A multi-cloud storage heterogeneous user intent recognition method and device based on LLM

Abstract

The application provides a multi-cloud storage heterogeneous user intention recognition method and device based on LLM, and belongs to the technical field of cloud computing. The method comprises the following steps: collecting user demand recognition data sets through a multi-cloud platform, and setting a unified data management format; constructing a professional keyword dictionary, and judging whether a user is a professional user according to user query content; dividing the data sets into four sub-data sets according to the interaction information field and the user background field in the data management format; performing data enhancement on the non-professional user sub-data set until the data set coverage reaches a set threshold; training an LLM model using the sub-data sets D1 and D2 and the data-enhanced sub-data sets D3 and D4 to obtain a fine-tuned model F, which is used for multi-cloud storage heterogeneous user intention recognition. The application takes into account the multi-source heterogeneous multi-cloud environment and the diversity of user intentions, and significantly improves the intention recognition ability of heterogeneous users in the multi-cloud storage environment.

Description

TECHNICAL FIELD

[0001] The application relates to an LLM-based multi-cloud storage heterogeneous user intention recognition method and device, and belongs to the technical field of cloud computing. BACKGROUND

[0002] In the field of cloud computing, in the face of the demand for future service customization network for cloud network integration, a multi-cloud platform emerges as the times require. Cloud resources usually include three categories of computing, storage and network. Among them, storage resources are an important component. At present, the management and control mode of cloud resources such as storage is mainly manual, which requires higher professional knowledge of users and is low in efficiency. Therefore, the intelligent management and control problem of multi-cloud storage needs to be solved urgently.

[0003] The prerequisite for intelligent management and control of multi-cloud storage is to correctly identify the real intention implied by the user. Multi-cloud platform users are not only professional personnel, but also non-professionals. These two types of users need the assistance of multi-cloud platform R&D personnel to a greater or lesser extent when deploying the multi-cloud platform. R&D personnel need to help users convert their intentions into specific multi-cloud parameters for the use of the multi-cloud platform. Due to the differences in user industry, knowledge background and application scenarios, the expression of their intentions is diverse.

[0004] At present, the related research on intention recognition usually adopts a method based on reinforcement learning and non-generative deep learning. Although this method has certain learning ability, it is difficult to adapt to the multi-source heterogeneous multi-cloud environment. Large language model (LLM) is a powerful artificial intelligence algorithm that learns the grammar, semantics and context information of language by training a large amount of text data, so as to model natural language text. This model has a wide range of applications in the field of natural language processing (NLP), including text generation, text classification, machine translation, sentiment analysis, etc. At the same time, some researches have adopted a method based on LLM, but the related methods mostly rely on advanced general LLMs such as GPT 4.0 and use the prompt method. Although this method can complete the function of intention recognition, its cost and controllability are limited. SUMMARY

[0005] The application aims to provide an LLM-based multi-cloud storage heterogeneous user intention recognition method and device, which improves the intention recognition ability of heterogeneous users in the multi-cloud storage environment.

[0006] To achieve the above-mentioned purpose, the following technical solutions are adopted:

[0007] An LLM-based multi-cloud storage heterogeneous user intention recognition method, comprising the following steps:

[0008] A user demand recognition dataset is collected through a multi-cloud platform, and a unified data management format is set, including general fields, interaction information fields, and user background fields;

[0009] A professional keyword dictionary is constructed, and the user query content is compared with the professional keyword dictionary to calculate the proportion of professional vocabulary. If the proportion exceeds a set threshold, the user background field is marked as a professional user, otherwise it is marked as a non-professional user.

[0010] According to the interaction information fields and the user background fields in the data management format, the dataset is divided into four sub-datasets: a single intent of a professional user is a sub-dataset D1, a composite intent of a professional user is a sub-dataset D2, a single intent of a non-professional user is a sub-dataset D3, and a composite intent of a non-professional user is a sub-dataset D4. The data in each sub-dataset includes user natural language intent expression, storage mode, and basic task.

[0011] Data augmentation is performed on the single intent sub-dataset D3 of non-professional users and the composite intent sub-dataset D4 of non-professional users until the coverage of the dataset reaches a set threshold.

[0012] The sub-dataset D1, the sub-dataset D2, and the data-augmented sub-dataset D3 and the sub-dataset D4 are used to train the LLM model to obtain a fine-tuned model F for multi-cloud storage heterogeneous user intent recognition.

[0013] Preferably, the general fields include timestamp, platform, user ID, session ID, and request ID; the interaction information fields include interaction type, query content, intent number, and intent recognition; and the intent recognition classification format is storage mode-basic task.

[0014] Preferably, the storage mode includes block storage, object storage, and file storage, and the basic task includes single basic task and multiple basic tasks.

[0015] Preferably, the data augmentation includes syntax structure transformation, vocabulary transformation, introduction of text errors, and mixed use of upper and lower case.

[0016] The syntax structure transformation includes adjusting the order of subject, predicate, and object.

[0017] The vocabulary transformation includes synonym replacement and word redundancy to generate non-professional expressions.

[0018] The introduction of text errors includes errors that violate Chinese or English grammar rules in written expressions.

[0019] Preferably, the coverage calculation formula is as follows:

[0020] ,

[0021] wherein, the number of types of enhanced transformations, for the data set the number of unique patterns in the data set, describing the actual diversity of the data set, the number of data corresponding to different sets of enhanced transformations, respectively, is the weight, indicating the importance of the type of transformation, indicating the number of transformations of each type of enhanced transformation.

[0022] Preferably, the LLM model training loss function is as follows:

[0023] ,

[0024] wherein, is the standard cross-entropy loss of the professional user data set, is the standard cross-entropy loss of the non-professional user data set, is the semantic alignment loss of the professional user data set, is the semantic alignment loss of the non-professional user data set, is the confidence regulation loss, , is the dynamic weight, , and indicate the weight balance parameters, and are used to adjust the influence of the semantic alignment loss of the professional user and the non-professional user, respectively, indicates the weight of the non-professional user confidence regulation.

[0025] Preferably, the weight balance parameters are calculated as follows:

[0026] ,

[0027] ,

[0028] The weight balance parameters are calculated as follows:

[0029] ,

[0030] ,

[0031] wherein, , , , The proportion of the middle sample of the data sets D1, D2, D3 and D4 respectively;

[0032] The weight balance parameter takes a value range of: .

[0033] Preferably, the standard cross-entropy loss of the professional user data set is as follows:

[0034]

[0035] The standard cross-entropy loss of the non-professional user data set is as follows:

[0036]

[0037] Wherein, represents the number of samples in the small batch data for calculating the classification loss of the professional user data set, represents the true probability value of the jth sample in the kth class, represents the probability that the jth sample of the professional user data set belongs to the kth class, represents the number of samples in the small batch data for calculating the classification loss of the non-professional user data set, represents the probability that the jth sample of the non-professional user data set belongs to the kth class, represents the number of elements in the basic task set;

[0038] The alignment semantic loss of the professional user data set is as follows:

[0039] ,

[0040] Wherein, represents the number of samples in the small batch data for calculating the classification loss of the professional user data set, is the semantic embedding of the professional user sample, corresponds to the target semantic embedding of the professional user sample;

[0041] The alignment semantic loss of the non-professional user data set is as follows:

[0042] ,

[0043] Wherein represents the number of samples in the small batch data for calculating the classification loss of the non-professional user data set, is the semantic embedding of the professional user sample, corresponds to the target semantic embedding of the professional user sample;

[0044] The confidence regulation loss is as follows:

[0045] ,

[0046] wherein, is a confidence threshold, is the highest class confidence in the prediction for the jth sample of the model pair.

[0047] Preferably, the dynamic weight calculation method is as follows:

[0048] ,

[0049] wherein, is the current training round, is the total training round, is the professional user initial weight, is the non-professional user initial weight, and .

[0050] A data set construction device of an LLM-based multi-cloud storage heterogeneous user intent recognition method, comprising a processor and a memory storing program instructions, the processor being configured to execute the program instructions when running, to perform the LLM-based multi-cloud storage heterogeneous user intent recognition method.

[0051] A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the LLM-based multi-cloud storage heterogeneous user intent recognition method.

[0052] The advantages of the present application are:

[0053] (1) Multi-level user intent recognition

[0054] By modeling the multi-target intent recognition problem, the model can effectively deal with the diversified user requests in the multi-cloud storage environment. This method not only improves the accuracy of the model, but also enhances the recognition ability of the model for different types of user demands, so as to realize more efficient resource management and service optimization in complex environments. Especially for non-professional users, it can better understand and respond to their diversified demands.

[0055] (2) Special optimization for non-professional users

[0056] For spelling errors, colloquial expressions and non-standard language that may occur in non-professional users, this method improves the understanding ability of non-professional user intent through data enhancement technology. By introducing a dynamic loss adjustment strategy for non-professional users, it can flexibly adapt to the user's expression method, ensuring accurate intent recognition even in the case of non-standard language.

[0057] (3) Flexible coverage evaluation mechanism

[0058] To ensure the diversity and coverage of the dataset, an innovative coverage evaluation mechanism is proposed. By evaluating the diversity of the dataset in terms of grammar, vocabulary, and expression methods, it is ensured that the model can cover as wide a range of user expression scenarios as possible. This mechanism enables the model to handle various potential language biases or non-standard expressions when dealing with real-world applications, improving the stability and reliability of the model.

[0059] (4) Adaptive fine-tuning strategy

[0060] The method uses an adaptive fine-tuning loss function to dynamically adjust the loss weight according to the user's background and intent type at different stages. This fine-tuning strategy allows the model to flexibly adjust the optimization direction according to the characteristics of the training data, thereby avoiding the overfitting or underfitting problems that may be caused by fixed loss weights in traditional methods. This adaptive fine-tuning approach not only improves accuracy but also enhances the model's generalization ability, enabling it to handle various complex real-world scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0061] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation on the present application.

[0062] Figure 1 The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0064] Embodiment 1

[0065] As shown in Figure 1 , a multi-cloud storage heterogeneous user intent recognition method based on LLM includes the following steps:

[0066] S1: Collect user demand recognition dataset through multi-cloud platform, and set unified data management format, i.e. log management structure, including general field, interaction information field and user background field.

[0067] As a refinement of the above embodiment, the specific goal of data collection is identified as user demand recognition, and the multi-cloud platform includes Alibaba Cloud, Tencent Cloud, Huawei Cloud, AWS Cloud, Microsoft Cloud, and Google Cloud. The specific interaction data types that need to be collected are conversation content and search queries, etc. Using common log aggregation tools (such as Elastic Stack, Splunk) and log management services provided by cloud platforms (such as AWS CloudTrail, Google Cloud Logging, Azure Monitor, etc.), the same log management structure is set for conversation and query records. This structure includes general fields and interaction information fields.

[0068] The general fields are further divided into the following parts: timestamp, platform, user ID, session ID, and request ID. Among them, the platform indicates the cloud platform where the log comes from, such as AWS, Azure, Huawei, etc., in order to distinguish between multi-cloud sources.

[0069] The interaction information field is divided into the following parts:

[0070] Interaction type: distinguish the type of interaction, such as "dialog query", "search query", "API request", etc.

[0071] Query content: specific content input by the user, such as messages in a conversation, search keywords, etc.

[0072] Number of intents: classify user intents into composite intents and single intents.

[0073] Intent identification: the intent identification corresponding to the user input content. The classification format is: storage mode (M) - basic task (t). For example: object storage - increase cloud disk.

[0074] Specifically, the basic task set of multi-cloud platform storage resources is defined as According to the storage type of the platform, it can be divided into three storage modes: block storage, object storage, and file storage. The basic task set related to block storage has the following tasks: increase cloud disk, mount cloud disk, unmount cloud disk, expand cloud disk, delete cloud disk, etc. The basic task set related to object storage can be divided into the following tasks: increase object storage bucket, set storage bucket access control permission, delete object storage bucket, etc. The basic task set of file storage can be divided into the following tasks: increase access group, modify access group, delete access group, etc. Then, according to the division of storage types, the result of user natural language intent expression x after recognition may be:

[0075] , where , is the fine-tuned model.

[0076] Since The block storage basic task may contain a single basic task, or two or more block storage basic tasks. That is It is possible to have a single intention It may also be a composite intention Then It can be further expressed as:

[0077] .

[0078] Among them, Corresponding to a single basic task of the block storage resource of the large-scale multi-cloud exchange platform, Corresponding to two or more basic tasks of the block storage resource of the large-scale multi-cloud exchange platform.

[0079] Similarly, Corresponding to The single basic task of the platform object storage and file storage, Corresponding to Multi-basic task of platform object storage and file storage resources. Therefore Corresponding to It can be further expressed as: ; .

[0080] S2: Build a professional keyword dictionary, compare user query content with the professional keyword dictionary, calculate the proportion of professional vocabulary, and if it exceeds the set threshold, mark the user background field as a professional user, otherwise mark it as a non-professional user. The threshold in this embodiment is 10%. In the application process, the threshold can be modified as needed.

[0081] As a refinement of the above embodiment, it specifically includes the following steps:

[0082] Build a professional keyword dictionary. Collect and organize professional terms in the field of cloud computing, including technical documents, professional papers, industry reports, and API documents. Classify the vocabulary according to the storage mode to facilitate subsequent management and expansion.

[0083] Define the professional threshold, which is defined as the proportion of professional keywords exceeding the total number of query words. If the user query content is considered a professional query.

[0084] Iterate through the segmentation results, compare each word with the professional keyword dictionary, and record the matched professional words.

[0085] Count the number of professional keywords matched in the query content and calculate their frequency.

[0086] If the proportion P of professional words in the query content exceeds a set threshold, it is marked as a "professional user", otherwise it is marked as a "non-professional user".

[0087] Mark the user background field as a professional user, otherwise mark it as a non-professional user.

[0088] S3: According to the interaction information field and the user background field in the data management format, the data set is divided into four sub-data sets: single intent of professional users for sub-data set D1, composite intent of professional users for sub-data set D2, single intent of non-professional users for sub-data set D3, and composite intent of non-professional users for sub-data set D4; the data in each sub-data set contains user natural language intent expression, storage mode and basic task.

[0089] As a refinement of the above embodiment, the log management structure is formatted after division, wherein the formatted data set refers to the format of the labeled data set adopted by the reference model.

[0090] S4: Data augmentation is performed on the single intent of non-professional users for sub-data set D3 and the composite intent of non-professional users for sub-data set D4 until the coverage of the data set reaches a set threshold, which is 10% in this embodiment. In application, the threshold can be modified as needed.

[0091] The purpose is to improve the coverage and diversity of the non-professional user sub-data set.

[0092] As a refinement of the above embodiment, the augmentation transformation set is: .

[0093] (1) Syntax structure transformation Syntax structure transformation refers to disrupting the order of the subject, predicate and object of a sentence, such as "help me increase cloud disk" → "cloud disk increase help me".

[0094] (2) Lexical transformation It refers to generating non-professional expressions through synonym replacement and word redundancy. The difference between professional expressions and non-professional expressions mainly lies in the accuracy, technicality and professionalism of the terms. Professional expressions usually use industry-recognized terms and concepts, which have clear definitions and uses, and can accurately describe cloud computing-related technologies, services and concepts. Non-professional expressions may use more colloquial and more colloquial words, which may not be accurate, but are easier for non-professionals to understand. For example: the professional expression is cloud storage service, and the corresponding non-professional expression is stored on the Internet.

[0095] Specifically, the division between professional and non-professional expressions is based on whether the expression is a term in the cloud computing keyword dictionary, such as Microsoft's Azure cloud computing dictionary or IBM's cloud computing glossary.

[0096] Word redundancy enables the conversion of professional expressions into non-professional expressions by explaining professional terms in more colloquial language without sacrificing information transmission. This method includes synonym replacement and also includes definitions, explanations, examples, or analogies to help understand complex or professional terms. Example: The professional expression API is converted to "application programming interface" through word redundancy, which is like a bridge between different software, allowing them to "talk" and exchange information.

[0097] (3) Introduce textual errors This includes but is not limited to improper use of words, improper collocation, incomplete components, and redundant components.

[0098] For example: storage bucket → storage, NFS → NFC, help me increase the cloud disk → help increase the disk, I plan to store data in the cloud → I plan to store data in the network cloud.

[0099] The enhanced non-professional user data sets D3 and D4 are marked as .

[0100] The coverage index R is used to measure whether the enhanced data sets meet the requirements of coverage diversity and complexity. The coverage calculation formula is as follows:

[0101] ,

[0102] where, the number of types of expression enhancement transformations, such as grammar, vocabulary, spelling, and colloquialization, is the number of unique patterns in the data set , which describes the actual diversity of the data set, is the number of data in the different enhancement transformation sets, is the weight, indicating the importance of the th transformation, is the number of transformations of each type of enhancement transformation, and the denominator represents the total number of all possible non-professional expression patterns. is the number of transformations of each type of enhancement transformation.

[0103] The closer R is to 1, the better the generated data set can simulate the diversity of non-professional expressions; the closer R is to 0, the more single the data set is and cannot represent the diverse expressions of non-professional users.

[0104] Set the coverage threshold If , the data augmentation ends. Otherwise, continue the data augmentation process.

[0105] S5: Train the LLM model using the sub-datasets D1, D2 and the data-augmented sub-datasets D3, D4 to obtain a fine-tuned model F for multi-cloud storage heterogeneous user intent recognition.

[0106] As a refinement of the above embodiment, the LLM model training loss function is as follows:

[0107] ,

[0108] wherein, is the standard cross-entropy loss of the professional user dataset (D1 and D2), is the standard cross-entropy loss of the non-professional user dataset (D3 and D4) that handles non-standard expressions; is the aligned semantic loss of the professional user dataset (D1 and D2), is the aligned semantic loss of the non-professional user dataset (D3 and D4), which is used to constrain the consistency of the output semantic embedding and the target semantic embedding; is the confidence regulation loss, which regulates the prediction confidence of non-professional users to avoid the model's false amplification of low-confidence samples; , is the dynamic weight, , and represent the weight balancing parameters, and are used to adjust the influence of the professional user and non-professional user semantic alignment loss, respectively, represents the weight of the non-professional user confidence regulation.

[0109] , are all greater than or equal to 0. There is no hard normalization requirement, but it can be set according to the proportion of the number of professional user datasets (D1 and D2) and non-professional user datasets (D3 and D4).

[0110]

[0111]

[0112]

[0113]

[0114] , , and are the proportions of the middle samples of the data sets D1, D2, D3 and D4 respectively

[0115] greater than or equal to 0, no hard normalization requirement, but usually the value range is .

[0116] The standard cross-entropy loss of the professional user data set is as follows:

[0117]

[0118] The standard cross-entropy loss of the non-professional user data set is as follows:

[0119]

[0120] wherein M represents the number of samples in the small batch data for calculating the classification loss, represents the number of elements in the basic task set, denotes the probability that the jth sample belongs to the kth class.

[0121] The aligned semantic loss of the professional user data set is as follows:

[0122] ,

[0123] wherein represents the number of samples in the small batch data for calculating the classification loss of the professional user data set, is the semantic embedding of the professional user sample, corresponds to the target semantic embedding of the professional user sample;

[0124] The aligned semantic loss of the non-professional user data set is as follows:

[0125] ,

[0126] wherein represents the number of samples in the small batch data for calculating the classification loss of the non-professional user data set, is the semantic embedding of the professional user sample, corresponds to the target semantic embedding of the professional user sample;

[0127] The confidence regulation loss is as follows:

[0128] ,

[0129] wherein is a confidence threshold, usually 0.9, The highest class confidence in the jth sample corresponding to the model. This mechanism is used to ensure that the data set of non-professional users (D3 and D4) will focus on improving the reliability of the prediction, rather than blindly increasing the recognition accuracy.

[0130] Specifically, the dynamic weight calculation method is as follows:

[0131] ,

[0132] Among them, is the current training round, is the total training round, is the initial weight of the professional user, is the initial weight of the non-professional user, and to ensure the training optimization of the non-professional user data set.

[0133] In the early training ( ), is small, is large, and the fine-tuning of the model pays more attention to the understanding and optimization of non-professional expressions In the later training ( ), and gradually approach each other, reach a balance, achieve balanced optimization of professional users and non-professional users, and enhance the generalization ability of the model.

[0134] Finally, the fine-tuned model F is obtained, and the following is obtained:

[0135] Among them,

[0136] The embodiments of the disclosure also provide a multi-cloud storage heterogeneous user intent recognition method based on LLM. The device comprises a processor and a memory. Optionally, the device can further comprise a communication interface and a bus. The processor, the communication interface and the memory can communicate with each other through the bus. The communication interface can be used for information transmission. The processor can call the logic instructions in the memory to execute the multi-cloud storage heterogeneous user intent recognition method based on LLM of the above-mentioned embodiments.

[0137] In addition, the logic instructions in the memory described above can be implemented in the form of a software functional unit and sold or used as an independent product. It can be stored in a computer readable storage medium.

[0138] The memory, as a computer readable storage medium, can be used to store software programs, computer executable programs, such as program instructions / modules corresponding to the method in the embodiments of the present disclosure. The processor executes the function application and data processing by running the program instructions / modules stored in the memory, that is, implements the method of the multi-cloud storage heterogeneous user intention recognition method based on LLM in the above embodiments.

[0139] The memory can include a program storage area and a data storage area, wherein the program storage area can store an operating system and application programs required by at least one function; the data storage area can store data created according to the use of the terminal device, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory.

[0140] Finally, it should be noted that: the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments or make equivalent replacements to some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying heterogeneous user intent in multi-cloud storage based on LLM, characterized in that, Includes the following steps: Collect user demand identification datasets through multi-cloud platforms and set a unified data management format, including general fields, interaction information fields and user background fields; Build a professional keyword dictionary, compare user queries with the professional keyword dictionary, calculate the proportion of professional words, and if it exceeds a set threshold, mark the user's background field as a professional user; otherwise, mark them as a non-professional user. Based on the interaction information field and user background field in the data management format, the dataset is divided into four subsets: single intent of professional users as subset D1, composite intent of professional users as subset D2, single intent of non-professional users as subset D3, and composite intent of non-professional users as subset D4; the data in each subset includes the user's natural language intent expression, storage mode, and basic task. Data augmentation was performed on the single intent subset D3 and the composite intent subset D4 of non-professional users until the dataset coverage reached a set threshold. The LLM model was trained using subsets D1, D2, and augmented subsets D3 and D4 to obtain a fine-tuned model F, which was used for heterogeneous user intent recognition in multi-cloud storage. The coverage calculation formula is as follows: ， in, Express the number of types of enhancement transformations, For dataset The number of unique patterns in the dataset describes the actual diversity of the dataset. The number of data points corresponding to different enhancement transformation sets, Let be the weight, representing the th The importance of class transformation This indicates the number of transformations for each type of enhancement transformation.

2. The method for identifying heterogeneous user intent in multi-cloud storage based on LLM according to claim 1, characterized in that, The general fields include timestamp, platform, user ID, session ID, and request ID; the interaction information fields include interaction type, query content, number of intents, and intent recognition; the intent recognition classification format is storage mode-basic task.

3. The method for identifying heterogeneous user intent in multi-cloud storage based on LLM according to claim 2, characterized in that, The storage modes include block storage, object storage, and file storage, and the basic tasks include single basic tasks and multiple basic tasks.

4. The method for identifying heterogeneous user intent in multi-cloud storage based on LLM according to claim 1, characterized in that, The data augmentation includes: grammatical structure transformation, lexical transformation, introduction of textual errors, and mixed use of uppercase and lowercase letters; The grammatical structure transformation includes adjusting the order of the subject, predicate, and object; The lexical transformations include synonym replacement and word redundancy generation of non-professional expressions; The introduced textual errors include errors that violate Chinese or English grammar rules in written expression.

5. The method for identifying heterogeneous user intent in multi-cloud storage based on LLM according to claim 4, characterized in that, LLM model training loss function as follows: ， in, For professional user datasets, the standard cross-entropy loss is used. For non-professional user datasets, the standard cross-entropy loss is used. For alignment semantic loss of professional user datasets, For alignment semantic loss of non-professional user datasets, To adjust for confidence level loss, , For dynamic weights, , and This represents the weight balancing parameter. and These are used to modulate the impact of semantic alignment loss on professional and non-professional users, respectively. This indicates the weighting of confidence levels for non-professional users.

6. The method for identifying heterogeneous user intent in multi-cloud storage based on LLM according to claim 5, characterized in that, Weighting balance parameters The calculation method is as follows: ， ， Weighting balance parameters The calculation method is as follows: ， ， in, , , , These represent the proportions of samples in datasets D1, D2, D3, and D4, respectively.

7. The method for identifying heterogeneous user intent in multi-cloud storage based on LLM according to claim 5, characterized in that, The standard cross-entropy loss for the professional user dataset is as follows: The standard cross-entropy loss for the non-professional user dataset is as follows: in, This represents the number of samples in the mini-batch used to calculate the classification loss in the professional user dataset. This represents the true probability value of the j-th sample in class k. Let represent the probability that the j-th sample in the professional user dataset belongs to the k-th class. This represents the number of samples in the mini-batch used to calculate the classification loss in a non-professional user dataset. Let represent the probability that the j-th sample in the non-professional user dataset belongs to the k-th class. Indicates the number of elements in the basic task set; The alignment semantic loss of the professional user dataset is as follows: ， in, This represents the number of samples in the mini-batch used to calculate the classification loss in the professional user dataset. Semantic embedding for professional user samples Target semantic embedding corresponding to professional user samples; The alignment semantic loss for the non-professional user dataset is as follows: ， in This represents the number of samples in the mini-batch used to calculate the classification loss in a non-professional user dataset. Semantic embedding for professional user samples Target semantic embedding corresponding to professional user samples; The confidence level adjustment loss is as follows: ， in, The confidence threshold is... The highest class confidence score in the predictions for the j-th sample of the model pair.

8. The method for identifying heterogeneous user intent in multi-cloud storage based on LLM according to claim 7, characterized in that, The dynamic weight calculation method is as follows: ， in, For the current training round, For the total number of training rounds, Initial weight for professional users, The initial weight is for non-professional users, and .

9. A dataset construction apparatus for a multi-cloud storage heterogeneous user intent recognition method based on LLM, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute the LLM-based heterogeneous user intent recognition method for multi-cloud storage as described in any one of claims 1-8 when running the program instructions.

Images