Method for business control based on large language model

Abstract

A method for business control based on a large language model, which performs user intent recognition on input text through a large language model, outputs a classification instruction according to the user intent recognition result, and then outputs a corresponding control instruction to a business execution module according to the classification instruction. The business execution module executes corresponding actions according to the received control instruction. The user intent recognition result outputs a classification instruction. First, the user intent is classified to determine whether the user intent belongs to a fixed intent or a variable intent. When the user intent belongs to a fixed intent, a fixed instruction is output. When the user intent belongs to a variable intent, a variable instruction is output. The present application uses a large model to recognize intent and further outputs system control instruction parameters through intent, thereby achieving the purpose of controlling the system with natural language. The effect of not being rigid and being controllable by the system is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligence, in particular to a method for business control based on a large language model. BACKGROUND

[0002] Large language models (LLM for short) are a type of artificial intelligence model designed to understand and generate human language. They are trained on large amounts of text data and can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and more. LLMs are characterized by their size, which helps them learn complex patterns in language data. These models are typically based on deep learning architectures and have achieved good performance on various NLP (Natural Language Processing) tasks.

[0003] The current market applications of large models are mainly in the form of question and answer, and most of the systems that can be controlled by natural language can only be controlled by fixed sentences or fixed sentence structures. The former makes the application scenario of large models too single, and the latter makes users need to learn and remember various instructions, thereby increasing the difficulty of use for users.

[0004] Therefore, in view of the deficiencies of the prior art, it is necessary to provide a method for business control based on a large language model to overcome the deficiencies of the prior art. SUMMARY

[0005] The purpose of the present application is to provide a method for business control based on a large language model to overcome the deficiencies of the prior art. The method uses a large model to identify the user's intent and further outputs system control instruction parameters according to the intent, thereby achieving natural language control of the system and achieving the effect of non-rigid instructions and system controllability.

[0006] The purpose of the present application is achieved through the following technical measures.

[0007] The present application provides a method for business control based on a large language model. The method uses a large language model to identify the user's intent based on the input text, outputs a classification instruction according to the user's intent recognition result, and then outputs the corresponding control instruction to the business execution module according to the classification instruction. The business execution module executes the corresponding action according to the received control instruction.

[0008] Preferably, in the method for business control based on a large language model, a plurality of intents are pre-set in the prompt words input to the large model.

[0009] Preferably, in the method for business control based on a large language model, the classification instruction is output according to the user's intent recognition result, and specifically includes:

[0010] First, the user intention is classified to determine whether the user intention belongs to a fixed intention or a variable intention. When the user intention belongs to a fixed intention, a fixed instruction is outputted;

[0011] When the user intention belongs to a variable intention, a variable instruction is outputted;

[0012] The fixed intention refers to an intention that can directly determine the instruction according to the intention,

[0013] The variable intention refers to an intention that needs other changes to determine the instruction after confirming the user intention;

[0014] The fixed instruction refers to a determined instruction, and the variable instruction refers to an instruction that needs to be determined according to other factors.

[0015] Preferably, the variable instruction is obtained in the following manner in the method for business control based on the large language model:

[0016] The large language model extracts specific content of the user input;

[0017] Then, the variable intention is determined according to the specific content of the user input extracted by the large language model;

[0018] And the variable instruction is outputted according to the specific content of the user input extracted by the large language model and the recognized variable intention.

[0019] Preferably, the variable instruction is obtained in the following manner in the method for business control based on the large language model:

[0020] The large language model extracts specific content of the user input;

[0021] The variable intention is determined according to the specific content of the user input extracted by the large language model;

[0022] The parameters of the business system module are obtained through other interfaces;

[0023] The variable instruction is outputted according to the specific content of the user input extracted by the large language model and the recognized variable intention.

[0024] Preferably, the method for business control based on the large language model further comprises a voice processing step. The voice input of the user is converted into text content by a voice-to-text module, and the converted text content is inputted into the large language model as text content.

[0025] The application is based on a method for business control by a large language model, which identifies user intent from input text by a large language model, outputs a classification instruction according to the user intent recognition result, and then outputs a corresponding control instruction to a business execution module according to the classification instruction. The business execution module executes corresponding actions according to the received control instruction. The application achieves the purpose of natural language control of the system by identifying intent by a large model and further outputting system control instruction parameters according to the intent, thereby achieving the effect of non-rigid instruction and system controllability. BRIEF DESCRIPTION OF DRAWINGS

[0026] The application will be further described with reference to the accompanying drawings, but the content of the drawings does not constitute any limitation on the application.

[0027] Figure 1 is a flowchart of a method for business control based on a large language model. DETAILED DESCRIPTION

[0028] The application will be further described with reference to the following examples.

[0029] Example 1.

[0030] A method for business control based on a large language model, as shown in Figure 1 , identifies user intent from input text by a large language model, outputs a classification instruction according to the user intent recognition result, and then outputs a corresponding control instruction to a business execution module according to the classification instruction. The business execution module executes corresponding actions according to the received control instruction.

[0031] The use scenario of this scheme is to process user voice output to control the business system to execute related business actions. As shown in Figure 1 , the voice processing module receives the user's voice input, and the voice-to-text module converts the user's voice input content into text content, which is input to the large language model as text content.

[0032] The large model processing module uses a large voice model to obtain user intent based on the excellent understanding ability of natural language, and outputs corresponding instructions according to user intent classification.

[0033] The large model processing module is the core module of the present embodiment, which performs the following operations by receiving input text:

[0034] (1) Intent recognition of input text. The specified multiple intents are pre-set in the prompt words of the large model, so that the large model matches the intent according to the user's input and outputs the intent, thereby achieving the effect of intent recognition.

[0035] The intent input to the large language model can be flexibly set according to actual needs, such as: when it is desired to identify a province, the intent can be "Guangdong Province", "Hunan Province", etc.; when it is desired to identify an action, the intent content can be "turn on the light", "increase the volume", etc.; when it is desired to query specified content, the intent can be "rainfall", "water flow", etc.

[0036] (2) Classify the intent. Because different intents correspond to different operations, a branch structure is needed to determine the subsequent process direction according to the intent output by the large model.

[0037] Classify the user intent to determine whether the user intent is a fixed intent or a variable intent, and output a fixed instruction when the user intent is a fixed intent, and output a variable instruction when the user intent is a variable intent.

[0038] Fixed intent refers to an intent that can directly determine the instruction according to the intent, and variable intent refers to an intent that requires other changes to determine the instruction after confirming the user intent. Fixed instruction refers to an instruction that is determined, and variable instruction refers to an instruction that needs to be determined according to other factors.

[0039] Figure 1 In the above, classification 1 is a path representing a fixed instruction, i.e., the instruction is directly determined according to the intent. The instruction here is fixed, such as "mute" and "unmute", and the business system module only needs to perform the "mute" and "unmute" operations to complete the task. Fixed intent and fixed instruction do not contain changes, so no further operation is needed.

[0040] Figure 1 In the above, classification 2 is another path representing a variable instruction, i.e., after confirming the intent, the specific content of the user input needs to be extracted, or even other interfaces need to be called to obtain parameters. For example, "sound to 20" (the specific content of the user input needs to be extracted) and "a little louder" (in addition to extracting the specific content of the user input, other interfaces need to be called to obtain the current volume parameter).

[0041] The method for controlling business based on a large language model, the acquisition of variable instructions needs to be determined in combination with specific circumstances, and one specific acquisition method is listed below:

[0042] The large language model extracts the specific content of the user input;

[0043] Then, according to the specific content of the user input extracted by the large language model, it is determined that it is a variable intent;

[0044] According to the specific content of the user input extracted by the large language model and the identified variable intent, a variable instruction is output.

[0045] For example, the user inputs the text "What is the average rainfall in Guangzhou this month?" and the large language model determines that it is a variable intent based on the specific content of the user input.

[0046] Large model parameter extraction: Extract the parameter content based on the user input text and the parameters that the developer has pre-set to extract.

[0047] For example, the user inputs the text "What is the average rainfall in Guangzhou this month?" and the developer has pre-set the large model to extract "startDate", "endDate", and "area". Then the large model extracts the parameters

[0048]

[0049] According to "startDate": "2024-11-01", "endDate": "2024-11-30", and "area": "guangzhou", output the variable instruction.

[0050] The following lists another way to obtain a variable instruction. In this case, the variable instruction is obtained as follows:

[0051] The large language model extracts the specific content of the user input;

[0052] According to the specific content of the user input extracted by the large language model, determine the variable intent;

[0053] Call other interfaces to obtain the parameters of the business system module;

[0054] According to the specific content of the user input extracted by the large language model and the identified variable intent, output the variable instruction.

[0055] For example, the user inputs the text "Increase the volume by 5".

[0056] The large model analyzes the text "Increase the volume by 5" and matches the intent to "Increase the volume".

[0057] This is a variable intent, and the developer has pre-set the large model to extract the parameter { "amount ": 5} based on the user input text.

[0058] However, the business system module does not have an interface to increase the volume, but only has an interface to set the specific volume. Therefore, when integrating the variable instruction, it is necessary to call the current volume interface to obtain the data supplement. Add the current volume value obtained by calling the current volume interface to the parameter "amount": 5 to get the specific volume that needs to be set.

[0059] Encapsulate and return the specific volume and system parameters such as identity verification according to the instruction unified structure.

[0060] Figure 1 The right part of the business system module is to receive the intended instruction, such as json input parameter, mechanical code, etc. to execute the operation of the business system.

[0061] The method for controlling business based on a large language model, the user intent is recognized by the large language model to the input text, the classification instruction is output according to the user intent recognition result, and the corresponding control instruction is output to the business execution module according to the classification instruction, and the business execution module executes the corresponding action according to the received control instruction. The present application realizes the purpose of controlling the system by natural language by identifying the intention of the large model and further outputting the system control instruction parameter, so that the system can execute the operation based on the intention recognition, such as "it is dark" corresponding to the operation "turn on the light". The system can be controlled by voice, text or other input forms. The change and adjustment of the function can be quickly modified by adjusting the prompt word of the large model. The effect of controlling the system can be achieved by non-fixed format natural language, and the effect of not rigid instruction and system controllable is achieved.

[0062] Embodiment 2.

[0063] A method for controlling business based on a large language model, other features are the same as embodiment 1, the difference is that in this embodiment, the user's voice input is not obtained by the voice processing module and converted into text content, but the user directly inputs the text content to the large model processing module for intent recognition and outputs the control instruction.

[0064] The user can directly input the text content through the keyboard, input the text content in the form of inputting text or transmitting pictures through the interactive interface, or input the corresponding text content in the form of operating the keys through other control interfaces.

[0065] Embodiment 3.

[0066] The effect of the method for controlling business based on a large language model is further described in combination with specific examples.

[0067] Pre-set scene:

[0068] The set intention scene: turn on the light, turn off the light, mute, unmute, adjust the volume, and broadcast the time.

[0069] The current volume is set to 20.

[0070] In this embodiment, the volume is fixed at 20, and the output instruction is replaced by natural language, which is convenient to see the effect.

[0071] The user inputs a series of voice, and the user intent recognized by the method of the present application and the related instruction formed are shown in Table 1.

[0072] Table I

[0073] Sequence User's voice input Recognized user intent Output instruction 1 Turn on the light Turn on the light Turn on the light 2 It's too dark Turn on the light Turn on the light 3 It's too noisy Mute Mute 4 Why is there no sound Unmute Unmute

[0074] It can be seen that the method of the present application can avoid fixed input to obtain return instructions, and the user can interact with the system as ordinary communication.

[0075] For the user's input of colloquial speech, the method of the present application can also recognize dynamic instructions, and the results are shown in Table II.

[0076] Table II

[0077] Sequence User's voice input Recognized user intent Output instruction 1 The sound is too small Adjust the volume Adjust the volume to 25 2 Make the sound a little louder Adjust the volume Adjust the volume to 25 3 The sound is too small Adjust the volume Adjust the volume to 30 4 What day is it today Announce the time It is now October 29, 2024 5 What time is it Announce the time It is now Tuesday, October 29, 2024 17:30

[0078] From Table II, it can be seen that for completely colloquial input, the large language model can also perform intent recognition according to the preset prompt word and infer information according to the user's input description, and then output dynamic instructions.

[0079] The method of the present application can realize system control through simple configuration setting of intent list and description for non-fixed sentence, colloquial natural language input.

[0080] Embodiment 4.

[0081] The method of the present application is further described in combination with specific examples.

[0082] Pre-set scene:

[0083] Set scene intent: weather query, temperature query, and outing suggestion.

[0084] Set the current area to be Guangzhou.

[0085] Set the current temperature to be 9℃ to 21℃.

[0086] Set the current time to be November 29, 2024, 10:00:00.

[0087] Set the current weather to be sunny to cloudy.

[0088] The user inputs a series of texts, and the user's intent recognized by the method of the present application and the related instructions formed are shown in Table III.

[0089] Table III

[0090]

[0091]

[0092] From Table Three, it can be seen that with completely oral input, the large model can also identify the intent according to the preset prompt word and infer information according to the user's input description, and can call other business interfaces to obtain supplementary information according to specific needs when integrating instructions, and then output dynamic instructions.

[0093] Table Three Sequence No. 1 Implementation Steps:

[0094] The large language model identifies the intent of the text "Will it rain today?" as "weather query".

[0095] The intent "weather query" is a variable intent that needs to be assembled into a dynamic instruction. The large model extracts the parameters date and location according to the developer's preset, with the default date being the current day and the default location being Guangzhou, so the extracted parameters are

[0096]

[0097] The date and location contain the requirements for assembling the weather query interface, and are combined with system parameters such as identity verification to encapsulate and return the instruction interface.

[0098] Table Three Sequence No. 3 Implementation Steps:

[0099] The large language model identifies the intent of the text "Should I take a coat?" as "outgoing suggestion".

[0100] The intent "outgoing suggestion" is a variable intent that needs to be assembled into a dynamic instruction. The large model extracts the parameters date, time, and location according to the developer's preset, with the default date being the current day, the default time being the current time, and the default location being Guangzhou, so the extracted parameters are

[0101]

[0102] The outgoing suggestion interface requires weather and temperature, so it needs to call the weather query interface and the temperature query interface to obtain. The parameters extracted by the large model can satisfy the instructions for assembling the weather query interface and the temperature query interface respectively, so at this time the instructions for assembling the weather query interface and the temperature query interface are assembled respectively and the above two interfaces are called to obtain weather and temperature data.

[0103] According to the weather and temperature data returned by the call to the business interface, the outgoing suggestion interface is encapsulated and returned according to the instruction.

[0104] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit the protection scope of the present application, and although the present application has been described in detail with reference to the preferred embodiments, it should be appreciated by those skilled in the art that the technical solutions of the present application can be modified or equivalently replaced without departing from the essence and scope of the technical solutions of the present application.

Claims

1. A method for business control based on a large language model, characterized in that, The large language model is used to identify user intent from the input text. Based on the user intent identification results, classification instructions are output. Then, based on the classification instructions, corresponding control instructions are output to the business execution module. The business execution module executes the corresponding actions according to the received control instructions. Variable instructions are obtained in the following way: Large language models extract specific content from user input; Then, variable intents are determined based on the specific content of user input extracted from the large language model. Then, based on the specific content of the user input extracted from the large language model and the identified variable intent, variable instructions are output; The prompts input to the large model pre-include multiple intentions; Based on the user intent recognition results, classification instructions are output, specifically including: First, classify the user intent to determine whether it is a fixed intent or a variable intent. When the user intent is a fixed intent, output a fixed instruction. When the user's intent is a variable intent, a variable instruction is output; A fixed intent refers to an intent that can directly determine instructions based on that intent. Variable intent refers to instructions whose intent requires further changes after the user's intent has been confirmed. Fixed instructions are instructions that are fixed, while variable instructions are instructions that need to be determined based on other factors. Variable instructions are obtained in the following way: Large language models extract specific content from user input; Determine variable intents based on the specific content of user input extracted from the large language model; Call other interfaces to obtain parameters from business system modules; Variable instructions are output based on the specific content of user input extracted from the large language model and the identified variable intent.

2. The method for business control based on a large language model according to claim 1, characterized in that: It also includes a voice processing step. The voice-to-text module converts the user's voice input into text content, which is then input into the large language model as text content.

3. The method for business control based on a large language model according to claim 1 or 2, characterized in that: Input text content into the large language model using text input or image input.

Images