Network management and control interaction method and device based on natural language

Abstract

This application discloses a network control interaction method and apparatus based on natural language. The method includes: a first agent using a pre-trained natural language generation model to convert the current network control intent into a natural language instruction and send it to a second agent; the second agent using a pre-trained natural language parsing model to extract the network control intent and corresponding slot information from the received natural language instruction to obtain corresponding network configuration parameters. Using this application can improve the scalability and robustness of network control interaction.

Description

Technical Field

[0001] This invention relates to network management technology, and in particular to a network control and interaction method and apparatus based on natural language. Background Technology

[0002] With the continuous rise in network demands and the ongoing development of network communication technology, future network communication will inevitably enter an era of intelligence and autonomy. Traditional network elements, primarily functioning as transmitters, receivers, and relays, will gradually evolve into new types of network intelligent agents with intelligent and autonomous interaction and decision-making capabilities. Intelligent interaction requires a unified interaction language and protocol suite as a foundation, while also meeting performance requirements such as high robustness, high efficiency, and high scalability. Scholars have conducted numerous effective studies on the question of which language should be used for interaction between multiple network elements.

[0003] Traditional network management solutions use artificial languages ​​(i.e., languages ​​that conform to the syntax specifications defined by network management protocols) to conduct interactions between network elements. For example, the Simple Network Management Protocol (SNMP) encapsulates the interaction messages between the manager and the agent into a standard language by defining Protocol Data Units (PDUs) and Basic Encoding Rules (BERs). The sender and receiver of the message only need to follow the pre-defined syntax specifications to complete the encapsulation and parsing of the message.

[0004] In the process of developing this invention, the inventors discovered that if traditional artificial language interaction methods are still used between novel network intelligent agents, there will be problems with poor scalability and robustness. A detailed analysis follows:

[0005] First, artificial languages ​​need to be implemented according to established grammatical rules. Once the grammatical rules are determined, the types of messages they represent will also be determined. When a new message requirement is raised, a corresponding grammatical rule needs to be generated for the new message requirement in order to meet the expression of the new message requirement. Therefore, due to the dependence on the grammatical rules of artificial languages, there is a problem of poor scalability.

[0006] Second, when transmitting messages based on artificial language through physical channels, error bits may appear due to channel noise. This can easily lead to the receiver either being unable to parse the message or parsing it incorrectly. Therefore, due to the presence of physical channel noise, there is a problem of poor robustness. Summary of the Invention

[0007] In view of this, the main objective of the present invention is to provide a network management and control interaction method and apparatus based on natural language, which can improve the scalability and robustness of network management and control interaction.

[0008] To achieve the above objectives, the technical solution proposed in this embodiment of the invention is as follows:

[0009] A natural language-based network management and interaction method includes:

[0010] The first agent uses a pre-trained natural language generation model to convert the current network control intent into natural language instructions and send them to the second agent.

[0011] The second agent uses a pre-trained natural language parsing model to extract the network control intent and corresponding slot information from the received natural language instructions, thereby obtaining the corresponding network configuration parameters.

[0012] This invention also proposes a network management and control interaction device based on natural language, comprising:

[0013] The natural language generation module is used to convert the current network control intent into natural language instructions and send them to the second intelligent agent using a pre-trained natural language generation model.

[0014] The natural language parsing module is used to extract network control intent and corresponding slot information from the natural language instructions received by the second agent using a pre-trained natural language parsing model, and obtain the corresponding network configuration parameters.

[0015] This invention also proposes a network management and control interaction device based on natural language, including a processor and a memory;

[0016] The memory stores an application program that can be executed by the processor, which enables the processor to execute the network management interaction method based on natural language as described above.

[0017] This invention also proposes a computer-readable storage medium storing computer-readable instructions for executing the natural language-based network management interaction method described above.

[0018] In summary, in the network control interaction scheme based on natural language proposed in this embodiment of the invention, the network control decision sending agent (i.e., the first agent) uses a pre-trained natural language generation model to convert the current network control intent into a natural language instruction and sends it to the network control decision execution agent (i.e., the second agent). Upon receiving the natural language instruction, the execution agent uses a pre-trained natural language parsing model to extract the network control intent and corresponding slot information from the natural language instruction, thereby obtaining the corresponding network configuration parameters for the natural language instruction. Thus, since the agents transmit natural language instructions with semantic information, even if transmission noise causes erroneous bits, it does not affect the correct understanding of the semantic information of the natural language instruction. Therefore, the semantic information of the natural language instruction can be used to reduce the impact of erroneous bits caused by transmission noise, obtaining the correct network control intent and key slot information. This allows for the use of the semantic robustness of natural language (i.e., the language used for human-to-human interaction) to improve the robustness of network control interaction. Furthermore, since there is no need to pre-design the interaction syntax and instruction set for intelligent agent interaction, but instead utilizes offline trained natural language generation and parsing models, natural language instructions containing control intentions are automatically generated at the sending end and sent to the receiving intelligent agent; at the receiving end, the natural language instructions are automatically parsed into specific network configuration actions. In this way, the powerful generation capabilities of the language model can be leveraged to meet the expression of new message requirements, thereby resisting the limitations of existing artificial protocol languages ​​and improving the scalability of network control interaction. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram illustrating a specific implementation of an embodiment of the present invention;

[0021] Figure 3 This is a schematic diagram of the device structure according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Figure 1 This is a schematic diagram of the network management and control interaction method based on natural language according to an embodiment of the present invention, such as... Figure 1 As shown, this embodiment mainly includes the following steps:

[0024] Step 101: The first agent uses a pre-trained natural language generation model to convert the current network control intent into a natural language instruction and sends it to the second agent.

[0025] Unlike existing network management protocols, this step involves the message-sending agent (the first agent) and the message-receiving agent (the second agent) interacting with natural language commands containing network management semantics, rather than network management commands constructed based on the syntax of a specific network management protocol. This leverages the semantic robustness of natural language commands to enhance the robustness of network management interactions. Furthermore, it utilizes the powerful generative capabilities of language models to improve the scalability of network management interactions, flexibly and conveniently meeting new message requirements. Unlike existing network management interaction schemes, which require updating the network management protocol and generating corresponding syntax for new message requirements, this significantly improves the scalability of network management interactions.

[0026] In practical applications, existing network communication protocols can be used to send natural language commands for network management to the second intelligent agent, which will not be elaborated here.

[0027] In one implementation, the natural language generation model can be trained in advance using the following steps a1 to a4:

[0028] Step a1: Input the network control intent sequence of the training sample into the first encoder of the natural language generation model for encoding to obtain the intent representation vector i.

[0029] This step is used to encode the network control intent sequence based on the training samples to obtain the corresponding intent representation vector i.

[0030] The first encoder can be specifically implemented as a bidirectional long short-term memory (LSTM) network.

[0031] For example, for a multi-intent control instruction "Move&Request_Report", the corresponding intent representation vector i, i∈R can be encoded. d×1 d is the length of the LSTM hidden layer vector.

[0032] Step a2: Input each intent and associated slot information in the network control intent sequence into the second encoder of the natural language generation model for encoding to obtain the intent and slot information representation vector s.

[0033] This step is used to encode the network control intent sequence based on the training samples and the slot information associated with each intent, to obtain the intent and slot information representation vector s.

[0034] The second encoder can be specifically implemented as a bidirectional long short-term memory network.

[0035] In this step, each intent and its associated slot information (including slot type and slot value) need to be concatenated together and output to the second encoder for encoding to obtain the corresponding intent and slot information representation vector s.

[0036] For example, the information input to the second encoder is "Move<Location=Left Front>&Request_Report<Info=SNR>", which, after encoding, yields an intent and slot information representation vector s, s∈R. d×1 d is the length of the LSTM hidden layer vector.

[0037] Step a3: After concatenating the intent representation vector i and the intent and slot information representation vector s, input the decoder of the natural language generation model for decoding. Based on the word distribution obtained from the decoding, obtain the natural language instruction corresponding to the network control intent sequence.

[0038] This step involves concatenating the hidden layer vectors i and s from the encoders in steps a1 and a2 above, inputting them into a decoder for decoding, and generating a word distribution p at each time step. t Thus, based on the word distribution obtained from decoding, the natural language instructions corresponding to the network control intent sequence can be obtained.

[0039] Step a4: Based on the output of the natural language generation model, semantic consistency constraints are applied to generate the total loss function of the natural language generation model.

[0040] In this step, in order to ensure that the natural language instructions can accurately express the network control intentions, based on the processing results obtained in steps a1 to a3, semantic consistency constraints are adopted to generate the total loss function of the natural language generation model.

[0041] In one implementation, step a4 can be specifically carried out according to the formula:

[0042] Generate the total loss function loss(θ) of the natural language generation model.

[0043] Where, p t The word distribution output by the natural language generation model, y t i represents the reference word distribution of the sample data. t+1 and i t Let i0 be the intent vectors to be expressed at time t+1 and time t, respectively, and T be the length of the word sequence output by the decoder. i0 is initialized with the intent representation vector i obtained by the intent encoder, i.e.: i0 = i, i T It is the intent vector to be expressed at the end of decoding. This is achieved through a controller σ. tRecord the intents that need to be decoded after time t, and complete i. t Update:

[0044] σ t =sigmoid(W y y t +W h h t-1 )

[0045] i t+1 =σ t ⊙i t

[0046] Among them, W y and W h For trainable model parameters, ⊙ represents pointwise multiplication of vectors.

[0047] τ and All are preset harmonic parameters, 0 < τ < 1. Specifically, those skilled in the art can set τ and within the above value range according to actual needs. A suitable value for τ is, for example, τ can be 0.90. It can be 1.01, but it is not limited to that.

[0048] The loss function described above consists of three parts, among which, For the correctness constraints of the decoder, To avoid redundant constraints, ||i T || represents the constraint of no omissions.

[0049] The aforementioned intention-driven natural language generation method, implemented using a natural language generation model, employs a sequence learning model to model and control the semantic sequence of intent representation. Furthermore, under the constraint of semantic consistency, it can further ensure that the generated natural language sentences accurately express the agent's interactive intent.

[0050] Step 102: The second agent uses a pre-trained natural language parsing model to extract the network control intent and corresponding slot information from the received natural language instructions, and obtains the corresponding network configuration parameters.

[0051] In this step, since the receiving agent receives natural language instructions, it is necessary to use a corresponding natural language parsing model to parse them, extract the network control intent and corresponding slot information, and thus obtain the corresponding network control configuration parameters.

[0052] Here, the natural language parsing model comprises two subtasks: intent recognition and slot labeling, which together constitute the semantic framework of natural language instructions. Intent recognition refers to identifying different types of control tasks, such as "ask" and "deny." Slot labeling refers to identifying the information slots contained in the natural language instruction and labeling their values, such as "IP address" and "SNR." Slot labeling, also known as slot filling, primarily uses sequence labeling methods to assign slot category labels to characters or words in the natural language instruction.

[0053] In actual command transmission, due to channel interference during communication, the text sequence received by the receiving agent may contain errors. Inputting a noisy text sequence into the natural language understanding module will cause deviations in the decoded natural language sequence. For example, "query signal-to-noise ratio" may be incorrectly decoded as "query signal-to-noise ratio", and "boot up network" may be incorrectly decoded as "boot up network".

[0054] To address the aforementioned issues and further improve the robustness of network control interactions, one implementation method involves pre-training a natural language parsing model using adversarial learning. This simultaneously enhances the robustness of natural language understanding through both large-scale pre-trained language models and adversarial learning. Preferably, this can be achieved using the following steps b1-b4.

[0055] Step b1: Input the sentence sequence X of the training samples into the encoder of the natural language parsing model to encode the sentence sequence and obtain the semantic vector h of the sentence sequence X. [cls] .

[0056] The encoder in this step can specifically be implemented as a BERT encoder.

[0057] Assume, sentence sequence Where T represents the length of the sentence sequence, the sentence is encoded by the BERT encoder to obtain the encoded output hidden layer vector. Specifically, it can be represented by the following formula:

[0058] H = BERT(X)

[0059] Among them, h [cls] ∈R d×1 This represents the semantic encoding result (i.e., semantic vector) of the sentence sequence, where d is the length of the semantic vector used by the encoder. This semantic encoding will be used in subsequent steps for sentence denoising and parsing.

[0060] Step b2: Transfer the semantic vector h [cls]The intention recognition network of the natural language parsing model is input to predict the combination of intentions and slot types contained therein; each combination of intentions and slot types consists of an intention and a slot type associated with that intention.

[0061] This step is used to identify the control intent and the slot type bound to the intent in natural language instructions. The corresponding model task is a sentence-to-intent-slot multi-classification task.

[0062] The intent recognition network can be specifically implemented as a fully connected network layer MLP. intent-slot Accordingly, this step requires extracting the semantic vector h of the sentence. [cls] The input is fed into the fully connected network layer MLP. intent-slot In this process, the possible intent-slot types are predicted, and this prediction can be specifically represented by the following formula:

[0063] I o =σ(W hi h [cls] +b i )

[0064] Among them, W hi and b i Here are the trainable model parameters, σ represents the activation function, and I... o ={I1, I2, ..., I I}, where I represents the total number of intents-slots. Let I... o Input the softmax function to obtain the predicted probability p for each intent-slot type. i .

[0065] Step b3, based on the semantic vector h [cls] The slot value is predicted for each intent-slot type combination using the slot value recognition network of the natural language parsing model.

[0066] This step generates corresponding slot values ​​for the identified intent-slot type. Given a sentence sequence X and an intent-slot type, predict the slot values ​​for that intent-slot type.

[0067] In one implementation, the slot identification network can be implemented as a classifier MLP. value For each type of intent-slot type I n Set up a classifier MLP value To predict the slot value corresponding to this intent-slot. The prediction in this step can be specifically represented by the following formula:

[0068] S o =σ(W hs h i_[cls] +b s )

[0069] Among them, W hs and b s Here are the trainable model parameters, σ represents the activation function, and S... o ={S1, S2, ..., S V}, where V represents the number of all candidate slot values. h i_[cls] To convert the intention-slot type I n The corresponding category label serves as the prefix of the sentence sequence X, and the semantic vector is obtained after BERT encoding. S o The softmax function is used to obtain the probability p of the candidate slot value corresponding to each intent-slot type. v .

[0070] Step b4: Based on the semantic vector h [cls] The noise text corresponding to the sentence sequence X semantic vectors The intent, slot type, and slot value are combined to calculate the total loss function value of the natural language parsing model; the total loss function value is then used to optimize and adjust the parameters of the natural language parsing model; wherein, the noisy text The semantic vector is generated based on the sentence sequence X in a noisy environment. For the noise text The encoder input is used to encode the data.

[0071] Here, to enhance the model's resilience to noise, the training for adversarial tasks needs to be considered when calculating the total loss function value of the natural language parsing model, i.e., based on the semantic vector h. [cls] and corresponding noisy text semantic vectors Calculate the loss function to obtain an encoder with noise reduction capabilities, so that the model can output the original correct instructions (i.e., natural language instructions before noise is added), thereby further improving the model's semantic understanding and symbol reconstruction capabilities in noisy environments.

[0072] In practical applications, the noisy text Existing methods can be used to generate it. Preferably, in one embodiment, the following steps d1 to d4 can be used to generate the noisy text.

[0073] Step d1: Convert the sentence sequence X into a 0-1 bit stream.

[0074] The aforementioned 0-1 bitstream is a bitstream composed of bits 1 and bits 0. Specifically, the sentence sequence X can be converted into a 0-1 bitstream using Unicode encoding, but it is not limited to this. Other methods that can obtain a 0-1 bitstream can also be used, which will not be elaborated here.

[0075] Step d2: Modulate the 0-1 bit stream using a sine wave.

[0076] Step d3: Using a channel that has introduced fading and noise, transmit the modulation result to the target receiver.

[0077] Step d4: The target receiver demodulates the received modulation result and decodes the demodulated result to obtain the noise text.

[0078] Using the above steps, in a noisy environment, the generated sentence sequence X corresponds to the noisy text. Next, the noisy text The input encoder encodes the data, i.e. Obtain the encoded output hidden layer vector Thus, the noisy sentence sequence can be obtained. semantic vectors

[0079] In one implementation, the total loss function value of the natural language parsing model can be calculated using the following method:

[0080] Step c1: Calculate the semantic vector h [cls] and the semantic vector The cosine loss function is used to obtain the first loss function.

[0081] This step is used to calculate the semantic vector h. [cls] and semantic vectors The cosine loss is used as the similarity loss function. Specifically, it can be represented by the following formula:

[0082]

[0083] Among them, L SIM Let h represent the first loss function. i Represents the semantic vector h [cls] The i-th element in Represents semantic vectors The i-th element in the vector, d, represents the semantic vector h. [cls] or The number of elements in the data.

[0084] Step c2: Calculate the second loss function value based on the combination of intent and slot type output by the intent recognition network.

[0085] For ease of implementation, in one embodiment, the cross-entropy loss function can be used to calculate the value of the second loss function. The specific formula for calculating the loss function is as follows:

[0086]

[0087] Among them, L Intent For the second loss function, p i For the predicted probability of intent-slot type, y i Let I be the true probability of an intent-slot type, and I be the number of all intent-slot types.

[0088] In practical applications, other loss functions can also be used for calculation, which will not be elaborated here.

[0089] Step c3: Calculate the third loss function value based on the slot value output by the intent recognition network.

[0090] For ease of implementation, in one embodiment, the cross-entropy loss function can be used to calculate the value of the third loss function. The specific formula for calculating the loss function is as follows:

[0091]

[0092] Among them, L Value For the third loss function, p v y is the predicted probability of the candidate slot value. v Let V be the true probability of a candidate slot value, and let V represent the total number of candidate slot values.

[0093] In practical applications, other loss functions can also be used for calculation, which will not be elaborated here.

[0094] Step c4: Summing the first loss function, the second loss function, and the third loss function to obtain the total loss function.

[0095] This step can be specifically represented by the following formula:

[0096] L = L Intent +L Value +L SIM

[0097] Where L is the total loss function.

[0098] Based on the pre-training method of the above natural language parsing model, it can be seen that when training the natural language parsing model, the input is a natural language instruction with physical channel noise introduced. The semantic representation of the instruction is obtained through the BERT coding layer, and the output includes: the instruction intent identified from the noisy instruction, the slot value identified from the noisy instruction, and the original correct instruction reconstructed from the noisy instruction. This natural language parsing model has the following advantages:

[0099] (1) Through multi-task learning, the similar representations of multiple related target tasks are effectively utilized, prompting the model to focus on the representation learning of general features of the task in the underlying encoder and on the representation learning of task-related features in the encoder part of each target task.

[0100] (2) By correctly identifying the instruction intent, slot value and restoring the noise input from the noise input, the model gains the ability to resist physical channel noise, thereby saving the error control overhead in traditional protocols and improving the transmission efficiency of instructions.

[0101] In summary, the pre-training method for the natural language parsing model described above, along with the multi-task learning framework, not only forces the natural language parsing model to acquire key semantic information, but also enables it to resist physical channel noise to the greatest extent possible, thereby improving the robustness of natural language-based network management protocols.

[0102] As can be seen from the above embodiments of the network management and control interaction method based on natural language, this invention utilizes pre-trained natural language generation and parsing models to transmit semantically informative natural language commands between agents, achieving human-like interaction among multiple network agents around network management and control tasks. Thus, on the one hand, the semantic robustness of natural language commands can be leveraged to improve the robustness of network management and control interaction; on the other hand, the powerful generation capabilities of language models can be utilized to meet the expression of new message requirements, effectively enhancing the scalability of network management and control interaction.

[0103] Figure 2 This is a schematic diagram illustrating a specific implementation of an embodiment of the present invention in a particular scenario. For example... Figure 2As shown, the sending agent uses a natural language generation model to encapsulate control intentions into natural language instructions, which are then converted into 0-1 bit sequences suitable for propagation through a physical channel by an encoder. These sequences are then transmitted to the receiving agent via the physical channel. At the receiving end, the natural language instructions are decoded, and the receiving agent uses a natural language parsing model to extract network configuration parameters, thus completing the network configuration. Multiple sending and receiving agents in the network share the same natural language generation and parsing models, endowing the network agents with the ability to express and understand complex task and environmental semantics, thereby enhancing their autonomous interaction capabilities.

[0104] Based on the above embodiments of the network management and control interaction method based on natural language, the present invention correspondingly proposes a network management and control interaction device based on natural language, such as... Figure 3 As shown, the device includes:

[0105] Natural language generation module 301 is used to convert the current network control intent into natural language instructions and send them to the second intelligent agent using a pre-trained natural language generation model;

[0106] The natural language parsing module 302 is used to extract the network control intent and corresponding slot information from the natural language instructions received by the second agent using a pre-trained natural language parsing model, and obtain the corresponding network configuration parameters.

[0107] The aforementioned natural language generation module 301 and natural language parsing module 302 are used for encapsulating sent information and parsing received information, respectively, and are therefore located in the sending agent and the receiving agent, respectively. Considering that in practical applications, an agent may often need to both send and receive natural language instructions, both modules will also be deployed simultaneously (e.g., ...). Figure 3 The middle modules 303 and 304 are used to process the sending and receiving of information, respectively.

[0108] It should be noted that the above methods and apparatus are based on the same inventive concept. Since the methods and apparatus solve problems in similar ways, the implementation of the apparatus and methods can refer to each other, and the repeated parts will not be described again.

[0109] Based on the above embodiments of the network management and control interaction method based on natural language, this invention also proposes a network management and control interaction device based on natural language, including a processor and a memory; the memory stores an application program that can be executed by the processor, for causing the processor to execute the network management and control interaction method based on natural language as described above. Specifically, a system or device equipped with a storage medium can be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or device can read and execute the program code stored in the storage medium. In addition, the operating system or other operating system on the computer can be instructed to perform some or all of the actual operations through instructions based on the program code. The program code read from the storage medium can also be written to a memory set in an expansion board inserted into the computer or to a memory set in an expansion unit connected to the computer, and then the CPU or other operating system installed on the expansion board or expansion unit can be instructed to perform some or all of the actual operations based on the instructions of the program code, thereby realizing the functions of any of the above embodiments of the network management and control interaction method based on natural language.

[0110] Specifically, the memory can be implemented as various storage media such as electrically erasable programmable read-only memory (EEPROM), flash memory, and programmable programmable read-only memory (PROM). The processor can be implemented as one or more central processing units (CPUs) or one or more field-programmable gate arrays (FPGAs), wherein the FPGA integrates one or more CPU cores. Specifically, the CPU or CPU core can be implemented as a CPU or an MCU.

[0111] This application also implements a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the network management interaction method based on natural language as described above.

[0112] It should be noted that not all steps and modules in the above processes and structural diagrams are mandatory; some steps or modules can be omitted as needed. The execution order of the steps is not fixed and can be adjusted as required. The division of modules is merely for the convenience of description and functional division. In actual implementation, a module can be implemented by multiple modules, and the functions of multiple modules can also be implemented by the same module. These modules can be located in the same device or in different devices.

[0113] The hardware modules in each embodiment can be implemented mechanically or electronically. For example, a hardware module may include specially designed permanent circuitry or logic devices (such as dedicated processors, such as FPGAs or ASICs) to perform specific operations. A hardware module may also include programmable logic devices or circuitry (such as general-purpose processors or other programmable processors) temporarily configured by software to perform specific operations. The choice between mechanical implementation, dedicated permanent circuitry, or temporarily configured circuitry (such as software-configured circuitry) can be made based on cost and time considerations.

[0114] In this document, "illustrative" means "serving as an example, illustration, or description," and any illustration or embodiment described herein as "illustrative" should not be construed as a preferred or more advantageous technical solution. For the sake of brevity, the figures only schematically represent the parts relevant to the invention and do not represent their actual structure as a product. Furthermore, for the sake of clarity and ease of understanding, in some figures, components with the same structure or function are only schematically depicted, or only one is labeled. In this document, "a" does not mean that the number of relevant parts of the invention is limited to "only one," and "a" does not exclude the possibility that the number of relevant parts of the invention is "more than one." In this document, terms such as "upper," "lower," "front," "rear," "left," "right," "inner," and "outer" are used only to indicate the relative positional relationship between relevant parts, and not to limit the absolute position of these relevant parts.

[0115] The solutions described in this specification and embodiments, if involving the processing of personal information, will be processed only under the premise of having a legal basis (such as obtaining the consent of the personal information subject, or being necessary for the performance of a contract), and will only be processed within the scope stipulated or agreed upon. A user's refusal to process personal information beyond what is necessary for basic functions will not affect the user's use of basic functions.

[0116] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A network management and control interaction method based on natural language, characterized in that, include: The first agent uses a pre-trained natural language generation model to convert the current network control intent into natural language instructions and send them to the second agent. The second intelligent agent uses a pre-trained natural language parsing model to extract network control intent and corresponding slot information from the received natural language instructions, thereby obtaining the corresponding network configuration parameters; wherein, the pre-training of the natural language generation model includes: The network control intent sequence of the training samples is input into the first encoder of the natural language generation model for encoding to obtain the intent representation vector i. Each intent and associated slot information in the network control intent sequence is input into the second encoder of the natural language generation model for encoding to obtain an intent and slot information representation vector s; The intent representation vector i and the intent and slot information representation vector s are concatenated and then input into the decoder of the natural language generation model for decoding. Based on the word distribution obtained from the decoding, the natural language instruction corresponding to the network control intent sequence is obtained. Based on the output of the natural language generation model, semantic consistency constraints are applied to generate the total loss function of the natural language generation model. The natural language parsing model is pre-trained using adversarial learning, including: input a sentence sequence X of a training sample into an encoder of the natural language parsing model, encode the sentence sequence to obtain a semantic vector of the sentence sequence X ; The semantic vector The intention recognition network of the natural language parsing model is input to predict the combinations of intentions and slot types it contains; each combination of intentions and slot types consists of an intention and a slot type associated with that intention. Based on the semantic vector Using the slot value recognition network of the natural language parsing model, the slot value corresponding to each combination of intent and slot type is predicted; Based on the semantic vector The noise text corresponding to the sentence sequence X semantic vectors The intent, slot type combination, and slot value are used to calculate the total loss function value of the natural language parsing model; the total loss function value is then used to optimize and adjust the parameters of the natural language parsing model; wherein, the noisy text The semantic vector is generated based on the sentence sequence X in a noisy environment. For the noise text The encoder input is used to encode the data.

2. The method of claim 1, wherein, according to Generate the total loss function of the natural language generation model. ;in, The word distribution output by the natural language generation model. For the reference word distribution of the sample data, The first The vector of intent to be expressed at any given moment Let be the vector of intent to be expressed at time t+1. and All are preset harmonic parameters. , T is the length of the word sequence output by the decoder.

3. The method of claim 1, wherein, The calculation of the total loss function value of the natural language parsing model includes: computing the semantic vector and a cosine loss function of the semantic vector to obtain a first loss function; The second loss function value is calculated based on the combination of intent and slot type output by the intent recognition network. Based on the slot value output by the intent recognition network, calculate the value of the third loss function; The total loss function is obtained by summing the first loss function, the second loss function, and the third loss function.

4. The method of claim 3, wherein, The generation of the noisy text includes: Convert the sentence sequence X into a 0-1 bit stream; The 0-1 bit stream is modulated using a sine wave; The modulation result is transmitted to the target receiver using a channel that introduces fading and noise. The target receiving end demodulates the received modulation result and decodes the demodulation result to obtain the noise text .

5. A natural language based network management interaction device, characterized in that, include: The natural language generation module is used to convert the current network control intent into natural language instructions and send them to the second intelligent agent using a pre-trained natural language generation model. The natural language parsing module is used to extract network control intent and corresponding slot information from the natural language instructions received by the second agent using a pre-trained natural language parsing model, and obtain the corresponding network configuration parameters. The pre-training of the natural language generation model includes: The network control intent sequence of the training samples is input into the first encoder of the natural language generation model for encoding to obtain the intent representation vector i. Each intent and associated slot information in the network control intent sequence is input into the second encoder of the natural language generation model for encoding to obtain an intent and slot information representation vector s; The intent representation vector i and the intent and slot information representation vector s are concatenated and then input into the decoder of the natural language generation model for decoding. Based on the word distribution obtained from the decoding, the natural language instruction corresponding to the network control intent sequence is obtained. Based on the output of the natural language generation model, semantic consistency constraints are applied to generate the total loss function of the natural language generation model. The natural language parsing model is pre-trained using adversarial learning, including: input a sentence sequence X of a training sample into an encoder of the natural language parsing model, encode the sentence sequence to obtain a semantic vector of the sentence sequence X ; The semantic vector The intention recognition network of the natural language parsing model is input to predict the combinations of intentions and slot types it contains; each combination of intentions and slot types consists of an intention and a slot type associated with that intention. based on the semantic vector , a slot value identification network using the natural language analysis model, predicts a slot value corresponding to each of the intent and slot type combination; Based on the semantic vector The noise text corresponding to the sentence sequence X semantic vectors The intent, slot type combination, and slot value are used to calculate the total loss function value of the natural language parsing model; the total loss function value is then used to optimize and adjust the parameters of the natural language parsing model; wherein, the noisy text The semantic vector is generated based on the sentence sequence X in a noisy environment. For the noise text The encoder input is used to encode the data.

6. A natural language based network management interaction device, characterized in that, Including processor and memory; The memory stores an application program that can be executed by the processor, which causes the processor to execute the natural language-based network management interaction method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores computer-readable instructions for performing the natural language-based network management interaction method as described in any one of claims 1 to 4.

Images