Control instruction generation method, system and equipment based on language model

Abstract

The invention relates to a control instruction generation method, system and equipment based on a language model, and the method comprises the following steps: receiving a control language input by a user, and carrying out the matching in a control instruction rule base according to the control language, and obtaining a matching result; the matching result comprises one or more of an operation intention vector, an initial instruction template vector and an initial association instruction vector; identifying an operation intention of the user according to a matching result, identifying a control parameter, and determining a closest instruction template and a related associated instruction according to the operation intention; and filling the control parameters into the closest instruction template, and splicing the control parameters with the related associated instructions to obtain a complete control instruction. The method can improve the reliability of instruction execution, and is applied to the technical field of equipment control.

Description

Technical Field

[0001] The present invention relates to the field of device control technology, and in particular to a method, system and device for generating control instructions based on a language model. Background Art

[0002] The information technology (IT) equipment monitoring and control systems widely used by the industry are mostly developed based on preset rules or fixed instruction sets. Users are typically required to enter standard instructions in a strict format through a command line interface or graphical interface. The system then maps these standard instructions into low-level operational instructions to obtain or control the device status. However, existing large-scale model-based instruction generation is mainly based on natural language descriptions. The generated instructions often fail to meet the precise parameters or logical requirements required for device execution. This is especially true in scenarios requiring strict technical specifications. This significantly limits practicality and reduces the reliability of instruction execution. Summary of the Invention

[0003] In view of this, an object of the present invention is to provide a method, system and device for generating control instructions based on a language model, which can improve the reliability of instruction execution.

[0004] In one aspect, the present invention provides a method for generating control instructions based on a language model, comprising the following steps:

[0005] receiving a control language input by a user, and matching the control language in a control instruction rule library to obtain a matching result; the matching result includes one or more of an operation intention vector, an initial instruction template vector, and an initial associated instruction vector;

[0006] Identifying the user's operation intention and control parameters based on the matching results, and determining the closest instruction template and related associated instructions based on the operation intention;

[0007] The control parameters are filled into the closest instruction template and spliced ​​with the related associated instructions to obtain a complete control instruction.

[0008] Optionally, the method further includes: determining whether the control parameters are complete; if the control parameters are incomplete, supplementing the missing control parameters by preset values ​​and / or re-identifying the missing control parameters in the operation intention and supplementing the missing control parameters.

[0009] Optionally, the method further comprises: verifying the parameter range, type or control logic of the complete control instruction according to a preset rule, and obtaining the complete control instruction if the complete control instruction passes the verification;

[0010] If the complete control instruction fails to be verified, the reason for the failure is inquired and the complete control instruction is corrected according to the reason for the failure to obtain a corrected complete control instruction.

[0011] Optionally, identifying the user's operation intention according to the matching result specifically includes:

[0012] Extracting intent information based on the matching result, and identifying a plurality of sub-intents and association relationships between the sub-intents based on the intent information;

[0013] An intention logic diagram is established based on the sub-intentions and the association relationship to obtain the user's operation intention; wherein, the nodes of the intention logic diagram represent sub-intentions, and the edges between the nodes represent the association relationship between two sub-intentions.

[0014] Optionally, determining the closest instruction template and related associated instructions according to the operation intention specifically includes:

[0015] Traverse the intention logic graph and determine the closest instruction template and related associated instructions for each sub-intention.

[0016] Optionally, the method further includes: serializing the complete control instruction into a preset instruction format, and sending the serialized complete control instruction to the business system so that the business system executes the serialized complete control instruction; the preset instruction format is an instruction format executable by the business system.

[0017] Optionally, the method further includes: conducting at least two rounds of dialogue with the user, determining the operation intention of the current round of dialogue based on the operation intention and control parameters of the related rounds of dialogue, and generating complete control instructions for the current round of dialogue.

[0018] Optionally, confirm whether the operation intention is a risk intention. If the operation intention is a risk intention, generate a risk prompt and return the risk prompt to the user.

[0019] On the other hand, the present invention provides a control instruction generation system based on a language model, the system comprising an embedding model, a language model and a business system, wherein:

[0020] The embedding model is used to receive a control language input by a user, and perform matching in a control instruction rule library according to the control language to obtain a matching result; the matching result includes one or more of an operation intention vector, an initial instruction template vector, and an initial associated instruction vector;

[0021] The language model is used to identify the user's operation intention and control parameters based on the matching results, and determine the closest instruction template and related associated instructions based on the operation intention; fill the control parameters into the closest instruction template and splice it with the related associated instructions to obtain a complete control instruction;

[0022] The business system is used to process and / or execute the complete control instruction.

[0023] On the other hand, the present invention provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the aforementioned method when executing the computer program.

[0024] The implementation of the present invention includes the following beneficial effects: the present invention first matches the operation intention vector, the initial instruction template vector and the initial associated instruction vector through the language input by the user, identifies the user's operation intention and control parameters according to the operation intention vector, and then determines the instruction template and associated instructions related to the intention. By filling the control parameters into the instruction template instead of directly identifying the control instructions in the control language, the instruction template has a clear structured format and direct executableness. Even if the user input language is relatively vague, it can meet the precise parameters or logic requirements required for device execution, thereby improving the reliability of the control instructions. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of the steps of a method for generating control instructions based on a language model provided by the present invention;

[0026] Figure 2 This is a system structure diagram of a control instruction generation method based on a language model provided by the present invention;

[0027] Figure 3 This is a flow chart for constructing a control instruction rule library provided by the present invention;

[0028] Figure 4 This is a flow chart of a second form of multi-round dialogue mechanism provided by the present invention;

[0029] Figure 5 This is a structural diagram of a control instruction generation system based on a language model provided by the present invention;

[0030] Figure 6 It is a structural schematic diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are provided for ease of description only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted based on the understanding of those skilled in the art.

[0032] The terms used in the embodiments of the present application are for the purpose of describing specific embodiments only and are not intended to limit the embodiments of the present application. The singular forms "a," "the," and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0033] When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Instead, they are merely examples of devices and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0034] In addition, in this application, unless otherwise specified, "plurality" refers to two or more. "And / or" describes the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. The character " / " generally indicates that the associated objects are in an "or" relationship.

[0035] In some embodiments, as Figure 1 and Figure 2 As shown, Figure 1 It is a flow chart of the steps of the control instruction generation method based on the language model. Figure 2 The present invention provides a method for generating control instructions based on a language model, which includes the following steps:

[0036] S100: Receive a control language input by a user, perform a match in a control instruction rule library according to the control language, and obtain a matching result.

[0037] The matching results include one or more of one or more operation intention vectors, one or more initial instruction template vectors, and one or more initial associated instruction vectors. The control language can be a spoken natural language. The control instruction rule library includes, but is not limited to, an operation intention vector library, an instruction template vector library, and an associated instruction vector library. Each vector library stores high-dimensional vector data.

[0038] like Figure 3 As shown, Figure 3 It is a flow chart for constructing a control instruction rule library. The present invention inputs the device control type template into the embedded model to obtain an operation intention vector library, wherein the device control type template includes but is not limited to the operation type and instruction description of the control instruction. The device control instruction template is input into the embedded model to obtain an instruction template vector library, wherein the device control instruction template includes but is not limited to specific control instructions, parameters in the control instructions, parameter rules (such as required, type, format), and natural language descriptions corresponding to the parameters. The instruction association relationship template is input into the embedded model to obtain an associated instruction vector library, wherein the instruction association relationship template includes but is not limited to specific instructions and instruction association relationships (for example, to restart a device in a certain alarm, it is necessary to first associate the instruction query device ID of the alarm).

[0039] Specifically, the control language input by the user is embedded into the model to obtain a vector of the control language. The similarity of the control language vector with vectors in the operation intention vector library, the instruction template vector library, and the associated instruction vector library is calculated, such as cosine similarity, and the operation intention vector, initial instruction template vector, and initial associated instruction vector whose similarity exceeds a similarity threshold are matched. The number of the operation intention vector, initial instruction template vector, and initial associated instruction vector is at least one.

[0040] It is worth noting that in addition to calculating cosine similarity, regular expressions can also be used to match the sentence structure of the control language to identify operation intention vectors, initial instruction template vectors, and initial associated instruction vectors with similar sentence structures, which can improve matching accuracy.

[0041] In addition, if the matched similarity is less than or equal to the similarity threshold, the first form of multi-round dialogue mechanism is triggered. This form of multi-round dialogue mechanism does not involve the recognition of operation intentions, but guides the user to re-enter more precise control language through a finite state machine.

[0042] S200: Identify the user's operation intention and control parameters according to the matching results, and determine the closest instruction template and related associated instructions according to the operation intention.

[0043] The language model may be, but is not limited to, a large language model.

[0044] Specifically, the operation intention vector, initial instruction template vector, and initial associated instruction vector are concatenated and fed into a language model. The language model identifies the user's operation intention using methods such as keyword matching. It then uses rule-based named entity recognition to identify entities such as device names, operation types, and parameters within the operation intention, and matches control parameters using regular expressions. Furthermore, control parameters can be verified to ensure their rationality, for example by checking their range.

[0045] Key-value matching can be used to match the command template closest to the intended operation from among several initial command templates, as well as the associated commands most closely related to the intended operation or the most relevant command template. Associated commands can be other control commands that are required to execute the intended operation, such as requiring a power-on command before executing a query. Command templates are standardized command formats that meet different intended operation requirements.

[0046] For identifying control parameters, it can not only be identified through the operation intention vector of a certain time, but it can also perform semantic understanding based on the operation intention vectors generated by multiple rounds of dialogues, automatically complete the control parameters missing in a certain round of dialogue, and finally generate structured and executable control instructions.

[0047] It is worth noting that after identifying the operation intention, this step can also confirm whether the operation intention is a risk intention. If the operation intention is a risk intention, a risk warning is generated and returned to the user. If the operation intention is not a risk intention, the closest instruction template and related associated instructions are determined.

[0048] In addition, the operation intention may include multiple sub-intentions, each of which represents a subtask. The sub-intentions may include, but are not limited to, querying the device, changing the status, and restarting the device. In this case, step S300 may include the following steps:

[0049] Extract the intent information from the matching results based on key-value matching, etc. The intent information includes but is not limited to information such as task objectives, operation objects, constraints, time expressions and control logic. Identify the several sub-intentions included in the intent information based on semantics and use keyword segmentation algorithms to identify the association between the sub-intentions. Among them, the association relationship may include but is not limited to sequential execution, parallel triggering, conditional jumps and rollback processing. For intents A and B, sequential execution means that A must be executed before B; parallel triggering means that A and B can be executed at the same time; conditional jump means jumping to B based on the execution result of A; rollback processing means returning to B for rollback processing when the execution of A fails or is triggered abnormally.

[0050] According to the association relationship, several sub-intentions are associated to establish an intention logic diagram to obtain the user's operation intention. Among them, the nodes of the intention logic diagram represent sub-intentions, and the edges between the nodes represent the association relationship between two sub-intentions. In this case, the nodes in the intention logic diagram are traversed, and for each sub-intention (node), the control parameters of each sub-intention are identified respectively, and the closest instruction template and related associated instructions of each sub-intention are determined respectively. Step S300 is executed for each instruction, and the control instructions of each sub-intention are spliced ​​to obtain a complete control instruction.

[0051] Let's take an example to illustrate the situation where the above operation intention can include multiple sub-intentions. The user inputs the control language: Restart the device whose CPU usage exceeds 90% in the system. This control language includes two sub-intentions. Figure 1 To generate a query command, Figure 2 To restart, the relationship between the two sub-intentions is used as a conditional jump, with the sub-intention Figure 1 The execution result is Figure 2 The execution conditions of Figure 1 , combined with the business system querying the current device status, there are two scenarios. Scenario 1: CPU usage is 90%. The device ID is returned. As context, the device information is used as control language input to supplement the large model and generate a restart instruction. Scenario 2: There are no devices in the current system with a CPU usage exceeding 90%. No precise instructions are generated, and the message "There are no devices in the current system with CPU usage exceeding 90%" is returned.

[0052] By utilizing an intent logic graph to represent the dependencies and concurrency between user sub-intents, and then using a task fusion mechanism to combine control instructions from multiple sub-intents into composite instructions with control logic, this system supports process control semantics such as sequential execution, parallel triggering, conditional jumps, and rollback processing. This mechanism reduces the user's cognitive burden of specialized terminology and system structure, enabling non-technical users to express and execute complex instructions through fuzzy natural language, thereby improving the system's universality, interaction efficiency, accuracy, and executability.

[0053] S300: Fill the control parameters into the closest instruction template and splice it with related associated instructions to obtain a complete control instruction.

[0054] Specifically, the control parameters may be embedded in the instruction template using, but not limited to, string formatting technology, and the closest instruction template after the control parameters are filled in is spliced ​​with the related associated instructions.

[0055] In this step, after filling the control parameters into the closest instruction template, it is also possible to determine whether the control parameters in the closest instruction template are complete. If the control parameters are incomplete, the missing control parameters can be supplemented by preset values; the operation intention can be re-identified, the control parameters omitted in step S200 can be extracted, and added to the closest instruction template; the operation intention can also be re-identified, the control parameters omitted in step S200 can be extracted, and added to the closest instruction template, and the remaining missing control parameters can be supplemented by preset values. Finally, the closest instruction template after the control parameters are supplemented is spliced ​​with the relevant associated instructions to obtain a complete control instruction. The aforementioned preset values ​​can be parameters commonly used to execute a certain control instruction, such as parameters in the configuration file for restarting a certain device when performing a restart.

[0056] After this step, the present invention further includes S400, verifying the complete control instruction, which specifically includes:

[0057] The complete control instruction's parameter range, type, or control logic are verified according to pre-set rules. If the complete control instruction passes verification, the complete control instruction is obtained. If the control parameters in the complete control instruction are within the correct range, the type is correct, and the control logic is correct, the verification passes; otherwise, it fails. This ensures that the instruction is compatible with the device control rules and detects and resolves logical conflicts between instructions.

[0058] If the complete control instruction verification fails, the cause of the failure is inquired and the complete control instruction is modified according to the cause of the failure to obtain a modified complete control instruction. Among them, modifying the complete control instruction according to the cause of the failure includes modifying the control parameters to be within the correct range, correcting the type, and deleting the instructions with control logic conflicts, etc.

[0059] In some embodiments, the present invention may further include serializing the complete control instruction into a preset instruction format, and issuing the serialized complete control instruction to the business system so that the business system executes the serialized complete control instruction. The preset instruction format is an instruction format executable by the business system. A complete control instruction is a control instruction that has passed verification and has complete control parameters.

[0060] In some embodiments, the present invention further includes: conducting at least two rounds of dialogue with the user, wherein the aforementioned first form of multi-round dialogue mechanism does not involve intent recognition, but aims to obtain accurate matching results. The dialogue mechanism here involves intent recognition, wherein each round of dialogue identifies the operation intention, determines the operation intention and control parameters of the current round of dialogue based on the operation intention and control parameters of the related round of dialogue, and generates complete control instructions for the current round of dialogue. The multi-round dialogue mechanism here is to use the operation intention and control parameters of the context to obtain more accurate operation intentions and missing or inaccurate control parameters of this round, automatically complete the control parameters, and finally generate structured and executable system instructions. This mechanism reduces the user's cognitive burden on professional terminology and system structure, allowing non-technical users to complete the expression and execution of complex instructions through fuzzy natural language, thereby improving the system's universality and interaction efficiency.

[0061] like Figure 4 As shown, Figure 4 This is a flow chart of a second-form multi-round dialogue mechanism. The present invention provides an example of a second-form multi-round dialogue mechanism:

[0062] S1: The user sends the control language that describes the device control, and the system receives the control language and sends it to the embedded model.

[0063] S2: The embedding model converts the control language into a vector and sends it to the device control command center for matching in the control command rule library. The control command rule library mainly includes the operation intention vector library, the command template vector library, and the associated command vector library.

[0064] S3: After receiving the user input vector, the device control command center performs a vector similarity search in the control command rule library. It then concatenates the retrieved operation intention vector, the initial command template vector, and the initial associated command vector to generate a "Context." This context is then sent to the "Big Language Model" for deeper understanding and command generation.

[0065] S4: The large language model receives the Context assembled by the device control command center, generates an accurate control command Request (steps S200-S300) and sends it to the "business system".

[0066] S5: The business system processes the control instruction and returns a Response.

[0067] S6: The large language model regenerates the user-oriented "natural language (Response)" based on the Response.

[0068] S7: By combining context and response information, the embedded model converts the user's subsequent control language (confirmation, modification, query) into a vector (including context information).

[0069] S8: The device control command center reassembles the data to form a context and sends it to the big prediction model.

[0070] S9: The large language model generates a new control instruction Request based on the new context (including the previous control intention and control parameters).

[0071] S10: After the business system executes the control instruction, it returns a new Response.

[0072] S11: Repeat the above steps. Finally, the large language model summarizes and organizes the responses generated by each round of dialogue, and outputs a complete natural language feedback to the user, including the execution results and execution content.

[0073] In the above step S5 or S10, the business system may add a verification process. After the control instruction is generated, if the control instruction is not executable, the user may be prompted (through the response of the control instruction) that the current control instruction is invalid, thereby triggering the next round of dialogue.

[0074] In some embodiments, as Figure 5 As shown, Figure 5The present invention provides a system for generating control instructions based on a language model. The system includes an embedded model, a language model, and a business system.

[0075] The embedding model is used to receive the control language input by the user, match the control language in the control instruction rule library, and obtain a matching result; the matching result includes one or more of the operation intention vector, the initial instruction template vector, and the initial associated instruction vector;

[0076] The language model is used to identify the user's operation intention and control parameters based on the matching results, and determine the closest instruction template and related associated instructions based on the operation intention; the control parameters are filled into the closest instruction template and spliced ​​with the related associated instructions to obtain a complete control instruction;

[0077] Business systems used to process and / or execute complete control instructions.

[0078] Specifically, the embedding model can be, but is not limited to, the GloVe (Global Vectors for Word Representation) model, the ELMo (Embeddings from Language Models) model, or the BERT (Bidirectional Encoder Representations from Transformers) model. The language model can be, but is not limited to, a large language model. The business system can be, but is not limited to, a computer integration system, a computer service system in an industrial control center, or a server cluster comprising multiple servers.

[0079] In some embodiments, such as Figure 6 As shown, Figure 6 1 is a structural diagram of an electronic device provided by the present invention. The present invention also provides an electronic device, which includes a processor 10 and a memory 11, wherein the memory 11 stores a computer program, and when the processor 10 executes the computer program, it implements any one of the methods described in the above method embodiments.

[0080] Among them, the memory is a non-transient computer-readable storage medium that can be used to store non-transient software programs and non-transient computer executable programs. The memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory optionally includes a remote memory remotely arranged relative to the processor, and these remote memories can be connected to the processor via a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0081] It is understood that all or some steps, systems in the disclosed method above can be implemented as software, firmware, hardware and appropriate combinations thereof. Some physical components or all physical components can be implemented as software by a processor, such as a central processing unit, a digital signal processor or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include computer storage media (or non-transitory media) and communication media (or temporary media). As known to those of ordinary skill in the art, the term computer storage medium is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules or other data) and is volatile and non-volatile, removable and non-removable media. Computer storage media includes but is not limited to RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassette, magnetic tape, disk storage or other magnetic storage device, or can be used to store desired information and any other medium that can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0082] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A control instruction generation method based on a language model, characterized in that: The following steps are involved: receiving a control language input by a user, and matching the control language in a control instruction rule library to obtain a matching result; the matching result includes one or more of an operation intention vector, an initial instruction template vector, and an initial associated instruction vector; Identifying the user's operation intention and control parameters based on the matching results, and determining the closest instruction template and related associated instructions based on the operation intention; The control parameters are filled into the closest instruction template and spliced ​​with the related associated instructions to obtain a complete control instruction.

2. The method according to claim 1, characterized in that The method further includes: determining whether the control parameters are complete; if the control parameters are incomplete, supplementing the missing control parameters by using preset values ​​and / or re-identifying the missing control parameters in the operation intention and supplementing the missing control parameters.

3. The method according to claim 1, characterized in that The method further includes: verifying the parameter range, type or control logic of the complete control instruction according to a preset rule, and obtaining the complete control instruction if the complete control instruction passes the verification; If the complete control instruction fails to be verified, the reason for the failure is inquired and the complete control instruction is corrected according to the reason for the failure to obtain a corrected complete control instruction.

4. The method according to claim 1, wherein The identifying the user's operation intention according to the matching result specifically includes: Extracting intent information based on the matching result, and identifying a plurality of sub-intents and association relationships between the sub-intents based on the intent information; An intention logic diagram is established based on the sub-intentions and the association relationship to obtain the user's operation intention; wherein, the nodes of the intention logic diagram represent sub-intentions, and the edges between the nodes represent the association relationship between two sub-intentions.

5. The method according to claim 4, characterized in that The determining of the closest instruction template and related associated instructions according to the operation intention specifically includes: Traverse the intention logic graph and determine the closest instruction template and related associated instructions for each sub-intention.

6. The method according to any one of claims 1 to 5, characterized in that The method also includes: serializing the complete control instruction into a preset instruction format, and sending the serialized complete control instruction to the business system so that the business system executes the serialized complete control instruction; the preset instruction format is an instruction format executable by the business system.

7. The method according to claim 6, characterized in that The method further includes: conducting at least two rounds of dialogue with the user, determining the operation intention of the current round of dialogue based on the operation intention and control parameters of the related rounds of dialogue, and generating complete control instructions for the current round of dialogue.

8. The method according to any one of claims 1 to 5, characterized in that Confirm whether the operation intention is a risk intention. If the operation intention is a risk intention, generate a risk prompt and return the risk prompt to the user.

9. A control instruction generation system based on a language model, characterized in that: The system includes an embedding model, a language model and a business system, wherein: The embedding model is used to receive a control language input by a user, and perform matching in a control instruction rule library according to the control language to obtain a matching result; the matching result includes one or more of an operation intention vector, an initial instruction template vector, and an initial associated instruction vector; The language model is used to identify the user's operation intention and control parameters based on the matching results, and determine the closest instruction template and related associated instructions based on the operation intention; fill the control parameters into the closest instruction template and splice it with the related associated instructions to obtain a complete control instruction; The business system is used to process and / or execute the complete control instruction.

10. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 8 when executing the computer program.

Images