Data processing method and device based on large language model, equipment and storage medium

Abstract

The invention relates to a data processing method and device based on a large language model, equipment and a storage medium. The method comprises the following steps: acquiring input data, matching the input data with a plurality of pre-constructed data matrixes to obtain a data matrix matched with the data type of the input data, and determining the matched data matrix as a current data matrix; constructing an inference tree model based on the input data and the current data matrix; determining a plurality of candidate sequences based on the inference tree model; and performing parallel verification on the plurality of candidate sequences, determining the output probability distribution of each candidate sequence, and obtaining an output sequence matched with the data content of the input data according to the output probability distribution. Compared with the prior art, the data processing efficiency can be improved, and the data output quality and accuracy can also be improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a data processing method, apparatus, device and storage medium based on a large language model. Background Art

[0002] Large Language Models (LLMs) can generate natural language text or understand the meaning of language text, and LLMs can handle a variety of natural language tasks, such as text classification, question and answer, dialogue, etc., and are an important path to artificial intelligence.

[0003] In recent years, LLMs have made breakthrough progress in the field of natural language processing, demonstrating powerful language understanding and generation capabilities. However, with the continuous expansion of model scale and the increasing complexity of application scenarios, LLMs still face many challenges in actual deployment and use.

[0004] On the one hand, with the surge in the number of parameters, the computational overhead of LLMs has increased exponentially, resulting in slow reasoning speed and difficulty in meeting the needs of real-time interaction. In particular, when generating long texts, the response speed seriously affects the efficiency of text generation. In addition, existing reasoning acceleration technologies such as model quantization and knowledge distillation can improve reasoning speed to a certain extent, but often at the expense of model performance, resulting in low reasoning accuracy. Summary of the invention

[0005] Based on this, it is necessary to provide a data processing method, device, equipment and storage medium based on a large language model that can improve data processing efficiency in response to the above technical problems.

[0006] In a first aspect, the present application provides a data processing method based on a large language model, the method comprising:

[0007] Acquire input data, match the input data with a plurality of pre-constructed data matrices, obtain a data matrix that matches the data type of the input data, and determine the matched data matrix as a current data matrix;

[0008] Building an inference tree model based on the input data and the current data matrix;

[0009] Based on the inference tree model, determining multiple candidate sequences;

[0010] The plurality of candidate sequences are verified in parallel to determine the output probability distribution of each candidate sequence, and an output sequence matching the data content of the input data is obtained according to the output probability distribution.

[0011] In one embodiment, the acquiring of input data and matching of the input data with a plurality of pre-constructed data matrices comprises:

[0012] Load multiple data matrices;

[0013] Input data is acquired, and the input data is input into a trained matching network, so that the input data is matched with the multiple data matrices respectively through the matching network.

[0014] In one embodiment, before inputting the input data into the pre-trained matching network, the method further includes:

[0015] Create multiple data matrices based on the data type of the sample data;

[0016] Initializing each data in the data matrix;

[0017] A pre-built matching network is trained based on the initialized data matrix to obtain a trained matching network.

[0018] In one embodiment, constructing an inference tree model based on the input data and the current data matrix includes:

[0019] Determine a key field of the input data, and determine the key field as a current field;

[0020] Creating a root node of the inference tree model, and determining the root node as the current node;

[0021] Assigning the current field to the current node;

[0022] Performing correlation matching between the current field and the data in the current data matrix to determine a plurality of secondary fields matching the current field;

[0023] According to the number of the secondary fields, subnodes corresponding to the secondary fields are created under the current node, the secondary fields are determined as the current fields and the subnodes are determined as the current nodes, and the step of assigning the current fields to the current node is returned to iterate until the end condition is reached to obtain the inference tree model.

[0024] In one embodiment, the determining of multiple candidate sequences based on the inference tree model includes:

[0025] Taking the root node as the starting point, traverse each node of the inference tree model;

[0026] When all current child nodes of a current node point to the same lower-level child node, the current child node and the lower-level child node are merged into a new child node, and the current node is pointed to the new child node, so as to merge and update each node of the inference tree model to obtain updated nodes; the current node is the root node or child node corresponding to the current traversal moment;

[0027] The updated nodes are respectively integrated in the traversal order to obtain multiple candidate sequences pointing from the root node to the lower-level child nodes step by step.

[0028] In one embodiment, the verifying the plurality of candidate sequences in parallel, determining the output probability distribution of each candidate sequence, and obtaining an output sequence matching the data content of the input data according to the output probability distribution includes:

[0029] Creating a zero matrix having the same number as the number of nodes in the inference tree model and initializing the zero matrix, and configuring the zero matrix according to the pointing relationship of each node in the inference tree model;

[0030] The zero matrix and multiple candidate sequences are input into a pre-constructed tree attention model, and the tree attention model is used to parallel verify each candidate sequence according to the pointing relationship of each node in the zero matrix to obtain the output probability distribution of each candidate sequence; the pointing relationship is used to indicate the attention allocation of the tree attention model when verifying the candidate sequences in parallel to obtain the output probability distribution of each candidate sequence.

[0031] In one embodiment, after obtaining the output sequence matching the data content of the input data according to the output probability distribution, the method further includes:

[0032] The current data matrix is ​​updated based on the output sequence.

[0033] In a second aspect, the present application further provides a data processing device based on a large language model, the device comprising:

[0034] A matrix matching module is used to obtain input data, match the input data with a plurality of pre-constructed data matrices, obtain a data matrix that matches the data type of the input data, and determine the matched data matrix as the current data matrix;

[0035] A model building module, used for building an inference tree model based on the input data and the current data matrix;

[0036] A sequence integration module, used for determining multiple candidate sequences based on the inference tree model;

[0037] The sequence verification module is used to verify the multiple candidate sequences in parallel, determine the output probability distribution of each candidate sequence, and obtain an output sequence that matches the data content of the input data according to the output probability distribution.

[0038] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned method when executing the computer program.

[0039] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described above are implemented.

[0040] In a fifth aspect, the present application also provides a computer program product, including a computer program, which implements the steps of the method described above when executed by a processor.

[0041] The above-mentioned data processing method, device, equipment and storage medium based on the large language model, when it is necessary to respond to the data processing needs of the user or terminal, obtains input data, matches the input data with multiple pre-constructed data matrices, obtains a data matrix that matches the data type of the input data, and determines the matched data matrix as the current data matrix, which can not only process different types of inputs, but also limit the processing range of the data to the same or similar range as the data type of the input data, reduce the amount of data processing, so as to reduce the calculation burden of the large language model and improve the data processing efficiency of the large language model; then, an inference tree model is constructed based on the input data and the current data matrix, and multiple candidate sequences are determined based on the inference tree model, and the multiple candidate sequences are verified in parallel to determine the output probability distribution of each candidate sequence, and an output sequence that matches the data content of the input data is obtained according to the output probability distribution. By processing the candidate sequences in parallel, the efficiency of data processing can be further improved to quickly respond to the generation of real-time data, and the candidate sequence is determined based on the matched current data matrix, so that the correlation between the output sequence and the input data can be improved, thereby improving the output quality and accuracy of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1An application environment diagram of a data processing method based on a large language model in one embodiment;

[0044] Figure 2 is a flow chart of a data processing method based on a large language model in one embodiment;

[0045] Figure 3 A schematic diagram of a process for constructing an inference tree model in one embodiment;

[0046] Figure 4 It is a flowchart of a data processing method based on a large language model in an application example;

[0047] Figure 5 is a structural block diagram of a data processing device based on a large language model in one embodiment;

[0048] Figure 6 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0050] The data processing method based on the large language model provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The server 104 obtains the data to be processed as input data, and matches the input data with multiple pre-built data matrices to obtain a data matrix that matches the input data. Subsequently, the server 104 constructs a step-by-step extended inference tree model based on the input data and the matched data matrix. Through the inference tree model, multiple candidate sequences that may match the input data can be determined, and then each candidate sequence is verified in parallel, and the output probability distribution of each candidate sequence is calculated. According to the output probability distribution, a sequence that matches the data content of the input data is determined as an output sequence, which is not only highly correlated with the input data, but also can improve the data processing efficiency. Among them, the terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car devices, projection devices, etc. Portable wearable devices can be smart watches, smart bracelets, etc. The server 104 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0051] In an exemplary embodiment, a data processing method based on a large language model is provided, wherein the method is applied to Figure 1 The server in the example is used for explanation. It is understandable that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. Figure 2 As shown, the method includes:

[0052] Step S201, obtaining input data, matching the input data with a plurality of pre-constructed data matrices, obtaining a data matrix matching the data type of the input data, and determining the matching data matrix as the current data matrix.

[0053] Among them, the input data is the data that needs to be processed by the large language model, usually text data, audio data, image and video data, structured data and other types of data.

[0054] Among them, the data matrix refers to a matrix used to store data of a specific data type. The data type can be the attribute, field, etc. of the data, which includes the transition probability, association relationship, etc. between the stored data. For example, for text data, the data stored in the data matrix may be the collocation information of vocabulary, and the vocabulary for different fields may be stored in different data matrices; for another example, for audio data, the data stored in the data matrix may be the gain coefficient of the data in the output channel, the sampling value in the time domain or frequency domain, etc.; the specific field can be science and technology, people, society, etc., and the corresponding data matrix contains the data corresponding to the field; for example, in the data matrix focusing on the field of science and technology news, it may contain commonly used words and phrases in science and technology news, and in the data matrix focusing on the field of legal text, it may contain legal terms and commonly used sentence patterns. In this way, each data matrix represents the data of a field or data attribute, which is conducive to improving the processing efficiency of the large language model for specific types of data, and processing input data based on the same type of data matrix can also improve the correlation between the output data and the input data.

[0055] Among them, matching means that the data type defined by the data matrix is ​​the same as or similar to the data type of the input data. Matching can be that the similarity with the data type of the input data is maximum or the similarity meets certain conditions, so that when the large language model performs data processing, it can process the matched data matrix to achieve the purpose of improving data processing.

[0056] The current data matrix refers to a matched data matrix that is identical or similar to the input data at least in terms of data attributes or domains, and serves as the processing basis for the large language model.

[0057] Exemplarily, when there is a need for data processing, the server obtains input data, and matches the data type of the input data with multiple pre-constructed data matrices one by one to determine a data matrix that matches the data type of the input data, and the data matrix is ​​the same or similar to the input data at least in data attributes or fields. Subsequently, the server determines the matching data matrix as the current data matrix for subsequent data processing.

[0058] Step S202, constructing an inference tree model based on the input data and the current data matrix.

[0059] Among them, the inference tree model refers to a tree model with multiple levels of nodes, each level of node represents a data, and the data of two adjacent layers of nodes have a corresponding association relationship; for example, for the collocation information in the text data, when the input data is "introduction to the large language model", when constructing the inference tree model, if the data of a certain node is "large language model", the collocations that may appear in its lower-level nodes are "algorithm", "learning", "training", etc., and by determining layer by layer, an inference tree model with multiple levels of nodes can be obtained.

[0060] Exemplarily, the server analyzes the input data according to the determined current data matrix, and successively selects data matching the input data from the data matrix to construct an inference tree model.

[0061] Step S203: determining multiple candidate sequences based on the inference tree model.

[0062] Among them, the candidate sequence refers to the sequence that may be used as the output of the input data. Usually, there are multiple candidate sequences, but there is generally only one sequence that is used as the output. By processing each branch of the inference tree model, multiple candidate sequences that are used as the output of the input data can be determined.

[0063] Exemplarily, the server determines multiple branches of the inference tree model based on the constructed inference tree model, and obtains candidate sequences based on the one-to-one correspondence between the branches.

[0064] Step S204, verifying multiple candidate sequences in parallel, determining the output probability distribution of each candidate sequence, and obtaining an output sequence matching the data content of the input data according to the output probability distribution.

[0065] Among them, parallel verification means that the sequence can simultaneously verify multiple candidate sequences during a forward propagation process to further improve the efficiency of data processing.

[0066] The output probability distribution refers to the probability that one of the multiple candidate sequences may be used as the output sequence given multiple candidate sequences and the current decoder state. Selecting the output sequence that matches the data content of the input data according to the output probability distribution can be to use the candidate sequence with the highest output probability distribution as the output sequence, or to use the candidate sequence whose output probability distribution satisfies certain conditions as the output sequence.

[0067] When verifying the candidate sequence and outputting the probability distribution, one or more implementations including but not limited to self-attention mechanism, tree attention mechanism, convolutional neural network, recurrent neural network and the like methods or models may be adopted.

[0068] Exemplarily, the server inputs multiple candidate sequences into a selected algorithm or model, and the multiple candidate sequences are verified in parallel through a forward propagation in the algorithm or model, and the output probability distribution of each candidate sequence is determined respectively through the selected algorithm or model, so as to select a candidate sequence that matches the data content of the input data as the output sequence according to the output probability distribution.

[0069] In the above-mentioned data processing method based on the large language model, when it is necessary to respond to the data processing needs of the user or terminal, by obtaining input data, matching the input data with multiple pre-constructed data matrices, obtaining a data matrix matching the data type of the input data, and determining the matching data matrix as the current data matrix, it is not only possible to process different types of inputs, but also possible to limit the processing range of the data to the same or similar range as the data type of the input data, reducing the amount of data processing, so as to reduce the calculation burden of the large language model and improve the data processing efficiency of the large language model; subsequently, an inference tree model is constructed based on the input data and the current data matrix, and multiple candidate sequences are determined based on the inference tree model, and the multiple candidate sequences are verified in parallel to determine the output probability distribution of each candidate sequence, and an output sequence matching the data content of the input data is obtained according to the output probability distribution. By processing the candidate sequences in parallel, the efficiency of data processing can be further improved to quickly respond to the generation of real-time data, and the candidate sequence is determined based on the matching current data matrix, so that the correlation between the output sequence and the input data can be improved, thereby improving the output quality and accuracy of the data.

[0070] In one embodiment, obtaining input data and matching the input data with a plurality of pre-constructed data matrices includes:

[0071] Load multiple data matrices; obtain input data, and input the input data into a trained matching network, so as to match the input data with the multiple data matrices respectively through the matching network.

[0072] Among them, the matching network refers to the ability to dynamically select a matching data matrix from multiple data matrices according to the data type of the input data; in specific implementation, the matching data matrix in this embodiment usually refers to the data matrix with the highest matching degree with the input data, so that the output of the large language model is strongly correlated with the input data. Of course, in some other embodiments, such as application scenarios with low accuracy requirements but high comprehensiveness requirements, several data matrices can also be determined from multiple data matrices according to the degree of matching as the current data matrix to ensure that the output content of the large language model is more comprehensive.

[0073] The matching network can be implemented using a reinforcement learning algorithm, which is a method that enables an agent to make optimal decisions through interactive learning with the environment. In specific implementation, the matching network can be implemented using one or more reinforcement learning algorithms including but not limited to Q learning, value iteration, policy gradient, multi-objective reinforcement learning, contextual bandit, etc.

[0074] Exemplarily, the server loads multiple pre-built data matrices from a data storage system and obtains input data; then, the input data is input into a trained matching network, and the input data is matched one by one with each loaded data matrix through the matching network to determine a data matrix that matches the data type of the input data.

[0075] In this embodiment, the input data is matched with multiple data matrices through a matching network constructed based on a reinforcement learning algorithm. The data matrix with the highest matching degree with the current input data can be dynamically selected according to the characteristics of the input data. While improving the data processing capability, it can also better adapt to different types of input data, thereby effectively improving the versatility and processing efficiency of the large language model.

[0076] In one embodiment, before inputting the input data into the pre-trained matching network, the method further includes:

[0077] Create multiple data matrices according to the data type of the sample data; initialize each data in the data matrix; train a pre-built matching network based on the initialized data matrix to obtain a trained matching network.

[0078] Among them, sample data refers to data belonging to various fields and various data attributes collected from various channels. The sample data has a data label, which can be the field label of the data or the attribute label of the type or attribute. When creating a data matrix, the sample data can be classified according to the data label by manual annotation, and data matrices can be constructed based on the classified sample data to obtain multiple data matrices; the sample data can also be clustered according to the data label of the sample data by a clustering algorithm to obtain multiple different data matrices; or a classification model can be trained by machine learning with the sample data and its data label as input to construct multiple different data matrices based on the classification results of the classification model.

[0079] Among them, the matching network is constructed based on the contextual gambling mechanism. Its agent program can select a data matrix from multiple data matrices according to the context information of the input data at each time step, and judge the expected value corresponding to the selection of the data matrix, so as to dynamically select the data matrix with the largest expected value among multiple data matrices as the current data matrix.

[0080] Exemplarily, the server collects a number of sample data and classifies the sample data according to data labels to obtain multiple data matrices; the data matrix contains the transition probability and correlation relationship between data in a certain field or a certain data type. Taking text data as an example, the data matrix contains commonly used vocabulary collocation information in a certain field, and the data matrix is ​​expressed as:

[0081] (1)

[0082] in, For the A data matrix, A vocabulary consisting of all sample data belonging to text data, which contains all possible tokens in various fields of text data (i.e., word units, the basic components of text data can be words, characters, phrases, etc.). is the number of Tokens in the vocabulary, that is, the size of the vocabulary, The number of candidates to keep for each Token.

[0083] Then, the server initializes each data (i.e., sample data) in the data matrix. During initialization, the server uses formula (2) to initialize the data matrix:

[0084] (2)

[0085] in, For the The first Line Column data (ie Token), is the conditional probability, Represents a dataset in a given field In the Appears in the vocabulary The probability of In the vocabulary Find the conditional probability of Token at maximum , in order to determine the most likely following of each Token in the corresponding field A subordinate token.

[0086] Finally, based on the initialized data matrix, the sample data and its corresponding data labels are input into the matching network constructed based on the context gambling machine to iteratively train the matching network to obtain a trained matching network. Specifically, the matching network is defined as , indicating that according to the input sample data features (such as fields, attributes, types, etc.), select the data matrix that matches the input data from all data matrices ; When the server trains the matching network, the training goal is to maximize the expected value, which is expressed as:

[0087] (3)

[0088] in, is the matching strategy parameter of the matching network, that is, the expected value corresponding to a data matrix selected by the matching network from multiple data matrices according to the context information of the input data. The sample data of the matching network for the current time step in the training phase (or the input data if in the application phase), is the data matrix selected at the current time step, For the sample data under the matching strategy of the matching network With data matrix The joint distribution of Based on sample data Select Data Matrix After the reward, According to the joint distribution Calculate the expectation.

[0089] In order to optimize the network parameters of the matching network, the server uses a gradient optimization strategy to adjust the network parameters so that the optimized matching network can accurately select a data matrix that matches the input data. The gradient optimization strategy is expressed as:

[0090] (4)

[0091] in, is the objective function About strategy parameters The gradient of Based on sample data Select Data Matrix The probability of , that is, the policy function, Based on sample data Select Data Matrix And follow the strategy Reward. According to the joint distribution Computed compliance strategy expectations.

[0092] In this embodiment, by training the matching network to dynamically select the data matrix that best suits the current input data, the relationship between different input types and the optimal matching strategy can be effectively captured, making the matching process more accurate and efficient.

[0093] In one embodiment, Figure 3 As shown, an inference tree model is constructed based on the input data and the current data matrix, including:

[0094] Step S301, determine the key field of the input data, and determine the key field as the current field.

[0095] Among them, the key field refers to the field that can represent the obvious characteristics of the input data. For example, in text data, the key field can be the most frequently appearing words in the input text data, the sentence that best represents the subject of the input text data, etc. When determining the key field, it is usually necessary to preprocess the input data. The preprocessing process includes but is not limited to the steps of word segmentation, stop word removal, stem extraction, filtering, etc., so as to accurately select the key field from the preprocessed data.

[0096] Exemplarily, the server preprocesses the input data to divide the input data into several fields, and determines the field with obvious characteristics of the input data as the key field by judging each field, and determines the key field as the current field.

[0097] Step S302, creating a root node of the inference tree model, and determining the root node as the current node.

[0098] The root node refers to the starting node of the inference tree model.

[0099] Exemplarily, the server creates a node as a root node of the inference tree model, and determines the root node as the current node.

[0100] Step S303, assigning the current field to the current node.

[0101] Among them, the current field refers to the field obtained at the current time step, and the current node refers to the node created at the current time step. Specifically, in the early stage of creating the inference tree model, the current field is the field corresponding to the root node, that is, the key field, and the current node is the root node; in each subsequent time step, the current field is a field selected from the current data matrix that has an associated relationship with the field of the parent node, and the current node is a child node created under the parent node.

[0102] Exemplarily, the server assigns the current field to the current node to use the selected current field as the value of the current node.

[0103] Step S304: perform correlation matching between the current field and the data in the current data matrix to determine a plurality of secondary fields matching the current field.

[0104] The relevance refers to whether a certain secondary field is similar or related to the current field.

[0105] Exemplarily, after the current field is determined and written into the current node, if the end condition has not been reached at this time, the server continues to select matching subsequent fields from the current data matrix as secondary fields for each current field. In this process, the server calculates the correlation between the current field and each data (Token) in the current data matrix, and selects data whose correlation meets the preset conditions as secondary fields. The correlation can be calculated using a word vector space model, Euclidean distance, cosine similarity, etc.

[0106] Step S305, according to the number of secondary fields, create child nodes corresponding to the secondary fields under the current node, determine the secondary field as the current field and the child node as the current node, and return to the step of assigning the current field to the current node to iterate until the end condition is reached to obtain the inference tree model.

[0107] The end condition may be reaching a predetermined number of layers, number of nodes, or a specific termination condition (such as no more secondary fields are available, reaching a set number of iterations, etc.).

[0108] Exemplarily, the server creates one-to-one child nodes under each corresponding current node according to the number of secondary fields selected from the current data matrix, determines the secondary field as the current field and the child node as the current node, returns to step S303 and loops through steps S303 to S305 until the end condition is met to obtain an inference tree model.

[0109] Among them, the inference tree model The collection of layer nodes is represented as:

[0110] (5)

[0111] in, The inference tree model A collection of layer nodes, For the The nodes of the layer, For the data matrix Select the previous Tokens, The first elements (when processing input data, it is usually processed into sequence data, The processed output data sequence is recorded as elements), is the depth of the inference tree model, that is, the number of layers of the inference tree model.

[0112] In this embodiment, by constructing an inference tree model to select data that matches the input data in terms of correlation from the current data matrix, it is possible to effectively split the current data matrix and simplify the inference tree model. Moreover, in the process of constructing the inference tree model, key information and patterns in the data can be quickly identified, which is conducive to further improving data processing efficiency.

[0113] In one embodiment, based on the inference tree model, multiple candidate sequences are determined, including:

[0114] Starting from the root node, each node of the inference tree model is traversed; when all current child nodes of a current node point to the same lower-level child node, the current child node and the lower-level child node are merged into a new child node, and the current node is pointed to the new child node, so as to merge and update each node of the inference tree model to obtain updated nodes; the updated nodes are respectively integrated in the traversal order to obtain multiple candidate sequences pointing from the root node to the lower-level child nodes step by step.

[0115] Among them, the current node is the root node or child node corresponding to the current traversal moment, and all current child nodes of the current node point to the same lower-level child node, which means that the current node is taken as the starting point, the current node is recorded as the Nth-level node, and all the next-level child nodes of the current node (that is, the N+1th-level node) are connected to the same next-level child node (that is, the N+2th-level node), that is, the fields of multiple N+1th-level nodes match the fields of the same N+2th-level node.

[0116] Among them, the new child node refers to merging all the above-mentioned multiple N+1-level nodes pointing to the same N+2-level node into one node as a new child node of the inference tree model, and updating the inference tree model based on the new child node to reduce the number of child nodes in the inference tree model, further reduce the amount of data processing, and improve processing efficiency.

[0117] The integration in the traversal order refers to connecting the updated nodes to the child nodes step by step from the root node downwards in the traversal order, and integrating them into a sequence in order according to the value (ie, field) of each node to obtain a candidate sequence.

[0118] Exemplarily, the server starts from the root node and traverses each node of the inference tree model layer by layer. If all current child nodes of a current node point to the same lower-level child node, then these current child nodes are merged into a new child node, and the current node directly points to the new child node, and then the new direct point directly points to the next child node, and the new child node is used to merge and update the inference tree model until the traversal is completed to obtain the updated nodes; then, the server connects the updated nodes to the child nodes layer by layer starting from the root node in the order of traversal, and integrates them into a sequence in order according to the value of each node (i.e., field) to obtain a candidate sequence.

[0119] In this embodiment, by merging the current sub-nodes, the redundancy of sub-nodes in the reasoning tree model can be reduced, thereby further simplifying the structure of the reasoning tree model and reducing the amount of calculation of the model during the reasoning process, which is conducive to further improving the response speed and processing efficiency of the model.

[0120] In one embodiment, multiple candidate sequences are verified in parallel, the output probability distribution of each candidate sequence is determined, and an output sequence matching the data content of the input data is obtained according to the output probability distribution, including:

[0121] A zero matrix with the same number of nodes as the inference tree model is created and initialized, and the zero matrix is ​​configured according to the pointing relationship of each node in the inference tree model; the zero matrix and multiple candidate sequences are input into a pre-built tree attention model, and each candidate sequence is verified in parallel according to the pointing relationship of each node in the zero matrix through the tree attention model to obtain the output probability distribution of each candidate sequence.

[0122] Here, a zero matrix refers to a matrix in which all elements are initialized to zero.

[0123] Among them, the pointing relationship of each node refers to the parent-child relationship or predecessor-successor relationship between the upper node and the lower node. The pointing relationship is used to indicate the attention allocation of the tree attention model when verifying the candidate sequences in parallel to obtain the output probability distribution of each candidate sequence.

[0124] The tree attention model refers to the ability to decompose the calculation of self-attention into multiple parallel tasks, each of which can process an input candidate sequence, thereby realizing parallel processing of multiple candidate sequences. Attention allocation means that when the tree attention model processes the input candidate sequence, it can dynamically focus on different candidate sequences and assign different attention weights according to the importance of the focused part.

[0125] Exemplarily, the server creates a zero matrix based on the updated inference tree model, where the number of elements in the zero matrix is ​​the same as the number of nodes in the updated inference tree model, and initializes the zero matrix, initializing all elements in the zero matrix to 0; then, for each pair of nodes (a, b) with a pointing relationship in the inference tree model, if node a is a child node of node b or node b is a parent node of node a, the server sets the corresponding position (a, b) or (b, a) in the zero matrix to 1 to determine the pointing relationship between each node; then, the server inputs each candidate sequence into the tree attention model, and passes the reset matrix as additional information to the tree attention model, so that the tree attention model can process each node in each candidate sequence in parallel according to the pointing relationship in the matrix, and calculate the output probability distribution of each candidate sequence according to the characteristics of each node in the candidate sequence and the pointing relationship between the nodes.

[0126] In this embodiment, by constructing a multi-layer inference tree structure and combining the merged nodes and the configured zero matrix, parallel processing can be performed on nodes with directed relationships, which can significantly improve the processing efficiency of the model. In addition, a large language model can verify multiple possible candidate sequences in parallel during the data generation process, which can effectively reduce processing delays and meet the needs of application scenarios that require real-time response.

[0127] In one embodiment, after obtaining the output sequence matching the data content of the input data according to the output probability distribution, the method further includes: updating the current data matrix based on the output sequence.

[0128] Exemplarily, after determining the output sequence, the server may also reversely transfer the output sequence to the current data matrix to continuously update each data matrix. The update formula for updating the current data matrix using the output sequence is as follows:

[0129] (6)

[0130] in, For the The output probability distribution of candidate sequences.

[0131] At the same time, each data matrix is ​​continuously updated based on the output sequence. Specifically, the update formula of each data matrix is ​​as follows:

[0132] (7)

[0133] in, is the learning rate for updating the data matrix, is the loss function to measure the performance of the matrix, is the gradient update strategy.

[0134] In addition, the server can also dynamically update the matching network based on the output sequence to continuously optimize and adjust the matching strategy of the matching network for the data matrix. The matching strategy is updated in the following way:

[0135] (8)

[0136] in, is the learning rate for updating the matching strategy.

[0137] In this embodiment, by recycling the generated output sequence into the data matrix, the data matrix can be adaptively updated, and by continuously updating the matching strategy, the data in the data matrix can dynamically adapt to new language patterns and expressions, especially when processing real-time dialogues or generation tasks, the latest language trends can be captured more accurately.

[0138] In an application example, Figure 4 As shown, a data processing method based on a large language model is provided. The method is based on context-aware expert selection and adaptive dynamic inference tree, which can accelerate the reasoning of the large language model. Among them, the method of this example involves the following structure:

[0139] Contextual Bandit: Contextual Bandit is a reinforcement learning problem where the agent chooses an action at each time step based on the current context and observes the corresponding reward. In this paper, it is used to model the expert matrix selection problem, enabling the system to dynamically select the optimal inference strategy based on the input features.

[0140] Expert Matrix: Expert Matrix is ​​a key data structure in this invention, which is used to store the token transition probability of a specific field or task. Each matrix represents the knowledge of an expert and contains the commonly used vocabulary collocation information in the field, which helps to improve the reasoning efficiency of specific types of input.

[0141] Tree Attention: Tree Attention is an innovative attention mechanism that allows the model to verify multiple possible token sequences in parallel in one forward propagation. By constructing a tree-structured attention mask, it significantly improves the parallelism and efficiency of the reasoning process.

[0142] Token Recycling: Token recycling is a knowledge updating mechanism proposed by this invention. It realizes the dynamic updating of the knowledge base by reintegrating the candidate tokens generated in the reasoning process into the expert matrix, so that the system can continuously adapt to new language patterns and expressions.

[0143] Token: Token is the basic processing unit in natural language processing, which is used to segment continuous text into discrete semantic segments. In a large language model, text is first converted into a series of tokens, which can be words, subwords or characters. This processing method simplifies text representation and improves the generalization ability and processing efficiency of the model.

[0144] Based on this, the method of this example includes the following steps:

[0145] Step 1: Preprocessing phase - building an expert knowledge base

[0146] 1. Initialization of expert matrix (ie data matrix)

[0147] Create multiple expert matrices for different task domains or input types. Each matrix All contain the collocation information commonly used in this field, including Glossary, is the number of candidates retained for each Token:

[0148] (9)

[0149] Each element of the matrix is ​​initialized by the following formula:

[0150] (10)

[0151] in, For the The expert matrix Line Column data (ie Token), is the conditional probability, Represents a dataset in a given field In the Appears in the vocabulary The probability of In the vocabulary Find the conditional probability of Token at maximum , in order to determine the most likely following of each Token in the corresponding field A subordinate token.

[0152] 2. Policy network (i.e. matching network) training

[0153] Training a policy network , which works based on the input Features, quickly select the most suitable expert matrix The training goal is to maximize the expected reward:

[0154] (11)

[0155] in, is the policy parameter of the policy network, that is, the expected value corresponding to a certain expert matrix selected by the policy network from multiple expert matrices according to the context information of the input data. is the input of the policy network, is the expert matrix selected at the current time step, For the policy network, under Policy, enter With Expert Matrix The joint distribution of Based on input Select Expert Matrix After the reward, According to the joint distribution Calculate the expectation.

[0156] Use the policy gradient method to optimize this network and continuously adjust the network parameters so that it can more accurately select the appropriate expert matrix:

[0157] (12)

[0158] in, is the objective function About strategy parameters The gradient of Based on input Select Expert Matrix The probability of , that is, the policy function, Based on input Select Expert Matrix And follow the strategy Reward. According to the joint distribution Computed compliance strategy expectations.

[0159] Step 2: Inference Phase - Dynamically Accelerate Token Generation

[0160] 1. Expert Matrix Selection

[0161] For input , select the expert matrix that can maximize the probability of the policy network output:

[0162] (13)

[0163] in, Output the expert matrix with the highest probability for the policy network.

[0164] 2. Build an inference tree

[0165] Build an inference tree. For depth and the current Token sequence , the inference tree The layer node set is defined as:

[0166] (14)

[0167] Select the most likely subsequent token from the expert matrix and build a multi-layer tree structure.

[0168] 3. Tree Attention Mechanism Verification

[0169] Inference Tree Compress to merge sequence And the corresponding tree attention mechanism , and then verify multiple candidate sequences simultaneously through a forward propagation:

[0170] (15)

[0171] in, For the The output probability distribution of candidate sequences.

[0172] 4. Select output and token recovery

[0173] Select the longest correct sequence as output and update the expert matrix with the newly generated candidate Token:

[0174] (16)

[0175] Step 3: Optimization Phase - Continuously Improve System Performance

[0176] 1. Dynamically update the policy network

[0177] According to the actual use effect, the selection strategy of the policy network is constantly adjusted:

[0178] (17)

[0179] in, is the learning rate of the policy network.

[0180] 2 Optimizing the Expert Matrix

[0181] At the same time, various expert matrices are continuously updated:

[0182] (18)

[0183] in, is the learning rate for updating the expert matrix, is the loss function to measure the performance of the matrix.

[0184] Based on the above, the beneficial effects of this example are:

[0185] 1. Context-aware expert matrix selection mechanism

[0186] The context-aware expert matrix selection mechanism in this example is an innovative large language model reasoning optimization technology that trains the policy network To dynamically select the expert matrix that best suits the current input, it effectively captures the relationship between different input types and the optimal reasoning strategy. This selection mechanism makes the reasoning process more accurate and efficient, especially when dealing with diverse input scenarios. Compared with the traditional fixed reasoning method, the context-aware selection mechanism improves the reasoning quality while better adapting to different types of inputs, thereby effectively improving the versatility and efficiency of the model.

[0187] 2. Adaptive knowledge update based on token recovery

[0188] In the reasoning stage, by analyzing the generated candidate tokens and recycling them into the expert matrix, the present invention realizes the adaptive update of knowledge by continuously updating the adjacency matrix. , which enables the expert knowledge base to dynamically adapt to new language patterns and expressions, especially when dealing with real-time dialogue or generation tasks, it can more accurately capture the latest language trends. Compared with traditional static knowledge bases, adaptive updates based on Token recycling provide a more flexible and real-time means of knowledge acquisition.

[0189] 3. Tree Attention Mechanism Reasoning Acceleration Algorithm

[0190] The tree attention mechanism inference acceleration algorithm is another core technology of the present invention. The algorithm significantly improves the inference efficiency of the model by constructing a multi-layer inference tree structure and combining compressed sequences and special attention masks, so that the large language model can verify multiple possible output sequences in parallel during the generation process, greatly reducing the inference delay. Compared with the traditional token-by-token generation method, the tree attention mechanism algorithm has significant advantages in improving the inference speed, and is particularly suitable for application scenarios that require real-time response.

[0191] This example not only significantly improves the model's adaptability to different inputs, but also greatly improves the reasoning speed by combining context-aware expert matrix selection, adaptive knowledge update based on token recycling, and tree attention mechanism reasoning acceleration algorithm, providing a new solution for the efficient deployment of large language models. The comprehensive application of these technical points enables the present invention to have significant technical advantages in improving the reasoning efficiency and adaptability of large language models, and is particularly suitable for application scenarios that require rapid response in a variety of input environments, such as real-time dialogue systems, dynamic content generation, etc. Through the above-mentioned reasoning framework, not only the reasoning speed bottleneck faced by large language models in practical applications is solved, but also a sustainable optimization mechanism is provided, so that the model can continuously adapt to new language patterns and expressions, thereby maintaining high efficiency and advancement in long-term use.

[0192] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0193] Based on the same inventive concept, the embodiment of the present application also provides a data processing device based on a large language model for implementing the data processing method based on a large language model involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the one or more embodiments of the data processing device based on a large language model provided below can refer to the limitations of the data processing method based on a large language model above, and will not be repeated here.

[0194] In an exemplary embodiment, Figure 5 As shown, a data processing device based on a large language model is provided, comprising: a matrix matching module 501, a model building module 502, a sequence integration module 503 and a sequence verification module 504, wherein:

[0195] The matrix matching module 501 is used to obtain input data, match the input data with a plurality of pre-constructed data matrices, obtain a data matrix that matches the data type of the input data, and determine the matched data matrix as the current data matrix.

[0196] The model building module 502 is used to build an inference tree model based on the input data and the current data matrix.

[0197] The sequence integration module 503 is used to determine multiple candidate sequences based on the inference tree model.

[0198] The sequence verification module 504 is used to verify multiple candidate sequences in parallel, determine the output probability distribution of each candidate sequence, and obtain an output sequence that matches the data content of the input data according to the output probability distribution.

[0199] In an optional embodiment, the matrix matching module 501 is further used to load multiple data matrices, obtain input data, and input the input data into a trained matching network, so as to match the input data with the multiple data matrices respectively through the matching network.

[0200] In an optional embodiment, the above-mentioned data processing device based on the large language model also includes a matrix construction module, which is used to create multiple data matrices according to the data type of the sample data; initialize each data in the data matrix; and train a pre-constructed matching network based on the initialized data matrix to obtain a trained matching network.

[0201] In an optional embodiment, the model building module 502 is also used to determine the key fields of the input data, and determine the key fields as the current fields; create the root node of the inference tree model, and determine the root node as the current node; assign the current field to the current node; match the current field with the data in the current data matrix for correlation, and determine multiple secondary fields that match the current field; according to the number of secondary fields, create child nodes corresponding to the secondary fields under the current node, determine the secondary fields as the current fields and the child nodes as the current node, and return to the step of assigning the current field to the current node to iterate until the end condition is reached to obtain the inference tree model.

[0202] In an optional embodiment, the sequence integration module 503 is also used to traverse each node of the inference tree model starting from the root node; when all current child nodes of a current node point to the same lower-level child node, the current child node and the lower-level child node are merged into a new child node, and the current node is pointed to the new child node, so as to merge and update each node of the inference tree model to obtain updated nodes; the current node is the root node or child node corresponding to the current traversal moment; the updated nodes are respectively integrated in the traversal order to obtain multiple candidate sequences pointing from the root node to the lower-level child nodes step by step.

[0203] In an optional embodiment, the sequence verification module 504 is also used to create a zero matrix with the same number of nodes as the inference tree model and initialize the zero matrix, and configure the zero matrix according to the pointing relationship between each node in the inference tree model; input the zero matrix and multiple candidate sequences into a pre-built tree attention model, and use the tree attention model to verify each candidate sequence in parallel according to the pointing relationship between each node in the zero matrix to obtain the output probability distribution of each candidate sequence; the pointing relationship is used to indicate the attention allocation of the tree attention model when verifying the candidate sequence in parallel to obtain the output probability distribution of each candidate sequence.

[0204] In an optional embodiment, the above-mentioned data processing device based on the large language model further includes an updating module, which is used to update the current data matrix based on the output sequence.

[0205] Each module in the above-mentioned data processing device based on the large language model can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0206] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 6 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as data matrices, matching networks, input data, and each node of the inference tree model. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a data processing method based on a large language model is implemented.

[0207] Those skilled in the art will understand that Figure 6The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0208] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the data processing method based on the large language model of the above embodiment is implemented.

[0209] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the data processing method based on a large language model of the above embodiment is implemented.

[0210] In one embodiment, a computer program product is provided, including a computer program, which, when executed by a processor, implements the data processing method based on a large language model of the above embodiment.

[0211] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0212] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0213] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0214] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A data processing method based on a large language model, characterized in that: The method comprises: Acquire input data, match the input data with a plurality of pre-constructed data matrices, obtain a data matrix that matches the data type of the input data, and determine the matched data matrix as a current data matrix; Constructing an inference tree model based on the input data and the current data matrix; Based on the inference tree model, determining multiple candidate sequences; The plurality of candidate sequences are verified in parallel to determine the output probability distribution of each candidate sequence, and an output sequence matching the data content of the input data is obtained according to the output probability distribution.

2. The method according to claim 1, characterized in that The obtaining of input data and matching the input data with a plurality of pre-constructed data matrices comprises: Load multiple data matrices; Input data is acquired, and the input data is input into a trained matching network, so that the input data is matched with the multiple data matrices respectively through the matching network.

3. The method according to claim 2, characterized in that Before inputting the input data into the pre-trained matching network, the method further includes: Create multiple data matrices based on the data type of the sample data; Initializing each data in the data matrix; A pre-built matching network is trained based on the initialized data matrix to obtain a trained matching network.

4. The method according to claim 1, characterized in that: The constructing of an inference tree model based on the input data and the current data matrix comprises: Determine a key field of the input data, and determine the key field as a current field; Creating a root node of the inference tree model, and determining the root node as the current node; Assigning the current field to the current node; Performing correlation matching between the current field and the data in the current data matrix to determine a plurality of secondary fields matching the current field; According to the number of the secondary fields, subnodes corresponding to the secondary fields are created under the current node, the secondary fields are determined as the current fields and the subnodes are determined as the current nodes, and the step of assigning the current fields to the current node is returned to iterate until the end condition is reached to obtain the inference tree model.

5. The method according to claim 4, characterized in that Determining a plurality of candidate sequences based on the inference tree model comprises: Taking the root node as the starting point, traverse each node of the inference tree model; When all current child nodes of a current node point to the same lower-level child node, the current child node and the lower-level child node are merged into a new child node, and the current node is pointed to the new child node, so as to merge and update each node of the inference tree model to obtain updated nodes; the current node is the root node or child node corresponding to the current traversal moment; The updated nodes are respectively integrated in the traversal order to obtain multiple candidate sequences pointing from the root node to the lower-level child nodes step by step.

6. The method according to claim 4, characterized in that The performing parallel verification on the plurality of candidate sequences, determining the output probability distribution of each candidate sequence, and obtaining an output sequence matching the data content of the input data according to the output probability distribution, comprises: Creating a zero matrix having the same number of nodes as the inference tree model and initializing the zero matrix, and configuring the zero matrix according to the pointing relationship of each node in the inference tree model; The zero matrix and multiple candidate sequences are input into a pre-constructed tree attention model, and the tree attention model is used to parallel verify each candidate sequence according to the pointing relationship of each node in the zero matrix to obtain the output probability distribution of each candidate sequence; the pointing relationship is used to indicate the attention allocation of the tree attention model when verifying the candidate sequences in parallel to obtain the output probability distribution of each candidate sequence.

7. The method according to any one of claims 1 to 6, characterized in that: After obtaining the output sequence matching the data content of the input data according to the output probability distribution, the method further includes: The current data matrix is ​​updated based on the output sequence.

8. A data processing device based on a large language model, characterized in that: The device comprises: A matrix matching module is used to obtain input data, match the input data with a plurality of pre-constructed data matrices, obtain a data matrix that matches the data type of the input data, and determine the matched data matrix as the current data matrix; A model building module, used for building an inference tree model based on the input data and the current data matrix; A sequence integration module, used to determine multiple candidate sequences based on the inference tree model; The sequence verification module is used to verify the multiple candidate sequences in parallel, determine the output probability distribution of each candidate sequence, and obtain an output sequence that matches the data content of the input data according to the output probability distribution.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images