponds to the degree of suitability (appropriateness) of the function to be used as the answer to the instruction text.

[0042] Furthermore, for example, the decision unit 207 according to the third embodiment may search for an appropriate function using, for example, the language model server 10. Below is an example of a prompt for searching for a function. "#Instruction From the list of functions below, please select the appropriate function to return an answer to the instruction text given next. #Instruction text [Query statement] #List of functions [Functions stored in the function database 208] #Answer format When answering, please return the executable form of the function function_retrieval.<function calling> { name: function_retrieval, properties: { function_name: { type: string description: The name of the function included in the list of functions.} aptitude: { type: float description: The degree to which the function is suitable to be used as the answer to the instruction text. It is expressed as a value between 0 and 1.} argument: { function_name, aptitude}}<function calling> "

[0043] The determination unit 207 determines whether the suitability of the retrieved function exceeds a predetermined threshold. When the suitability of the function exceeds the threshold, the execution unit 204 generates a query string that includes the name of the retrieved function and instructs an answer according to the format of the function related to the name. This query string does not necessarily include the definition of the function. When the suitability of the function does not exceed the threshold, the execution unit 204 generates a format that can be analyzed for the response string in the same manner as in the second embodiment. In this case, the execution unit 204 includes the definition of the function and generates a query string that instructs an answer according to the format of the function.

[0044] <Other Embodiments> Although one embodiment has been described in detail above with reference to the drawings, the specific configuration is not limited to the above, and various design changes and the like are possible. That is, in other embodiments, the order of the above processes may be appropriately changed. Also, some processes may be executed in parallel. The language model server 10 and the front server 20 according to the above-described embodiment may each be configured by a single computer, or may be divided and arranged on a plurality of computers by cloud computing or the like, and the plurality of computers may cooperate with each other to function as the language model server 10 or the front server 20.

[0045] The front server 20 according to the above-described embodiment provides a discussion application, but is not limited to this in other embodiments, and may provide any application service using the language model server 10. Therefore, the processing method of each value obtained from the response string is not limited to the above-described method. For example, the front server 20 according to another embodiment may draw a figure based on the value obtained from the response string. Also, for example, the front server 20 according to another embodiment may calculate a control amount for controlling a machine actuator based on the value obtained from the response string and control the machine according to the control amount.

[0046] <Computer Configuration> Figure 6 is a schematic block diagram showing the configuration of a computer according to at least one embodiment. The computer 90 includes a processor 91, main memory 92, storage 93, and interface 94. The language model server 10, front server 20, and user terminal 30 described above are each implemented in the computer 90. The operation of each of the above-described processing units is stored in the storage 93 in the form of a program. The processor 91 reads the program from the storage 93, expands it into the main memory 92, and executes the above-described processing according to the program. The processor 91 also allocates storage areas in the main memory 92 corresponding to each of the above-described storage units according to the program. Examples of the processor 91 include a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), and a microprocessor.

[0047] The program may be for implementing a part of the functions to be performed by the computer 90. For example, the program may perform functions in combination with other programs already stored in storage, or in combination with other programs implemented on other devices. In other embodiments, the computer 90 may include a custom LSI (Large Scale Integrated Circuit) such as a PLD (Programmable Logic Device) in addition to, or instead of, the above configuration. Examples of PLDs include PAL (Programmable Array Logic), GAL (Generic Array Logic), CPLD (Complex Programmable Logic Device), and FPGA (Field Programmable Gate Array). In this case, some or all of the functions implemented by the processor 91 may be implemented by the integrated circuit. Such an integrated circuit is also included as an example of a processor. In other embodiments, the computer 90 may be virtualized on one or more computers.

[0048] Examples of the storage 93 include a magnetic disk, a magneto-optical disk, an optical disk, a semiconductor memory, and the like. The storage 93 may be an internal medium directly connected to the bus of the computer 90, or an external medium connected to the computer 90 via the interface 94 or a communication line. Also, when this program is distributed to the computer 90 via a communication line, the computer 90 that has received the distribution may expand the program in the main memory 92 and execute the above processing. In at least one embodiment, the storage 93 is a non-transitory tangible storage medium.

[0049] Also, the program may be for realizing a part of the functions described above. Further, the program may be a so-called difference file (difference program) that realizes the functions described above in combination with other programs already stored in the storage 93.

[0050] Regarding the above embodiments, the following additional notes are further disclosed.

[0051] (Additional Note 1) A language processing system including one or more processors, wherein the one or more processors receive an input of a query string and output a response string following the query string, and are adjusted to output a response string in a predetermined format including an argument to be passed to a function when a calling condition of a predefined function is satisfied, and give a first query string including the name of the function and instructing an answer according to the format of the function related to the name to a language model; and automatically analyze the first response string output from the language model according to the query string and extract a value.

[0052] (Additional Note 2) The language processing system according to Additional Note 1, wherein the one or more processors execute arithmetic processing based on the variable type specified in the format using the value extracted from the first response string.

[0053] (Note 3) The language processing system according to Note 2, wherein the one or more processors calculate a control amount for the machine using the value extracted from the first response string and execute control of the machine.

[0054] (Note 4) A non-temporary storage medium storing a language processing program executable by a computer to execute a language processing method, wherein the language processing method includes: a language model that accepts a query string input and outputs a response string following the query string, and is adjusted to output a response string in a predetermined format that includes arguments to be passed to a predefined function when the conditions for calling the function are met, and provides the language model with a first query string that includes the name of the function and instructs the response to follow the format of the function related to that name; and automatically parses the first response string output from the language model in response to the first query string according to the format and extracts a value.

[0055] 1...Language processing system 10...Language model server 20...Front server 201...Input unit 202...Query acquisition unit 203...Role storage unit 204...Execution unit 205...Response acquisition unit 206...Output unit 207...Decision unit 208...Function database 30...User terminal 90...Computer 91...Processor 92...Main memory 93...Storage 94...Interface

Claims

1. A language processing system comprising: an execution unit that provides a language model that accepts a query string input and outputs a response string following the query string, and is adjusted to output a response string in a predetermined format that includes arguments to be passed to a predefined function when the conditions for calling the function are met, with a first query string that includes the name of the function and instructs the response to follow the format of the function related to that name; and a response acquisition unit that parses the first response string output from the language model in response to the first query string according to the format and extracts values.

2. The language processing system according to claim 1, wherein the execution unit provides the language model with the first query string, which includes a string representing the format of the function.

3. A language processing system according to claim 2, comprising: a query acquisition unit that acquires a query statement representing a question to the language model; and a determination unit that determines the format of a response string based on the query statement, wherein the execution unit provides the language model with the first query string, which includes the query statement and a function definition string that includes the determined format.

4. The language processing system according to claim 3, wherein the determination unit provides the language model with a second query string that includes an instruction to determine a format suitable as a string to respond to the query sentence, and determines the format based on the second response string for the second query string.

5. The language processing system according to claim 3, wherein the determination unit determines from among a plurality of formats a format suitable as a string to respond to the query sentence.

6. A language processing system according to claim 5, wherein each of the plurality of formats is associated with a function call condition related to that format, and the determination unit determines the format based on the similarity between a query vector obtained by projecting the query text onto a vector space and a condition vector obtained by projecting the call conditions related to the plurality of formats onto a vector space.

7. A language processing method comprising: a step of providing a first query string to a language model that accepts a query string input and outputs a response string following the query string, the language model being adjusted to output a response string in a predetermined format that includes arguments to be passed to a predefined function when the conditions for calling the function are met, the first query string including the name of the function and instructing a response in accordance with the format of the function related to that name; and a step of parsing the first response string output from the language model in response to the first query string according to the format and extracting values.

8. A language processing program that causes a computer to perform the following steps: providing a first query string to a language model that accepts a query string input and outputs a response string following the query string, and is configured to output a response string in a predetermined format that includes arguments to be passed to a predefined function when the conditions for calling the function are met, the first query string that includes the name of the function and instructs the computer to provide a response in accordance with the format of the function related to that name; and parsing the first response string output from the language model in response to the first query string according to the format and extracting values.

Images