A multi-client additional writing processing method and system

Abstract

The present invention discloses a multi-client append write processing method and system, including: first obtaining concurrent append write operations of multiple clients on a target file, generating and processing a global append write lock request, and persisting the data in a direct I / O write mode after determining the file offset. After completion, the metadata is updated, unlocked, and the result is returned. When the client receives the result, the file classification model is called to identify the target file metadata type. If it is consistent with the original type, the storage location is maintained. If it is inconsistent, it is transferred to the corresponding location for storage, thereby ensuring the efficiency and accuracy of data processing.

Description

Technical Field

[0001] The present invention relates to the field of file processing technology, and in particular to a multi-client additional writing processing method and system. Background Art

[0002] In distributed file systems, there is an increasing demand for multiple clients to perform additional write operations on the same file. Traditional methods for processing additional writes have problems such as imperfect file lock management, low cache mode efficiency, and complex logic, which makes it difficult to ensure data consistency and poor performance. Summary of the invention

[0003] The object of the present invention is to provide a multi-client additional writing processing method and system.

[0004] In a first aspect, an embodiment of the present invention provides a multi-client append write processing method, comprising:

[0005] Get concurrent append write operations of multiple clients on the target file;

[0006] Generate a global append write lock request according to the target append write operation, and call the metadata server corresponding to the target file to lock the target file; the target append write operation is a concurrent append write operation of the target client on the target file among the concurrent append write operations of the multiple clients on the target file, and the target client is any client among the multiple clients;

[0007] When the global append write lock is locked successfully, the file size of the target file is obtained from the metadata server, and the file offset is determined according to the file size;

[0008] Performing the target append write operation on the target file in a direct I / O write mode according to the file offset, and persisting the write data corresponding to the target append write operation to the data server;

[0009] In the case where data persistence is complete, the metadata corresponding to the target file in the metadata server is updated, and the metadata server initiates an unlock request to release the global append write lock, and returns the append write result corresponding to the target append write operation to the target client;

[0010] In the case that the plurality of clients all receive corresponding append writing results, calling a pre-trained file classification model to perform type identification on the metadata of the target file;

[0011] When the type identification result of the metadata of the target file is consistent with the original file type of the metadata of the target file, maintaining the storage location of the target file;

[0012] When the type identification result of the metadata of the target file is inconsistent with the original file type of the metadata of the target file, the target file is transferred to a storage location corresponding to the type identification result for storage.

[0013] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is used to execute the method described in the first aspect.

[0014] Compared with the prior art, the beneficial effects provided by the present invention include: adopting a multi-client append write processing method and system disclosed in the present invention, by obtaining concurrent append write operations of multiple clients on the target file, generating and processing a global append write lock request, and persisting the data in a direct I / O write mode after determining the file offset. After completion, the metadata is updated, unlocked, and the result is returned. When the client receives the result, the file classification model is called to identify the target file metadata type. If it is consistent with the original type, the storage location is maintained. If it is inconsistent, it is transferred to the corresponding location for storage, thereby ensuring the efficiency and accuracy of data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0016] Figure 1 A schematic diagram of the steps of the multi-client append writing processing method provided by an embodiment of the present invention;

[0017] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0019] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0020] In order to solve the technical problems in the aforementioned background technology, Figure 1This is a flow chart of a multi-client append write processing method provided in an embodiment of the present disclosure. The multi-client append write processing method is introduced in detail below.

[0021] Step S201, obtaining concurrent append write operations of multiple clients on a target file;

[0022] Step S202, generating a global append write lock request according to a target append write operation, calling a metadata server corresponding to the target file to lock the target file; the target append write operation is a concurrent append write operation of a target client on the target file among concurrent append write operations of the multiple clients on the target file, and the target client is any client among the multiple clients;

[0023] Step S203, when the global append write lock is locked successfully, obtaining the file size of the target file from the metadata server, and determining the file offset according to the file size;

[0024] Step S204, performing the target append write operation on the target file in a direct I / O write mode according to the file offset, and persisting the write data corresponding to the target append write operation to the data server;

[0025] Step S205: When data persistence is complete, the metadata corresponding to the target file in the metadata server is updated, and the metadata server initiates an unlock request to release the global append write lock, and returns the append write result corresponding to the target append write operation to the target client;

[0026] Step S206, when the plurality of clients have all received corresponding append writing results, calling a pre-trained file classification model to perform type identification on the metadata of the target file;

[0027] Step S207, when the type identification result of the metadata of the target file is consistent with the original file type of the metadata of the target file, maintaining the storage location of the target file;

[0028] Step S208: When the type identification result of the metadata of the target file is inconsistent with the original file type of the metadata of the target file, the target file is transferred to a storage location corresponding to the type identification result for storage.

[0029] In the embodiment of the present invention, for example, it is assumed that there is a distributed file system, and the server is responsible for processing requests from multiple clients. At a certain moment, the server simultaneously receives concurrent append write operations from clients A, B, and C for the same target file "document.txt". Client A wants to add a text "The weather is really good today" at the end of the file, client B is preparing to add "suitable for going out for a walk", and client C plans to add "but the wind is a bit strong". After the server receives these concurrent append write operations, take the operation of client A (target append write operation) as an example. The server generates a global append write lock request and calls the metadata server corresponding to "document.txt". After receiving the lock request, the metadata server checks whether there are other clients currently holding the lock of the file. If not, "document.txt" is successfully locked for client A, and the server is notified of the successful lock. After the server learns that the lock is successful, it immediately obtains the current file size of "document.txt" from the metadata server, assuming that the current file size is 500 bytes. Then, the file offset is calculated based on this size to determine that the starting position of the client A's append write should be after 500 bytes. According to the calculated offset, the server uses direct I / O write to persist the text "The weather is great today" that client A wants to add to the corresponding location in the data server. The direct I / O write method ensures that the data is written directly to the data server without passing through the server's cache, which improves the efficiency and reliability of data writing. After the written data of client A is successfully persisted to the data server, the server updates the metadata of "document.txt" in the metadata server, such as updating the file size from 500 bytes to 520 bytes (assuming that the added text occupies 20 bytes), and at the same time modifies the last modification time and other related metadata information. Then, the metadata server initiates an unlock request to release the global append write lock set for client A. Finally, the server returns the result of the successful append write to client A. Assume that clients A, B, and C have successfully received their respective append write results. At this time, the server calls the pre-trained file classification model to identify the type of the metadata of "document.txt". The file classification model analyzes the file's format, content characteristics, and other metadata information to determine the file type. If the identification result of the file classification model shows that "document.txt" still belongs to its original text file type and has not changed, the server will maintain "document.txt" in the current storage location and will not perform any migration operations. Suppose that after identification by the file classification model, it is found that "document.txt" has changed its content and format due to multiple appends and is no longer a simple text file, but more in line with a specific binary file type.The server will transfer "document.txt" to the storage location corresponding to the binary file type to ensure that the storage and management of files are more reasonable and efficient.

[0030] As a more complex example, suppose the target file is a large database file "data.db" and multiple clients are performing append write operations on it at the same time. Client D wants to add some new user records, client E wants to update some data fields, and client F wants to insert new index information. The server receives these different types of append write requests at the same time. The server generates a global append write lock request for client D's operation and successfully locks "data.db" on the metadata server. The current size of "data.db" is obtained to be 10GB, and the offset of client D's append write is calculated. The new user record of client D is accurately written to the specified location according to the direct I / O write method. The append write operations of client E and client F are processed in turn. When all clients complete the append write and receive the results, the file classification model can find that due to the large number of different types of data appends, the type of "data.db" has changed from a normal database file to an optimized database file with a specific structure. The server migrates "data.db" to a storage location specially set for this optimized database file to improve subsequent access and processing efficiency.

[0031] Through the above detailed scenario examples, we can more clearly understand the workflow and effects of the multi-client append write processing method in actual applications, ensuring data consistency, reliability and efficient processing.

[0032] In the embodiment of the present invention, the file classification model is obtained in the following manner.

[0033] Acquire a first file data instance corresponding to a first file format, wherein the first file data instance has a first file type target value, and the first file type target value is used to indicate an actual file type of the first file data instance;

[0034] Performing a feature extraction operation on the first file data instance through a first feature extraction network to obtain a first vector representation, wherein the first feature extraction network is used to convert the file data corresponding to the first file format to a target feature domain, wherein the target feature domain is a feature domain where the vector representation corresponding to the file data of the second file format is located, wherein the first file format and the second file format are different file formats, and the complexity of the file data corresponding to the first file format is less than the complexity of the file data corresponding to the second file format;

[0035] Based on the first vector representation, classify the file using a basic file classification model to obtain an inferred file type;

[0036] Based on the deviation between the inferred file type and the first file type target value, the model network parameters of the basic file classification model are optimized to obtain a file classification model, so that the file classification model is used to classify the type of file data corresponding to the second file format.

[0037] In an embodiment of the present invention, illustratively, it is assumed that the server is processing the training of a file classification model. First, the server obtains a large amount of data in a first file format, such as a simple file data instance in a plain text format (.txt). These file data instances contain various topics and contents, such as an article about tourism, a technical description document, or a fragment of a novel. Each such first file data instance is marked with a clear first file type target value, which accurately indicates the actual file type of the file data instance. For example, a travel article is marked as "tourism text", a technical description document is marked as "technical text", and a novel fragment is marked as "literary text". The server uses a specially designed first feature extraction network to process these first file data instances. This feature extraction network can convert relatively simple data in the first file format (such as .txt text) into a specific target feature domain. Assume that the second file format is a complex multimedia format, such as a composite file format containing audio, video, and images. The target feature domain is the feature space in which the file data in this complex second file format is located when performing feature representation. For example, for a travel article in .txt format, the first feature extraction network will analyze the features of the article, such as vocabulary, sentence structure, paragraph distribution, etc., and convert these features into a vector representation in the target feature domain. Although this vector representation comes from a simple text file, it has a certain correlation and comparability with the feature representation of a complex multimedia file in the feature space. The server then uses the basic file classification model to classify the first vector representation. This basic file classification model can be a neural network model based on deep learning, which has undergone preliminary training but is not yet perfect. Taking the vector representation of the travel article extracted previously as an example, the basic file classification model will analyze and calculate this vector and try to predict the file type to which it belongs. It may come up with inferred file types such as "travel text" and "life text". The server will compare the inferred file type obtained by the basic file classification model with the first file type target value initially set. If the inferred type is inconsistent with the target value, the deviation between the two will be calculated. For example, for the above travel article, if the basic file classification model infers it as "life text" and the target value is "travel text", the server will adjust the network parameters of the basic file classification model according to this deviation. By continuously inputting a large number of first file data instances and their corresponding target values, the model parameters are continuously optimized, making the model's inference results more and more accurate. After multiple optimizations and adjustments, the basic file classification model gradually evolves into a file classification model that can accurately classify. This final file classification model can be used to process data in complex second file formats (such as multimedia files) and accurately classify them.To give a more specific example, assume that the first file format is a simple CSV (comma separated value) file, which contains information such as the product name, price and description. The second file format is a complex e-commerce product page file containing product pictures, video introductions and related comments. The server first obtains a large number of CSV files as first file data instances and marks their actual file types, such as "electronic products", "clothing", "home furnishings", etc. Then, the data in the CSV file, such as the vocabulary features of the product name, the price range features, the text features of the description, etc., are converted to the same target feature domain as the complex e-commerce product page file through the first feature extraction network to obtain the first vector representation. Next, the basic file classification model classifies these vectors, which may result in inference errors, such as misjudging "electronic products" as "home furnishings". The server optimizes the parameters of the model based on this deviation, and after repeated training and optimization, finally obtains a file classification model that can accurately classify. This model can accurately judge the type of complex e-commerce product page files, such as determining that they belong to categories such as "hot electronic products", "fashionable clothing", and "high-end home furnishings". Through such detailed scenario examples, you can see how the server gradually acquires and optimizes the file classification model to improve the classification ability of different file formats. Let's continue to give you an example. Suppose the server is processing a file classification task about academic research papers. The first file format is a paper abstract in plain text format, while the second file format is a complete academic paper file containing the full text of the paper, charts, references, etc. The server obtains a large number of paper abstracts as first file data instances. Each abstract is marked with the subject field to which it belongs, such as "physics", "biology", "computer science", etc. This is the first file type target value. The first feature extraction network extracts features from these paper abstracts, such as extracting keywords, counting the frequency of occurrence of specific field terms, analyzing the structure and semantics of sentences, etc., and converts these features into the same target feature domain as the complete academic paper file to form a first vector representation. The basic file classification model may have deviations when classifying these first vector representations, such as misclassifying a paper abstract that originally belongs to the field of "computer science" as the field of "mathematics". Based on this deviation, the server optimizes the basic file classification model by adjusting the weights, biases and other parameters of the neural network. After many iterations, the model continues to learn and improve until it can accurately infer the correct subject area based on the vector representation of the paper abstract. When the model is trained to maturity, it can be applied to classify complete academic paper files (second file format), determine the specific subject area to which they belong, and provide effective support for the management and retrieval of academic resources. For example, in the scenario of image classification, the first file format can be a simple grayscale image, while the second file format is a color image containing multiple complex elements.The server obtains a large number of grayscale images as first file data instances and marks their categories, such as "landscape", "person", "animal", etc. The first feature extraction network extracts the pixel value distribution, edge features, etc. of the grayscale image and converts it to the same target feature domain as the complex color image. The basic file classification model classifies these converted vectors, which may result in misclassification. The server optimizes the model parameters according to the deviation, and finally obtains a file classification model that can accurately classify complex color images. In the example of audio file classification, the first file format can be a monophonic, low-sampling audio clip, and the second file format is a multi-channel, high-sampling complex audio file. The server obtains many monophonic, low-sampling audio clips as first file data instances, marked as "music", "speech", "ambient sound", etc. The first feature extraction network extracts the frequency features, duration features, etc. of the audio, and converts them to the same target feature domain as the complex audio file. The basic file classification model classifies these vectors, and then the server optimizes the model according to the deviation, so that the model can accurately classify complex multi-channel high-sampling audio files. Through these rich and diverse scenario examples, we can have a deeper understanding of the specific operations and roles of the server in the process of obtaining and optimizing the file classification model, as well as the application potential of the model in different types of file classification tasks. Suppose we are dealing with file classification in the medical field. The first file format can be a simple medical record text description, which only contains the patient's basic symptoms and preliminary diagnosis, while the second file format is a complete medical file containing detailed examination reports, imaging materials, treatment records, etc. The server obtains a large number of medical record text descriptions as first file data instances, and each instance is marked with a specific disease category, such as "cold", "pneumonia", "diabetes", etc. This is the first file type target value. The first feature extraction network extracts features from these medical record texts, such as extracting disease keywords, patient age characteristics, keyword frequency of symptom descriptions, etc., and converts these features into the same target feature domain as the complete medical file to form a first vector representation.

[0038] The basic file classification model may have deviations when classifying these vectors, such as misclassifying "pneumonia" as "bronchitis". Based on this deviation, the server optimizes the basic file classification model by adjusting the model's neuron connection weights, thresholds and other parameters. After multiple training and optimizations, the model can accurately determine the disease category based on the vector representation of the medical record text. When the model is mature, it can be used to classify complete medical records (second file format) containing rich information, providing strong support for the management and analysis of medical data. Let's look at an example in the financial field. The first file format can be a simple list of transaction records, which only contains basic information such as transaction amount, time and transaction object, while the second file format is a comprehensive financial file containing detailed financial statements, risk assessment reports, market analysis, etc. The server obtains a large number of transaction record lists as first file data instances and marks the financial business type to which they belong, such as "stock trading", "bond investment", "insurance business", etc. The first feature extraction network extracts the amount distribution characteristics, transaction time rules, transaction object categories, etc. in the transaction records, and converts them to the same target feature domain as the comprehensive financial file. The basic file classification model may classify these vectors inaccurately. The server optimizes the model parameters according to the deviation, and finally obtains a file classification model that can accurately classify comprehensive financial files, which provides assistance for financial data analysis and risk management. In the field of education, the first file format can be a simple list of students' grades, which only contains grades in various subjects, while the second file format is a comprehensive learning archive containing students' homework, classroom performance, teacher comments, etc. The server obtains a large number of student grades as first file data instances, marked as different subject categories or learning levels, such as "excellent in mathematics" and "medium in Chinese". The first feature extraction network extracts the distribution characteristics of grades, high-scoring and low-scoring subjects, etc., and converts them to the same target feature domain as the comprehensive learning archive. The basic file classification model is classified, and then the server optimizes the model according to the deviation, so that the model can accurately classify the comprehensive learning archive, providing a basis for education evaluation and personalized teaching. Through the above detailed scenario examples of many different fields and types, the comprehensiveness and versatility of the server in the process of obtaining and optimizing file classification models, as well as the important value of the model in various practical applications, are fully demonstrated.

[0039] In the embodiments of the present invention, the following implementation modes are also provided.

[0040] Acquire first inferred file data corresponding to the second file format;

[0041] performing a feature conversion operation on the first inferred file data through a second feature extraction network to obtain a second vector representation, wherein the second feature extraction network is used to convert the file data corresponding to the second file format into the target feature domain;

[0042] Based on the second vector representation, classification is performed using the file classification model to obtain the type of the first inferred file data.

[0043] In the embodiment of the present invention, for example, it is assumed that the system where the server is located processes the classification task of image files, the first file format is a low-resolution black and white image, and the second file format is a high-resolution color image. First, the server obtains the first inference file data corresponding to the second file format. For example, a high-resolution color landscape picture is obtained, which is the first inference file data that needs to be classified. Then, the server performs a feature conversion operation on the landscape picture through a second feature extraction network. This second feature extraction network is specifically used to process data in the second file format of high-resolution color images. It may analyze complex features such as color distribution, texture features, object shape, etc. in the image, and convert these features into a second vector representation in the target feature domain. For example, the color ratio of the blue sky and green grass in the image, the features of the mountain outline, the regularity of the tree texture, etc. are extracted, and this information is converted into a specific vector. Next, based on the obtained second vector representation, the server classifies it through a trained file classification model. This file classification model has learned the feature patterns of different types of images in previous training. For the second vector representation of this landscape picture, the model will calculate and analyze it and finally determine its type. Assume that the model classifies this picture as a "natural scenery" image. For another example, when processing the classification of audio files, the first file format is a monophonic, low-sampling audio clip, and the second file format is a multi-channel, high-sampling complex audio file. The server obtains a multi-channel, high-sampling complex music audio as the first inferred file data corresponding to the second file format. Then, it is processed by a second feature extraction network designed specifically for this complex audio. It may extract features such as fundamental frequency, harmonics, rhythm changes, and sound intensity dynamics in the audio, and convert them into vector representations in the target feature domain. Then, based on this second vector representation, it is classified by the file classification model. The model may classify this audio as a "classical music" type. Another example is in the document classification scenario, where the first file format is a simple plain text format, and the second file format is a complex document containing multiple fonts, formats, and charts. The server obtains a complex financial report document containing rich formats and charts as the first inferred file data corresponding to the second file format. The second feature extraction network extracts features such as text content features, chart types and contents, and format layout in the document, and converts them into vectors in the target feature domain. Finally, the file classification model classifies the document as a "financial report" document based on this vector representation. In the video classification scenario, assume that the first file format is a low-frame-rate, low-definition video clip, and the second file format is a high-frame-rate, high-definition complete video. The server obtains a high-frame-rate, high-definition movie video as the first inferred file data corresponding to the second file format.Through a special second feature extraction network, the picture change frequency, scene switching mode, character action features, audio features, etc. in the video are extracted and converted into a vector representation of the target feature domain. Based on this vector, the file classification model classifies the movie as an "action movie" type. In the task of processing web page classification, the first file format can be a simple web page text content, and the second file format is a complete web page containing various scripts and multimedia elements. The server obtains a complex e-commerce web page containing multiple scripts and multimedia elements as the first inferred file data corresponding to the second file format. The second feature extraction network extracts features such as link structure, page layout, multimedia element type and distribution, and text keywords in the web page, and converts them into vectors of the target feature domain. Based on this vector, the file classification model classifies the web page as an "e-commerce shopping" web page. Through the examples of various specific scenarios above, the server can effectively use the second feature extraction network and the file classification model to accurately classify different types of second file format data to meet the classification needs in various practical applications.

[0044] In an embodiment of the present invention, the first feature extraction network includes a first file format feature extraction network and a second file format feature extraction network, the first file format feature extraction network is used to convert file data corresponding to the first file format into the target feature domain, and the second file format feature extraction network is used to convert file data corresponding to the second file format into the target feature domain;

[0045] The performing of a feature extraction operation on the first file data instance by using a first feature extraction network to obtain a first vector representation may be implemented through the following example.

[0046] Performing a feature extraction operation on the first file data instance through the first file format feature extraction network to obtain a first vector representation;

[0047] The embodiments of the present invention also provide the following implementation modes.

[0048] Acquire first inferred file data corresponding to the second file format;

[0049] Performing a feature extraction operation on the first inferred file data through the second file format feature extraction network to obtain a second vector representation;

[0050] Based on the second vector representation, classification is performed using the file classification model to obtain the type of the first inferred file data.

[0051] In an embodiment of the present invention, illustratively, it is assumed that in a multimedia data processing environment, the first file format is a low-definition black-and-white photo, and the second file format is a high-definition color photo. The server first obtains a first file data instance corresponding to the first file format, such as a low-definition black-and-white photo of a person. The first file format feature extraction network performs a feature extraction operation on the photo. This network is specifically used to process low-definition black-and-white photos. It may extract features such as the outline of the person and the gray value distribution in the photo, and convert these features to the target feature domain, thereby obtaining a first vector representation. Next, the server obtains the first inferred file data corresponding to the second file format, such as a high-definition color landscape photo. Then, the second file format feature extraction network performs a feature extraction operation on the landscape photo. The network processes the high-definition color photo, and may extract the color distribution, detailed features of the scene, depth of field information, etc. in the photo, and convert these features to the target feature domain to obtain a second vector representation. Based on the obtained second vector representation, the server performs classification through a trained file classification model. Assuming that the file classification model has learned the feature patterns of different types of photos, for the second vector representation of this landscape photo, the model, after calculation and analysis, finally determines that its type is a "natural scenery" photo. For another example, in the audio processing scenario, the first file format is a monophonic, low-sampling-rate speech clip, and the second file format is a multi-channel, high-sampling-rate music audio. The server obtains a monophonic, low-sampling-rate speech clip as the first file data instance corresponding to the first file format. Through the first file format feature extraction network designed specifically for this format, the intonation, speech speed, basic audio features, etc. in the speech are extracted and converted to the target feature domain to obtain the first vector representation. Subsequently, the server obtains a multi-channel, high-sampling-rate music audio as the first inferred file data corresponding to the second file format. Using the second file format feature extraction network, the features such as the type of instrument, melody pattern, and harmonic structure in the audio are extracted and converted to the second vector representation of the target feature domain. Based on this second vector representation, classification is performed through the file classification model. The model may classify this music audio as "classical music". In the field of document processing, the first file format can be a simple report in plain text, and the second file format can be a professional paper containing charts and complex formats. The server obtains a simple market research report in plain text as the first file data instance corresponding to the first file format. Using the first file format feature extraction network, the features such as keywords, sentence structure, and paragraph theme in the text are extracted and converted to the target feature domain to obtain the first vector representation. Then, the server obtains an academic paper containing a large number of charts and complex formats as the first inferred file data corresponding to the second file format.Through the second file format feature extraction network, the features such as the type and content of the charts, the citation format, and the distribution of professional terms in the paper are extracted and converted into the second vector representation of the target feature domain. Based on this second vector representation, the file classification model classifies this paper as a "scientific research" document. In the case of video processing, it is assumed that the first file format is a short video with a low frame rate and low resolution, and the second file format is a movie clip with a high frame rate and high resolution. The server obtains a short video with a low frame rate and low resolution as the first file data instance corresponding to the first file format. Through the corresponding first file format feature extraction network, the main action, simple scene features, etc. in the video are extracted and converted to the target feature domain to obtain the first vector representation. Then, the server obtains a movie clip with a high frame rate and high resolution as the first inference file data corresponding to the second file format. With the help of the second file format feature extraction network, the features such as the picture details, special effects application, and the coordination of audio and picture in the clip are extracted and converted into the second vector representation of the target feature domain. Based on the second vector representation, the file classification model classifies this movie clip as an "action movie". In the scenario of web page processing, the first file format can be a simple web page with pure text, and the second file format is a web page containing multimedia elements and complex layout. The server obtains a simple news web page with pure text as the first file data instance corresponding to the first file format. Through the first file format feature extraction network, the text content theme, keyword frequency and other features in the web page are extracted, and converted to the target feature domain to obtain the first vector representation. Next, the server obtains an e-commerce web page containing a large number of pictures, videos and dynamic effects as the first inferred file data corresponding to the second file format. Using the second file format feature extraction network, the multimedia element type and distribution, page interaction mode, commodity classification information and other features in the web page are extracted, and converted into the second vector representation of the target feature domain. Based on this second vector representation, the file classification model classifies this e-commerce web page as a "shopping consumption" web page. Through the above rich and diverse scenario examples, the server can effectively use different feature extraction networks to process file data of different formats, and accurately classify through the file classification model to meet the needs of various application scenarios.

[0052] In the embodiments of the present invention, the following implementation modes are also provided.

[0053] Acquire at least two file data instance data groups, wherein the file data instance data groups include a first sub-file data instance corresponding to the first file format and a second sub-file data instance corresponding to the second file format, wherein the first sub-file data instance and the second sub-file data instance in the same file data instance data group represent the same file information;

[0054] For a target file data instance data group in the at least two file data instance data groups, performing a feature extraction operation on a first sub-file data instance in the target file data instance data group by using a basic first file format feature extraction network to obtain a first sub-vector representation;

[0055] Performing a feature extraction operation on the second sub-file data instance in the target file data instance data group through a basic second file format feature extraction network to obtain a second sub-vector representation;

[0056] Determine the at least two file data instance data groups as the target file data instance data groups respectively, and obtain at least two first sub-vector representations and at least two second sub-vector representations;

[0057] Following a training strategy of enhancing the similarity of a target sub-vector representation array and reducing the similarity of a non-target sub-vector representation array, the model network parameters of the basic first file format feature extraction network and the model network parameters of the basic second file format feature extraction network are optimized to obtain the first file format feature extraction network and the second file format feature extraction network, wherein the first sub-vector representation and the second sub-vector representation included in the target sub-vector representation array are obtained based on the same file data instance data group, and the first sub-vector representation and the second sub-vector representation included in the non-target sub-vector representation array are not obtained based on the same file data instance data group.

[0058] In the embodiment of the present invention, illustratively, assuming that it is in an environment of image and text processing, the first file format is a simple text description, and the second file format is a corresponding image. The server first obtains multiple file data instance data groups. For example, in a set of file data instance data groups, the first sub-file data instance is a simple text description of a picture of a cat, such as "a black and white cat with blue eyes, squatting on the ground", and the second sub-file data instance is the actual image of the cat. For the target file data instance data group in these file data instance data groups, the server performs a feature extraction operation on the first sub-file data instance in the group (i.e., the above-mentioned text description) through the basic first file format feature extraction network. This network may extract key information in the text, such as "cat", "black and white", "blue eyes", etc., and convert it into a first sub-vector representation. At the same time, the server performs a feature extraction operation on the second sub-file data instance in the target file data instance data group (i.e., the image of the cat) through the basic second file format feature extraction network. The network may extract features such as color, shape, texture, etc. in the image and convert them into a second sub-vector representation. Then, the server determines the multiple file data instance data groups as target file data instance data groups, thereby obtaining multiple first sub-vector representations and multiple second sub-vector representations. During the training process, the server follows the strategy of enhancing the similarity of the target sub-vector representation array and reducing the similarity of the non-target sub-vector representation array. For the first sub-vector representation and the second sub-vector representation (i.e., the target sub-vector representation array) from the same file data instance data group, the server strives to make them closer in the feature space, indicating that they represent the same information. For example, the vector representations of the text description and the image of the above cat will be more similar after training. For the first sub-vector representation and the second sub-vector representation (i.e., the non-target sub-vector representation array) that are not from the same file data instance data group, the server tries to keep them away in the feature space to distinguish different information. For example, the vector of the text description of the dog in another file data instance data group and the vector of the image of the above cat will be quite different after training. By continuously optimizing the model network parameters of the basic first file format feature extraction network and the basic second file format feature extraction network, the server finally obtains the first file format feature extraction network and the second file format feature extraction network that can accurately convert different format files of the same information into similar vector representations. For another example, in the scenario of audio and music score processing, the first file format is an audio file of a piece of music, and the second file format is a corresponding music score file. The server obtains multiple file data instance data groups, one of which may include an audio of a piece of classical music and its corresponding music score. For the target file data instance data group, the server performs feature extraction on the audio file through the basic first file format feature extraction network, and can extract the frequency, rhythm, pitch and other features of the audio and convert them into the first sub-vector representation.The feature extraction of the score file is performed through the basic second file format feature extraction network, and the features such as notes, beats, and modes are extracted and converted into the second sub-vector representation. The server processes multiple such file data instance data groups to obtain multiple first sub-vector representations and second sub-vector representations. During training, for the vector representations of audio and music scores from the same group (target sub-vector representation array), the server enhances their similarity and makes them close in the feature space. For the vector representations of audio and music scores from different groups (non-target sub-vector representation array), the server reduces their similarity and makes them far away in the feature space. After repeated optimization, the server obtains a feature extraction network that can effectively convert audio and music scores into similar or different vector representations. In the scenario of web page and related document processing, the first file format is the HTML code of the web page, and the second file format is a detailed text description document of the content of the web page. The server obtains multiple file data instance data groups, such as the HTML code of an e-commerce web page and the text description document of its product introduction. For the target file data instance data group, the server extracts layout, links, element attributes and other features of the HTML code through the basic first file format feature extraction network and converts them into the first sub-vector representation. The features such as keywords, paragraph structure, and semantic information extracted from the text description document are converted into a second sub-vector representation through the basic second file format feature extraction network. The server processes multiple such groups to obtain numerous first sub-vector representations and second sub-vector representations. During training, the server enhances the similarity between the HTML code and the vector representation of the text description document of the same web page (target sub-vector representation array). For the HTML code of different web pages and the vector representation of the text description document of other web pages (non-target sub-vector representation array), the server reduces the similarity. Finally, the server optimizes the feature extraction network that can accurately reflect the relationship between the web page and its related text description. Through the above rich scenario examples, the server can use the combination of file data in different formats, and by optimizing the parameters of the feature extraction network, it can achieve effective feature extraction and similarity distinction of file data in different formats but related.

[0059] In an embodiment of the present invention, if the file data corresponding to the first file format is document data, and the file data corresponding to the second file format is multimedia data, the following implementation manner is also provided.

[0060] Obtain at least two second file type target values and a first guiding content framework including a file type target value placeholder, where the second file type target values are used to describe the types of file data corresponding to the second file format, and the first guiding content framework is used to prompt a pre-trained NLG model to generate semantic information of at least two multimedia data segments, and the at least two multimedia data segments described by the semantic information of the at least two multimedia data segments constitute target multimedia data related to the file type target value corresponding to the file type target value placeholder;

[0061] Configure the at least two second file type target values into the file type target value placeholder in the first guiding content framework to obtain first guiding content;

[0062] Based on the first guiding content, generate the first file data instance through the pre-trained NLG model, where the first file data instance is semantic information of multimedia data segments related to the at least two second file type target values;

[0063] Determine the at least two second file type target values as the first file type target values of the first file data instance.

[0064] In an embodiment of the present invention, exemplarily, it is assumed that in a multimedia content management system, the file data corresponding to the first file format is document data in text form, such as articles, reports, etc.; the file data corresponding to the second file format is multimedia data, such as video, audio, etc. The server first obtains at least two second file type target values. For example, the two file type target values ​​of "educational video" and "entertainment video" are obtained. At the same time, the server also obtains a first guide content framework including a file type target value placeholder. The function of this framework is to prompt the pre-trained natural language generation (NLG) model to generate specific multimedia data fragment semantic information. Then, the server configures the two second file type target values ​​of "educational video" and "entertainment video" to the file type target value placeholder in the first guide content framework, thereby obtaining the first guide content. Next, based on the obtained first guide content, the server generates a first file data instance through a pre-trained NLG model. Assume that the NLG model generates semantic information of multimedia data segments about mathematics course explanations according to the target value of "educational video", including the theme, content points, examples, etc. of the explanation; and generates semantic information of multimedia data segments about movie clips according to the target value of "entertainment video", including plot overview, character introduction, etc. These generated semantic information of multimedia data segments constitute the first file data instance, and they are all related to the configured second file type target value. Finally, the server determines the two second file type target values ​​of "educational video" and "entertainment video" as the first file type target value of the first file data instance for subsequent model training and classification operations. For another example, assume that the first file format is an academic research document and the second file format is music audio. The server obtains the two second file type target values ​​of "classical music" and "popular music", as well as the first guide content framework containing the file type target value placeholder. The two target values ​​are filled into the placeholder to obtain the first guide content. Based on this guide content, the NLG model generates semantic information about multimedia data segments of classical music concerts, such as performance repertoire, musician introductions, etc.; and also generates semantic information about multimedia data segments of pop music concerts, such as singer performance, stage effects, etc. The server determines "classical music" and "pop music" as the first file type target values ​​of these generated multimedia data segment semantic information (i.e., the first file data instance). For example, in the image classification scenario, the first file format is image description text, and the second file format is image. The server obtains the two second file type target values ​​of "landscape image" and "character image", as well as the corresponding first guide content framework. After configuring the target value to obtain the first guide content, the NLG model generates semantic information about multimedia data segments of beautiful natural scenery, such as descriptions of mountains and rivers; and semantic information about multimedia data segments of character portraits, including character appearance, expressions, etc.The server sets "landscape image" and "character image" as the first file type target value of the first file data instance. In the scenario of video games, the first file format is the game strategy document, and the second file format is the game video. The server obtains the two second file type target values ​​of "action game video" and "strategy game video" and combines them with the first guide content framework. The NLG model generates multimedia data segment semantic information of the exciting battle scenes of action games and multimedia data segment semantic information of the layout planning of strategy games based on the guidance. The server determines "action game video" and "strategy game video" as the first file type target value of the first file data instance. Through the above multiple specific scenario examples, the server can effectively utilize the second file type target value and the first guide content framework, generate relevant multimedia data segment semantic information through the NLG model, and accurately determine its file type target value, which provides strong support for subsequent processing and analysis.

[0065] In an embodiment of the present invention, if the first guiding content framework is also used to prompt the pre-trained NLG model to generate other file type target values ​​related to the target multimedia data, then the determination of the at least two second file type target values ​​as the first file type target value of the first file data instance can be implemented through the following example.

[0066] The at least two second file type target values ​​and the other file type target values ​​are determined as the first file type target value of the first file data instance.

[0067] In an embodiment of the present invention, illustratively, it is assumed that in a multimedia content creation and management system, the first file format corresponds to a text description document, such as a text version of a movie script, and the second file format corresponds to multimedia data, such as a movie clip. The server obtains at least two second file type target values, such as "comedy movie" and "action movie". At the same time, there is a first guiding content framework, which is not only used to prompt the pre-trained natural language generation (NLG) model to generate multimedia data segment semantic information related to these two target values, but also used to prompt the NLG model to generate other file type target values ​​related to the target multimedia data, such as "science fiction movie". In this case, the server determines the three file type target values ​​of "comedy movie", "action movie" and "science fiction movie" as the first file type target value of the first file data instance. For example, the NLG model generates a movie clip semantic information full of humorous plots and funny character interactions based on the target value of "comedy movie"; generates semantic information of fierce fighting and chasing scenes based on "action movie"; and generates semantic information containing future technology elements and fantasy scenes due to the target value of "science fiction movie". For the first file data instance composed of the semantic information of these generated multimedia data segments, the server determines "comedy movie", "action movie" and "science fiction movie" as its first file type target value. For another example, in a music creation scene, the first file format is the text conception of the music creation, and the second file format is the actual music segment. The server obtains the two second file type target values ​​of "rock music" and "pop music", and the first guide content framework also prompts the NLG model to generate another file type target value of "electronic music". Based on this, the NLG model generates semantic information about the strong rhythm and passionate performance of rock style, the smooth melody and easy-to-sing characteristics of pop music, and the unique synthetic sound effects and dynamic beats of electronic music. For the first file data instance composed of the semantic information of these generated music segment, the server determines "rock music", "pop music" and "electronic music" as its first file type target value. For another example, in the field of image design, the first file format is the text description of the image design, and the second file format is the image work. The server obtains the two second file type target values ​​of "landscape image" and "person image", and the first guide content framework prompts the NLG model to generate the other file type target value of "abstract image". Based on these, the NLG model generates the semantic information of the beautiful natural landscape details in the landscape image, the vivid character image and expression semantic information in the person image, and the unique shape and color combination semantic information in the abstract image. The server determines "landscape image", "person image" and "abstract image" as the first file type target value of this first file data instance.In the context of video game development, the first file format is a game design document, and the second file format is a video clip in the game. The server obtains the two second file type target values ​​of "adventure game" and "role-playing game", and the first guide content framework prompts the generation of the other file type target value of "strategy game". The NLG model generates semantic information of stimulating levels and exploration elements in adventure games, semantic information of character growth and plot development in role-playing games, and semantic information of layout planning and tactical decision-making in strategy games. For the first file data instance composed of the video clip semantic information generated thereby, the server determines "adventure game", "role-playing game" and "strategy game" as its first file type target value. Through the above multiple specific and detailed scenario examples, the server can accurately set the type for the first file data instance based on the prompts of the first guide content framework and the combination of multiple file type target values, providing a clear and comprehensive classification basis for subsequent multimedia data processing and management.

[0068] In an embodiment of the present invention, the first guidance content framework is also used to prompt the pre-trained NLG model not to generate a first file data instance related to the target multimedia data when the at least two second file type target values ​​contradict each other. The embodiment of the present invention also provides the following implementation method.

[0069] The at least two second file type target values ​​are obtained by selecting from a plurality of second file type target values.

[0070] In the embodiment of the present invention, illustratively, it is assumed that in a multimedia content creation system, the server is processing different types of multimedia file generation tasks. First, the server has multiple optional second file type target values, for example, in the video category, there are "horror movie", "love movie", "comedy movie", "science fiction movie", "suspense movie", etc. In the first guide content framework obtained by the server, it is clearly stipulated that the pre-trained natural language generation (NLG) model does not generate a first file data instance related to the target multimedia data when at least two second file type target values ​​contradict each other. For example, if the server selects the two target values ​​of "horror movie" and "comedy movie" from multiple second file type target values. Since the two types of horror and comedy often conflict with each other in atmosphere, plot and expression techniques, it is difficult to be reflected in one multimedia data at the same time. According to the prompt of the first guide content framework, the NLG model will not generate relevant first file data instances. In order to avoid this contradiction, the server needs to carefully select from multiple second file type target values. For example, after carefully analyzing the current creation needs, user preferences, and system resources and capabilities, the server decides to select "love movie" and "comedy movie" as two target values ​​with certain compatibility and relevance. In this way, the NLG model can generate the semantic information of the multimedia data segment that meets the requirements according to these two clear and non-contradictory target values, and then form the first file data instance. For another example, in the scenario of music generation, the multiple second file type target values ​​faced by the server may include "heavy metal music", "classical music", "folk music", "electronic music", etc. If the server initially selects "heavy metal music" and "classical music", because they are very different and contradictory in terms of music style, rhythm, instrument use, etc., the NLG model will not generate the first file data instance according to such a contradictory combination. In order to ensure effective generation, the server may reselect, such as selecting "folk music" and "electronic music". Although these two types of music are different, there is a certain possibility of fusion in some elements, and no obvious contradiction will occur. Then, the NLG model can generate the corresponding music segment semantic information according to the two selected target values ​​to form the first file data instance. For another example, in the field of image generation, possible second file type target values ​​include "realistic landscape image", "abstract art image", "cartoon character image", "surrealist image", etc. If the server accidentally selects "realistic landscape images" and "surrealist images", the NLG model will not generate relevant first file data instances because they have great conflicts in expression and creative concepts. To avoid this, the server may adjust the selection, such as selecting "cartoon character images" and "abstract art images".These two types have certain commonalities in some creativity and expression techniques. Based on such selection, the server can let the NLG model generate meaningful image fragment semantic information to obtain the first file data instance. In the video game scenario, the second file type target value may include "adventure puzzle game", "competitive battle game", "simulation business game", "role-playing game", etc. If the server initially selects "adventure puzzle game" and "competitive battle game", the NLG model cannot generate the first file data instance based on them because their game mechanisms and player experience focuses are completely different and contradictory. In order for the model to work smoothly, the server may reselect, such as "simulation business game" and "role-playing game". These two types have certain compatibility in terms of game elements and player interaction. The NLG model can generate relevant game scenes and gameplay semantic information based on these two choices to form the first file data instance. Through the above various detailed scenario examples, it can be seen that the server makes reasonable selections from many second file type target values ​​under the rules of the first guided content framework to ensure that the NLG model can effectively generate valuable first file data instances and avoid generation failures caused by contradictions in target values.

[0071] In the embodiment of the present invention, the classification based on the first vector representation is performed through a basic file classification model to obtain an inferred file type, which can be implemented through the following examples.

[0072] Based on the first vector representation, performing a feature mapping operation through an element mapping model to obtain a first mapping vector representation, wherein the element mapping model is used to map the vector representation from the target feature domain to a model feature domain, wherein the model feature domain is a feature domain that can be classified by the file classification model;

[0073] Based on the first mapping vector representation, classify using the basic file classification model to obtain an inferred file type;

[0074] The embodiments of the present invention also provide the following implementation modes.

[0075] Acquire first inferred file data corresponding to the second file format;

[0076] Performing a feature extraction operation on the first inference file data through the first feature extraction network to obtain a second vector representation;

[0077] Based on the second vector representation, performing a feature mapping operation through the element mapping model to obtain a second mapping vector representation;

[0078] Based on the second mapping vector representation, classification is performed using the file classification model to obtain the type of the first inferred file data.

[0079] In the embodiment of the present invention, illustratively, it is assumed that in a multimedia file processing system, the first file format is a simple text description and the second file format is a complex video file. The server first obtains a first file data instance corresponding to the first file format, such as a short text introduction to a certain movie. The text introduction is subjected to feature extraction through a first feature extraction network to obtain a first vector representation. Then, based on the first vector representation, the server performs a feature mapping operation through an element mapping model. The element mapping model can map the first vector representation from the target feature domain (generated by the first feature extraction network) to the model feature domain. It is assumed that the model feature domain is a feature space specially designed for the file classification model and more suitable for classification. After mapping, a first mapping vector representation is obtained. Then, based on the first mapping vector representation, classification is performed through a basic file classification model to obtain an inferred file type. For example, the classification model infers that the movie may belong to the "action movie" type. On the other hand, the server obtains the first inferred file data corresponding to the second file format, such as a complete movie video file. The video file is subjected to a feature extraction operation through a first feature extraction network to obtain a second vector representation. Based on this second vector representation, the feature mapping operation is performed again through the element mapping model to obtain the second mapping vector representation. Finally, based on this second mapping vector representation, the file classification model is used for classification to obtain the type of the first inferred file data (movie video). Assume that the classification result is the "comedy" type. For another example, in the image classification scenario, the first file format is a text description of a picture, and the second file format is the actual picture file. The server obtains a text description of a landscape picture as the first file data instance, and obtains the first vector representation through feature extraction. Then it is mapped to the first mapping vector representation through the element mapping model, and then classified into the "natural scenery" type by the basic file classification model. For the first inferred file data corresponding to the second file format, that is, the actual landscape picture, the first feature extraction network extracts features to obtain the second vector representation, and the element mapping model obtains the second mapping vector representation, and finally the file classification model classifies it into the "tourist attraction scenery" type. In the case of audio classification, the first file format is a text description of a piece of music, and the second file format is the audio file itself. The server obtains a text description of a piece of music, extracts features and maps them, and the classification model infers that it is "classical music". For actual audio files, the same process is followed and they are classified as "symphony" type. In the document classification scenario, the first file format is the summary of the document, and the second file format is the complete document. The server processes and classifies the document summary, for example, as "scientific report". For the complete document, after a series of operations, it is classified as "computer technology research report". In the web page classification scenario, the first file format is a simple description of the web page, and the second file format is the complete web page.The server processes the webpage description and classifies it as a "shopping website". For the complete webpage, it is finally classified as a "clothing shopping website". Through the above detailed scenario examples, the server can effectively use feature extraction, element mapping and file classification models to accurately infer and classify file data in different formats.

[0080] In the embodiments of the present invention, the following implementation modes are also provided.

[0081] Acquire a second file data instance corresponding to the first file format, wherein the second file data instance has a third file type target value;

[0082] Based on the second file data instance, performing a feature extraction operation on the second file data instance through the first feature extraction network to obtain a third vector representation;

[0083] Based on the third vector representation, performing a feature mapping operation through a basic element mapping model to obtain an inferred mapping vector representation;

[0084] Based on the inferred mapping vector representation, classify using the basic file classification model to obtain second inferred file data;

[0085] Based on the deviation between the second inferred file data and the third file type target value, the model network parameters of the basic file classification model are fixed, and the model network parameters of the basic element mapping model are optimized to obtain the element mapping model.

[0086] In the embodiment of the present invention, illustratively, it is assumed that in an image classification system, the first file format is a simple text description of the image, and the second file format is the image itself. The server first obtains the second file data instance corresponding to the first file format, such as a detailed text description of a photo of a person. This second file data instance has a clear third file type target value, such as "person portrait". Based on this second file data instance, the server performs a feature extraction operation on it through the first feature extraction network. This network may extract key features in the text description, such as the feature description of the person, background information, etc., to obtain a third vector representation. Next, based on this third vector representation, the server performs a feature mapping operation through the basic element mapping model to obtain an inferred mapping vector representation. Then, based on this inferred mapping vector representation, classification is performed through the basic file classification model to obtain the second inferred file data. Assume that the classification result is "artistic person photo". Due to the deviation between the second inferred file data and the third file type target value ("person portrait"), the server fixes the model network parameters of the basic file classification model, because it is believed that the judgment ability of the classification model is relatively reliable at this time. Then, the server focuses on optimizing the model network parameters of the basic element mapping model. By continuously adjusting the parameters of the model, more accurate mapping results can be obtained based on the same type of input in the future, and finally an optimized element mapping model is obtained. For example, in the audio classification scenario, the first file format is a text description of a piece of audio, such as "a piece of pop music with a cheerful rhythm", and its third file type target value is "pop music". The server processes this text description through the first feature extraction network to obtain a third vector representation. After the basic element mapping model obtains the inferred mapping vector representation, and then classifies it through the basic file classification model, the second inferred file data obtained can be "electronic pop music". Due to the deviation from the third file type target value, the server fixes the parameters of the basic file classification model and optimizes the network parameters of the basic element mapping model to improve the accuracy of subsequent feature mapping. In the example of document classification, the first file format is a summary of the document, such as "a technical analysis on the development of artificial intelligence", and the third file type target value is "scientific and technological documents". After a series of operations, the server obtains the second inferred file data, such as "computer technology documents". When there is a deviation, the parameters of the basic file classification model are fixed, and the basic element mapping model is optimized to obtain a more accurate element mapping model. In the scenario of web page classification, the first file format is a brief introduction of the web page, such as "a website page providing travel guides", and the third file type target value is "travel web page". After processing and classification, the second inferred file data obtained can be "overseas travel web page". The basic element mapping model is optimized according to the deviation so that it can map features more accurately.Through the above rich and diverse scenario examples, the server can adjust the model parameters in a targeted manner during the training and optimization process to improve the accuracy and reliability of classification.

[0087] In the embodiments of the present invention, the following implementation modes are also provided.

[0088] Acquire a second guide content framework and an element mapping request, wherein the second guide content framework includes a vector representation placeholder and an element mapping request placeholder, and the element mapping request is used to call the basic file classification model to generate the second inferred file data;

[0089] Configuring the inferred mapping vector representation to a vector representation placeholder in the second guided content frame, configuring the element mapping request to an element mapping request placeholder in the second guided content frame, and obtaining a second guided content;

[0090] The method of performing classification based on the inferred mapping vector representation through the basic file classification model to obtain the second inferred file data can be implemented through the following example.

[0091] Based on the second guidance content, classification is performed using the basic file classification model to obtain the second inferred file data.

[0092] In an embodiment of the present invention, illustratively, it is assumed that in a text classification system, the server is processing classification tasks for different types of texts. The server first obtains a second guide content framework and an element mapping request. For example, the second guide content framework is a template in a specific format, which includes a vector representation placeholder and an element mapping request placeholder. The element mapping request is an instruction for calling the basic file classification model to generate second inferred file data. Assume that the server obtains a text about a scientific article as an object to be classified. After performing a feature extraction operation on this text through the first feature extraction network, an inferred mapping vector representation is obtained. Then, the server configures this inferred mapping vector representation to the vector representation placeholder in the second guide content framework, and configures the element mapping request to the element mapping request placeholder, thereby obtaining a complete second guide content. Next, based on this second guide content, classification is performed through the basic file classification model. After receiving this guide content, the basic file classification model processes and analyzes the vector representation and request therein. For example, in the classification of this scientific article, the model may determine that this text belongs to the category of "computer technology research" based on the features such as keywords, sentence structure, and professional terms in the text, combined with the element mapping request configured in the guide content, and finally obtain the second inferred file data. For another example, in an image description classification scenario, the server obtains a detailed description text of a picture. After feature extraction and mapping, the inferred mapping vector representation is obtained. According to the format of the second guide content framework, the vector representation and the element mapping request are configured to form a complete guide content. The basic file classification model classifies the image description according to this guide content. For example, it is classified as "city landscape description" to obtain the second inferred file data. In an example of audio feature classification, the server obtains the feature description information of a piece of audio. After processing, the inferred mapping vector representation is obtained and configured in the second guide content framework. Based on this guide content, the basic file classification model may classify this audio as "rock music performance", which is the second inferred file data obtained. In a web page content classification scenario, the server obtains the main content description of a web page. After the same process, the basic file classification model may classify this webpage as "online education platform" according to the second guidance content, which is the second inferred file data obtained. Through the above various specific scenario examples, the server can effectively use the second guidance content framework and element mapping request to enable the basic file classification model to accurately classify different types of file data and obtain reliable second inferred file data.

[0093] In the embodiment of the present invention, the optimization of the model network parameters of the basic file classification model based on the deviation between the inferred file type and the first file type target value to obtain the file classification model can be implemented through the following examples.

[0094] Based on the deviation between the inferred file type and the first file type target value, the model network parameters of the basic file classification model are optimized to obtain a file classification model; and based on the deviation between the inferred file type and the first file type target value, the model network parameters of the element mapping model are optimized to obtain an updated element mapping model.

[0095] In the embodiment of the present invention, illustratively, it is assumed that in a multimedia file classification system, the server is processing the classification task of audio files. First, the server obtains a large number of audio files as first file data instances corresponding to the first file format. These audio files may include various types, such as pop music, classical music, rock music, etc. Each audio file is marked with its actual first file type target value, for example, a certain audio is marked as "pop music". The server extracts features from these audio files through a first feature extraction network to obtain corresponding vector representations. Then, these vector representations are mapped to feature representations suitable for processing by the basic file classification model through a basic element mapping model. Next, the basic file classification model classifies these feature representations to obtain inferred file types. Assume that for a certain audio file, the basic file classification model infers that it is "rock music", while its actual first file type target value is "pop music", which produces a deviation. Based on this deviation, the server starts to optimize the model. On the one hand, the server adjusts the model network parameters of the basic file classification model. For example, if the model has problems in the weight allocation of certain features, resulting in classification errors, the server will adjust the weights of these features to make the model more accurate in subsequent classifications. Through continuous adjustment and optimization, a file classification model that can be more accurately classified is finally obtained. On the other hand, the server will also optimize the model network parameters of the element mapping model based on this deviation. For example, when the element mapping model converts the original audio feature vector into a feature representation suitable for the classification model, some information is lost or distorted, causing the classification model to make an incorrect judgment. The server will adjust the parameters of the element mapping model to ensure that it can more accurately convert the original features into feature representations that are more conducive to the classification model. After optimization, an updated element mapping model is obtained. For another example, in the image classification scenario, the server processes painting images of different styles. The first file type target value can be "Impressionist painting", "Realistic painting", etc. When the file type inferred by the basic file classification model is inconsistent with the actual first file type target value, the server will optimize the parameters of the basic file classification model and the element mapping model at the same time. For the basic file classification model, the sensitivity to certain features such as color, line, composition, etc. may be adjusted; for the element mapping model, the extraction and conversion methods of image features may be improved to reduce the adverse effects on the classification results. In the document classification scenario, it is assumed that the first file type target value is "scientific and technological papers", "literary reviews", etc. If the inference results of the classification model deviate, the server will optimize the basic file classification model and element mapping model based on the deviation. The basic file classification model may pay more attention to features such as professional terminology and citation format in the document; the element mapping model will optimize the processing and conversion of features such as text structure and vocabulary distribution.In the video classification scenario, the first file type target value is, for example, "comedy short film" or "documentary". When a classification deviation occurs, the server adjusts the parameters of the basic file classification model and the element mapping model. The basic file classification model may pay more attention to features such as the plot development and picture style of the video; the element mapping model will improve the integration and conversion methods of video frame features, audio features, etc. Through the above detailed and specific scenario examples, the server can comprehensively optimize the basic file classification model and the element mapping model based on the deviation between the inferred file type and the first file type target value, thereby improving the accuracy and performance of the entire classification system.

[0096] In an embodiment of the present invention, if the file data corresponding to the first file format is document data, and the file data corresponding to the second file format is multimedia data, then the second file data instance includes at least two multimedia data fragment semantic information, and the third file type target value is the global semantic information of the multimedia data corresponding to the second file data instance or the at least two multimedia data fragment semantic information.

[0097] In an embodiment of the present invention, illustratively, it is assumed that the system is in a multimedia data processing system with rich content, wherein the file data corresponding to the first file format is detailed document data, such as academic research reports, technical description documents, etc.; and the file data corresponding to the second file format is multimedia data, such as a combination of video, audio and image. The second file data instance acquired by the server contains at least two multimedia data fragment semantic information. For example, in a project about natural science, the second file data instance may include a video clip introducing the growth process of plants, a picture showing animal ecology, and an audio clip describing natural phenomena. For this second file data instance, its third file type target value is the global semantic information of these multimedia data. Assume that this global semantic information is marked as "natural science education resources". The server first processes these multimedia data fragments. For video fragments, key frames, object features, action sequences and other information may be extracted through image recognition and video content analysis technology; for pictures, color distribution, object shape, texture and other features may be extracted; for audio fragments, sound frequency, rhythm, intonation and other features may be extracted. Then, the server integrates and analyzes the extracted feature information. Through the first feature extraction network, these complex multimedia features are converted into vector representations. Next, based on these vector representations, feature mapping operations are performed through the basic element mapping model to obtain inferred mapping vector representations. Then, classification is performed through the basic file classification model. During the classification process, the model will comprehensively consider the semantic information of all multimedia data segments, not just the features of a single segment. Since the classification result is based on the global semantics of the entire second file data instance, even if individual multimedia data segments may have different local features, the final classification result should still conform to the global semantic information of "natural science education resources". For example, if during the processing, a certain video segment may be more like entertainment content when viewed alone, but combined with the overall information of the picture and audio segments, and the global semantics they jointly constitute, the final classification result is still related to natural science education. In another scenario, assume that the second file data instance includes a video introducing a historical event, relevant historical figures pictures, and the background audio at the time. Its third file type target value is "historical education materials". The server also extracts and integrates features for each multimedia data segment. After a series of processing and classification operations, the final classification result should be consistent with the global semantic information of "historical education materials". Through the above specific scenario examples, the server can accurately process the second file data instance containing semantic information of multiple multimedia data segments, and effectively classify and process it according to its global semantic information.

[0098] In the embodiments of the present invention, the following implementation modes are also provided.

[0099] Obtain important information of the first file data instance;

[0100] Perform feature extraction operations on the important information to obtain an important vector representation;

[0101] Based on the first vector representation, classify through a basic file classification model to obtain an inferred file type, which can be implemented through the following examples.

[0102] Based on the first vector representation and the important vector representation, classify through the basic file classification model to obtain the inferred file type.

[0103] In an embodiment of the present invention, illustratively, it is assumed that in a document classification system, a server is processing a large number of text files. The first file format corresponds to various types of document data. The server obtains a first file data instance, such as an academic paper on medical technology research. Then, the server obtains important information from it. In this example, the important information may include the core method of the research, the main experimental results, and the key conclusions drawn. The server performs feature extraction operations on these important information. For example, through word frequency analysis, keyword extraction, semantic understanding and other technologies, the important information is converted into an important vector representation. Then, the server has performed a feature extraction operation on the entire first file data instance through the first feature extraction network to obtain a first vector representation. When the basic file classification model is used for classification, it is no longer based solely on the first vector representation, but combined with the important vector representation obtained previously. The basic file classification model will consider the feature information in these two vector representations at the same time. For example, the first vector representation may contain general features such as the overall structure and language style of the paper, while the important vector representation highlights key features such as professional terms related to medical technology research and specific data patterns. After the model comprehensively analyzes the two vector representations, it infers the file type. For example, this academic paper is classified into the category of "Frontier Research in Medical Technology". For another example, in an image classification scenario, the first file data instance is a complex picture containing multiple elements. The server obtains important information in the picture, which can be the main person, key items or significant scene features in the picture. The important information is extracted to obtain important vector representations. At the same time, the first vector representation is obtained by extracting features from the entire picture. The basic file classification model combines the first vector representation and the important vector representation for classification. The picture may be classified as "artworks of a specific theme" based on the overall color, composition and characteristics of important elements of the picture. In an audio classification example, the first file data instance is a piece of music. The server obtains important information in the music, such as the characteristics of the main melody, the rhythm pattern of the climax part or a unique instrument performance segment. After feature extraction, the important vector representation is obtained, combined with the first vector representation obtained by extracting general features of the entire music. After comprehensive consideration, the basic file classification model classifies this piece of music as "music works of a certain style". In a web page classification scenario, the first file data instance is a web page with rich content. The server extracts important information from the web page, such as the core product introduction, the main terms of service or the key points in user comments. After obtaining the important vector representation, together with the first vector representation extracted from the overall layout and link structure of the web page, the basic document classification model finally classifies the web page as a "specific type of commercial web page".Through the various detailed scenario examples above, the server can make full use of the important information of the first file data instance, combine the overall characteristics, and perform more accurate classification through the basic file classification model to meet the needs of different application scenarios.

[0104] In the embodiments of the present invention, the following implementation modes are also provided.

[0105] Acquire a third guided content framework and an inference request, wherein the third guided content framework includes a vector representation placeholder and an inference request placeholder, and the inference request is used to call the basic file classification model to generate the inferred file type;

[0106] Configuring the first vector representation to a vector representation placeholder in the third guidance content frame, configuring the inference request to an inference request placeholder in the third guidance content frame, and obtaining third guidance content;

[0107] The method of performing classification based on the first vector representation through a basic file classification model to obtain an inferred file type can be implemented through the following example.

[0108] Based on the third guidance content, classification is performed using a basic file classification model to obtain an inferred file type.

[0109] In the embodiment of the present invention, illustratively, it is assumed that in a multimedia file classification system, the server is processing different types of multimedia files. The server first obtains a third guide content framework and an inference request. For example, the third guide content framework is a template with a specific format and structure, which contains a vector representation placeholder and an inference request placeholder. The inference request is an instruction for instructing the basic file classification model to perform file type inference. Assume that the server obtains a video file as a first file data instance. The first feature extraction network performs a feature extraction operation on the video file to obtain a first vector representation. Then, the server configures the first vector representation to the vector representation placeholder in the third guide content framework, and configures the inference request to the inference request placeholder, thereby obtaining a complete third guide content. For example, this video file is a wonderful collection of a football game. The first vector representation obtained after feature extraction contains feature information such as player movements, game scenes, and audience reactions. These feature information are filled into the vector representation placeholder, and the inference request (such as "please infer the file type of this video") is filled into the inference request placeholder. Next, based on this third guide content, the basic file classification model begins to perform classification operations. After receiving the third guidance content, the basic file classification model will read the vector representation and inference request therein, and infer the type of the video file based on its internal algorithm and learned patterns. Assume that after analysis and calculation, the model finally infers the file type as "sports event video". For another example, in the scenario of processing audio files, the server obtains a song as the first file data instance. After feature extraction, the first vector representation is obtained, which includes the melody, rhythm, harmony and other features of the music. In the same way, the first vector representation and inference request are configured in the third guidance content framework to form the third guidance content. Based on this guidance content, the basic file classification model infers that the type of the audio file is "pop music". In the scenario of image classification, the server obtains a landscape photo as the first file data instance. After feature extraction, the first vector representation is obtained, which includes the color, composition, object shape and other features of the image. The first vector representation and inference request are configured in the third guidance content framework, and the basic file classification model infers that the file type is "natural scenery image". In the scenario of document classification, it is assumed that the first file data instance is a scientific paper. After the feature extraction obtains the first vector representation, it is configured into the third guide content framework. The basic file classification model finally infers that the type of this document is "computer science research paper". Through the above detailed scenario examples, the server can effectively use the third guide content framework and the basic file classification model to accurately infer the type of different types of multimedia files.

[0110] In the embodiment of the present invention, the third guide content frame further includes an important information placeholder. The embodiment of the present invention also provides the following implementation manner.

[0111] Obtaining important information of the first file data instance;

[0112] Performing a feature extraction operation on the important information to obtain an important vector representation;

[0113] The configuring of the first vector representation to the vector representation placeholder in the third guided content frame and the configuring of the inference request to the inference request placeholder in the third guided content frame to obtain the third guided content may be implemented through the following example.

[0114] The first vector representation is configured to the vector representation placeholder in the third guide content frame, the inference request is configured to the inference request placeholder in the third guide content frame, and the important vector representation is configured to the important information placeholder in the third guide content frame to obtain the third guide content.

[0115] In the embodiment of the present invention, illustratively, it is assumed that in a multimedia file processing system, the server is processing various types of multimedia files. First, the server obtains an image file as a first file data instance. This image can be a complex urban landscape. The server obtains a third guide content framework, which contains not only a vector representation placeholder and an inference request placeholder, but also an important information placeholder. Then, the server obtains important information from the image file. In this urban landscape, important information may include iconic buildings, major traffic routes, or unique urban landscape elements. Feature extraction operations are performed on these important information to obtain important vector representations. At the same time, feature extraction is performed on the entire image file through the first feature extraction network to obtain a first vector representation. Next, the first vector representation is configured to the vector representation placeholder in the third guide content framework, the inference request (such as "infer the file type of this image") is configured to the inference request placeholder, and the important vector representation is configured to the important information placeholder, so as to obtain a complete third guide content. For example, the first vector representation may contain general information such as the overall color distribution and texture features of the image, while the important vector representation highlights key information such as the shape features of the iconic buildings and the direction of the traffic routes. In another scenario, assume that the first file data instance is an audio file, such as a symphony. The server obtains important information from this symphony, which may be the melody direction of the main melody, the combination of instruments in the climax part, or a unique playing technique. After feature extraction, an important vector representation is obtained, and the first vector representation of the entire audio is also obtained. When configuring the third guide content, the first vector representation, the inference request, and the important vector representation are placed in the corresponding placeholders respectively. For another example, the first file data instance is a video file, and the content is a stage performance. The server obtains the wonderful performance clips of the main actors on the stage, the unique stage setting, etc. as important information, and converts them into important vector representations. Combined with the first vector representation of the entire video, it is configured into the third guide content framework as required. Through the above multiple specific scenario examples, the server can make full use of the important information of the first file data instance, together with the vector representation obtained by general feature extraction, to accurately construct the third guide content, providing a more comprehensive and accurate information basis for subsequent file type inference.

[0116] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned multi-client append writing processing method. Figure 2 As shown, Figure 2The block diagram of the computer device 100 provided in the embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112 and a communication unit 113. To achieve data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected to each other. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0117] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A multi-client additional writing processing method, characterized in that: include: Get concurrent append write operations of multiple clients on the target file; Generate a global append write lock request according to the target append write operation, and call the metadata server corresponding to the target file to lock the target file; the target append write operation is a concurrent append write operation of the target client on the target file among the concurrent append write operations of the multiple clients on the target file, and the target client is any client among the multiple clients; When the global append write lock is locked successfully, the file size of the target file is obtained from the metadata server, and the file offset is determined according to the file size; Performing the target append write operation on the target file in a direct I / O write mode according to the file offset, and persisting the write data corresponding to the target append write operation to the data server; In the case where data persistence is complete, the metadata corresponding to the target file in the metadata server is updated, and the metadata server initiates an unlock request to release the global append write lock, and returns the append write result corresponding to the target append write operation to the target client; In the case that the plurality of clients all receive corresponding append writing results, calling a pre-trained file classification model to perform type identification on the metadata of the target file; When the type identification result of the metadata of the target file is consistent with the original file type of the metadata of the target file, maintaining the storage location of the target file; When the type identification result of the metadata of the target file is inconsistent with the original file type of the metadata of the target file, transferring the target file to a storage location corresponding to the type identification result for storage; The file classification model is obtained by: Acquire a first file data instance corresponding to a first file format, wherein the first file data instance has a first file type target value, and the first file type target value is used to indicate an actual file type of the first file data instance; Performing a feature extraction operation on the first file data instance through a first feature extraction network to obtain a first vector representation, wherein the first feature extraction network is used to convert the file data corresponding to the first file format to a target feature domain, wherein the target feature domain is a feature domain where the vector representation corresponding to the file data of the second file format is located, wherein the first file format and the second file format are different file formats, and the complexity of the file data corresponding to the first file format is less than the complexity of the file data corresponding to the second file format; Based on the first vector representation, classify the file using a basic file classification model to obtain an inferred file type; Based on the deviation between the inferred file type and the first file type target value, the model network parameters of the basic file classification model are optimized to obtain a file classification model, so that the file classification model is used to classify the type of file data corresponding to the second file format.

2. The method according to claim 1, characterized in that The method further comprises: Acquire first inferred file data corresponding to the second file format; performing a feature conversion operation on the first inferred file data through a second feature extraction network to obtain a second vector representation, wherein the second feature extraction network is used to convert the file data corresponding to the second file format into the target feature domain; Based on the second vector representation, classification is performed using the file classification model to obtain the type of the first inferred file data.

3. The method according to claim 1, characterized in that The first feature extraction network includes a first file format feature extraction network and a second file format feature extraction network, the first file format feature extraction network is used to convert file data corresponding to the first file format into the target feature domain, and the second file format feature extraction network is used to convert file data corresponding to the second file format into the target feature domain; The step of performing a feature extraction operation on the first file data instance through a first feature extraction network to obtain a first vector representation includes: Performing a feature extraction operation on the first file data instance through the first file format feature extraction network to obtain a first vector representation; The method further comprises: Acquire first inferred file data corresponding to the second file format; Performing a feature extraction operation on the first inferred file data through the second file format feature extraction network to obtain a second vector representation; Based on the second vector representation, classify using the file classification model to obtain the type of the first inferred file data; The first file format feature extraction network and the second file format feature extraction network are obtained by: Acquire at least two file data instance data groups, wherein the file data instance data groups include a first sub-file data instance corresponding to the first file format and a second sub-file data instance corresponding to the second file format, wherein the first sub-file data instance and the second sub-file data instance in the same file data instance data group represent the same file information; For a target file data instance data group in the at least two file data instance data groups, performing a feature extraction operation on a first sub-file data instance in the target file data instance data group by using a basic first file format feature extraction network to obtain a first sub-vector representation; Performing a feature extraction operation on the second sub-file data instance in the target file data instance data group through a basic second file format feature extraction network to obtain a second sub-vector representation; Determine the at least two file data instance data groups as the target file data instance data groups respectively, and obtain at least two first sub-vector representations and at least two second sub-vector representations; Following a training strategy of enhancing the similarity of a target sub-vector representation array and reducing the similarity of a non-target sub-vector representation array, the model network parameters of the basic first file format feature extraction network and the model network parameters of the basic second file format feature extraction network are optimized to obtain the first file format feature extraction network and the second file format feature extraction network, wherein the first sub-vector representation and the second sub-vector representation included in the target sub-vector representation array are obtained based on the same file data instance data group, and the first sub-vector representation and the second sub-vector representation included in the non-target sub-vector representation array are not obtained based on the same file data instance data group.

4. The method according to claim 1, characterized in that: If the file data corresponding to the first file format is document data, and the file data corresponding to the second file format is multimedia data, the method further includes: Acquire at least two second file type target values ​​and a first guide content framework including a file type target value placeholder, wherein the second file type target value is used to describe the type of file data corresponding to the second file format, and the first guide content framework is used to prompt a pre-trained NLG model to generate at least two multimedia data segment semantic information, wherein the at least two multimedia data segments described by the at least two multimedia data segment semantic information constitute target multimedia data related to the file type target value corresponding to the file type target value placeholder; Allocate the at least two second file type target values ​​to the file type target value placeholders in the first guide content frame to obtain the first guide content; Based on the first guide content, generating the first file data instance through the pre-trained NLG model, the first file data instance being multimedia data segment semantic information related to the at least two second file type target values; Determine the at least two second file type target values ​​as first file type target values ​​of the first file data instance; If the first guiding content framework is also used to prompt the pre-trained NLG model to generate other file type target values ​​related to the target multimedia data, then determining the at least two second file type target values ​​as the first file type target value of the first file data instance includes: Determine the at least two second file type target values ​​and the other file type target values ​​as the first file type target value of the first file data instance; The first guiding content framework is further used to prompt the pre-trained NLG model not to generate a first file data instance related to the target multimedia data when the at least two second file type target values ​​are contradictory. The method further includes: The at least two second file type target values ​​are obtained by selecting from a plurality of second file type target values.

5. The method according to claim 1, characterized in that The step of performing classification based on the first vector representation by using a basic file classification model to obtain an inferred file type includes: Based on the first vector representation, performing a feature mapping operation through an element mapping model to obtain a first mapping vector representation, wherein the element mapping model is used to map the vector representation from the target feature domain to a model feature domain, wherein the model feature domain is a feature domain that can be classified by the file classification model; Based on the first mapping vector representation, classify using the basic file classification model to obtain an inferred file type; The method further comprises: Acquire first inferred file data corresponding to the second file format; Performing a feature extraction operation on the first inference file data through the first feature extraction network to obtain a second vector representation; Based on the second vector representation, performing a feature mapping operation through the element mapping model to obtain a second mapping vector representation; Based on the second mapping vector representation, classify using the file classification model to obtain the type of the first inferred file data; The method further comprises: Acquire a second file data instance corresponding to the first file format, wherein the second file data instance has a third file type target value; Based on the second file data instance, performing a feature extraction operation on the second file data instance through the first feature extraction network to obtain a third vector representation; Based on the third vector representation, performing a feature mapping operation through a basic element mapping model to obtain an inferred mapping vector representation; Based on the inferred mapping vector representation, classify using the basic file classification model to obtain second inferred file data; Based on the deviation between the second inferred file data and the third file type target value, fixing the model network parameters of the basic file classification model, optimizing the model network parameters of the basic element mapping model, and obtaining the element mapping model; The method further comprises: Acquire a second guide content framework and an element mapping request, wherein the second guide content framework includes a vector representation placeholder and an element mapping request placeholder, and the element mapping request is used to call the basic file classification model to generate the second inferred file data; Configuring the inferred mapping vector representation to a vector representation placeholder in the second guided content frame, configuring the element mapping request to an element mapping request placeholder in the second guided content frame, and obtaining a second guided content; The step of performing classification based on the inferred mapping vector representation by using the basic file classification model to obtain second inferred file data includes: Based on the second guidance content, classification is performed using the basic file classification model to obtain the second inferred file data.

6. The method according to claim 5, characterized in that The step of optimizing the model network parameters of the basic file classification model based on the deviation between the inferred file type and the first file type target value to obtain the file classification model comprises: Based on the deviation between the inferred file type and the first file type target value, the model network parameters of the basic file classification model are optimized to obtain a file classification model; and based on the deviation between the inferred file type and the first file type target value, the model network parameters of the element mapping model are optimized to obtain an updated element mapping model.

7. The method according to claim 6, characterized in that If the file data corresponding to the first file format is document data, and the file data corresponding to the second file format is multimedia data, then the second file data instance includes at least two multimedia data fragment semantic information, and the third file type target value is the global semantic information of the multimedia data corresponding to the second file data instance or the at least two multimedia data fragment semantic information.

8. The method according to claim 1, characterized in that The method further comprises: Obtaining important information of the first file data instance; Performing a feature extraction operation on the important information to obtain an important vector representation; The step of performing classification based on the first vector representation by using a basic file classification model to obtain an inferred file type includes: Based on the first vector representation and the important vector representation, classify using the basic file classification model to obtain the inferred file type; The method further comprises: Acquire a third guided content framework and an inference request, wherein the third guided content framework includes a vector representation placeholder and an inference request placeholder, and the inference request is used to call the basic file classification model to generate the inferred file type; Configuring the first vector representation to a vector representation placeholder in the third guidance content frame, configuring the inference request to an inference request placeholder in the third guidance content frame, and obtaining third guidance content; The step of performing classification based on the first vector representation by using a basic file classification model to obtain an inferred file type includes: Based on the third guidance content, classify the file using a basic file classification model to obtain an inferred file type; The third guide content frame also includes an important information placeholder, and the method further includes: Obtaining important information of the first file data instance; Performing a feature extraction operation on the important information to obtain an important vector representation; The configuring the first vector representation to a vector representation placeholder in the third guided content frame, configuring the inference request to an inference request placeholder in the third guided content frame, and obtaining third guided content, comprises: The first vector representation is configured to the vector representation placeholder in the third guide content frame, the inference request is configured to the inference request placeholder in the third guide content frame, and the important vector representation is configured to the important information placeholder in the third guide content frame to obtain the third guide content.

9. A server system, characterized in that: The method comprises a server, wherein the server is used to execute the method described in any one of claims 1 to 8.

Images