Production portal information recommendation method and system based on large model driving

Abstract

The invention provides a production portal information recommendation method and system based on large model driving, and relates to the technical field of data processing. The method comprises the steps of obtaining current consultation data of a user; performing word segmentation processing on the current consultation data; constructing a word embedding vector sequence; analyzing a current consulting intention of a user through a large model based on a multi-layer stacked gating circulation unit; according to the current consultation intention, multiple pieces of production portal information matched with the current consultation data are determined through a multi-layer perceptron; obtaining a consultation intention abstract formed by the user in the historical consultation process, wherein the consultation intention abstract comprises a plurality of typical consultation intentions of the user in the historical consultation process; updating the consultation intention abstract of the user according to the similarity between the current consultation intention and each typical consultation intention in the consultation intention abstract; according to the updated consultation intention abstract, sorting the successfully matched multiple pieces of production portal information; and recommending the sorted multiple pieces of production portal information to the user.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a production portal information recommendation method and system based on large model driving. Background Art

[0002] In today's digital age, users are faced with a huge amount of information, and it has become very difficult to manually filter out valuable information. Currently, users can clarify their needs by typing keywords or inputting voice data. The recommendation system automatically filters and provides the most relevant information based on user needs, helping users save time and energy.

[0003] Current recommendation systems mainly use keyword matching to make recommendations. However, the same keyword may have different meanings in different contexts, and there are many contents with different expressions but similar meanings. The keyword-based recommendation method ignores the contextual relevance, which can easily lead to a large difference between the recommended content and customer needs, too single recommended content, low personalization, and failure to meet the diverse needs of users. Summary of the invention

[0004] In order to solve the technical problems that the prior art recommends based on keyword matching, ignores context relevance, easily leads to a large difference between recommended content and customer needs, the recommended content is too single, the degree of personalization is low, and cannot meet the diverse needs of users, the present invention provides a production portal information recommendation method and system based on large model drive.

[0005] The technical solution provided by the embodiment of the present invention is as follows:

[0006] First aspect

[0007] An embodiment of the present invention provides a production portal information recommendation method based on large model driving, comprising:

[0008] S1: Get the user's current consultation data;

[0009] S2: performing word segmentation processing on the current consultation data;

[0010] S3: construct word embedding vectors for each word segment and form a word embedding vector sequence;

[0011] S4: Analyze the user's current consultation intention according to the word embedding vector sequence through a large model based on multi-layer stacked gated recurrent units;

[0012] S5: According to the current consultation intention, determine multiple production portal information matching the current consultation data through a multi-layer perceptron;

[0013] S6: Obtain a consultation intention summary formed by the user in the historical consultation process, where the consultation intention summary includes multiple typical consultation intentions of the user in the historical consultation process;

[0014] S7: updating the user's consultation intention summary according to the similarity between the current consultation intention and each typical consultation intention in the consultation intention summary;

[0015] S8: sorting and processing the multiple successfully matched production portal information according to the updated consultation intention summary;

[0016] S9: Recommending multiple sorted production portal information to the user.

[0017] Second aspect

[0018] An embodiment of the present invention provides a production portal information recommendation system based on large model drive, including:

[0019] processor;

[0020] A memory having computer-readable instructions stored thereon, wherein when the computer-readable instructions are executed by the processor, the method for recommending production portal information based on large model driving as described in the first aspect is implemented.

[0021] The third aspect

[0022] An embodiment of the present invention provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method for recommending production portal information based on large model driving as described in the first aspect is implemented.

[0023] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0024] In the present invention, a large model based on multi-layer stacked gated recurrent units is used to effectively capture the user's contextual information and analyze the user's deep consulting intention. At the same time, the user's consulting intention summary is continuously updated to dynamically capture changes in user interests. A multi-layer perceptron is used to perform multi-level feature extraction and matching processing on the current consulting intention, which helps to find the production portal information that best meets the user's needs, improve the accuracy and quality of the recommended content, and make the recommendation results more in line with the user's long-term preferences and have a higher degree of personalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0026] Figure 1 A flowchart of a production portal information recommendation method based on large model drive provided by an embodiment of the present invention;

[0027] Figure 2 A schematic diagram of the structure of a production portal information recommendation method based on large model drive provided by an embodiment of the present invention;

[0028] Figure 3 A schematic diagram of the structure of a production portal information recommendation system driven by a large model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0030] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either of the two.

[0031] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same. "of", "corresponding, relevant" and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same.

[0032] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0033] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0034] Reference Manual Attached Figure 1 , showing a flow chart of a production portal information recommendation method based on large model drive provided in an embodiment of the present invention.

[0035] Reference Manual Attached Figure 2 , showing a structural schematic diagram of a production portal information recommendation method based on large model drive provided by an embodiment of the present invention.

[0036] The embodiment of the present invention provides a production portal information recommendation method based on a large model drive, which can be implemented by a production portal information recommendation device based on a large model drive, and the production portal information recommendation device based on a large model drive can be a terminal or a server. The processing flow of the production portal information recommendation method based on a large model drive can include the following steps:

[0037] S1: Get the user's current consultation data.

[0038] Alternatively, the consultation data may be text data typed by a user or voice data for input.

[0039] Furthermore, the voice data can be converted into text data through a voice recognition algorithm.

[0040] S2: Perform word segmentation on the current consultation data.

[0041] In a possible implementation, S2 specifically includes: performing word segmentation processing on the current consultation data through regular expressions, maximum matching method, bidirectional maximum matching method, hidden Markov model and / or conditional random field.

[0042] Among them, regular expression is a tool for matching strings. By defining character patterns, specific text fragments can be found and processed. It can be used to identify word boundaries, punctuation marks, numbers, etc., so as to perform word segmentation.

[0043] Among them, the maximum matching method is a word segmentation method based on a dictionary, also known as the forward maximum matching method. It starts from the beginning of the text and tries to match the longest word. In each step, it starts from the current position and tries to match the longest word possible until it can no longer match.

[0044] Among them, the bidirectional maximum matching method combines the forward maximum matching method and the reverse maximum matching method (matching from right to left), and improves the accuracy of word segmentation by executing these two methods at the same time and comparing the results. If the word segmentation results obtained by the two methods are different, the result with fewer word segmentations is usually selected, or a decision is made based on other strategies.

[0045] Among them, the hidden Markov model is a probabilistic model that generates an observation sequence (words) through a state sequence (usually part of speech) and handles the word segmentation problem through statistical learning. When segmenting words, each word is regarded as an observation state, and the optimal word segmentation result is determined by maximizing the overall probability.

[0046] Among them, conditional random field is an undirected graph model, which is often used in sequence labeling tasks, such as word segmentation. CRF predicts the label of each character (such as the beginning, middle or end of a word) given the context by learning the interdependence between features, thereby achieving word segmentation.

[0047] It should be noted that the above word segmentation methods have their own advantages and disadvantages. In practical applications, the appropriate method is often selected according to specific needs and scenarios. Regular expressions and maximum matching methods are suitable for simple scenarios, while hidden Markov models and conditional random fields are suitable for complex scenarios that require high-precision word segmentation. The bidirectional maximum matching method can achieve a certain compromise effect in both simple and complex scenarios.

[0048] S3: Construct word embedding vectors for each word segment and form a word embedding vector sequence.

[0049] In a possible implementation, S3 specifically includes: constructing word embedding vectors for each word segment through Word2Vec, GloVe and / or FastText, and forming a word embedding vector sequence.

[0050] Among them, Word2Vec is a word embedding technology based on neural network. The core idea is to predict the target word (Skip-gram) through the context word or to predict the context word (CBOW, Continuous Bag of Words) through the target word, so as to learn the vector representation of words.

[0051] Among them, GloVe is a global word embedding method. Different from the local context prediction of Word2Vec, GloVe generates word embeddings by factoring based on the global co-occurrence matrix. GloVe uses global statistical information to better capture the co-occurrence relationship of words.

[0052] Among them, FastText is different from Word2Vec and GloVe. FastText decomposes words into character n-grams to capture the morphological information within the words, enabling FastText to generate more fine-grained word embeddings and process unregistered words.

[0053] In practical applications, you can choose the appropriate word embedding technology according to specific needs, or combine these methods to achieve the best effect.

[0054] S4: According to the word embedding vector sequence, the user’s current consultation intention is analyzed through a large model based on multi-layer stacked gated recurrent units.

[0055] Among them, the Gated Recurrent Unit (GRU) is an improved recurrent neural network (RNN) architecture that aims to solve the gradient vanishing and gradient exploding problems that occur in standard RNNs when processing long sequence data. The basic unit of GRU contains two gates: the reset gate and the update gate, which control how information flows at each time step. GRU retains the ability to remember long-term dependencies by simplifying the gating mechanism while reducing computational complexity.

[0056] Optionally, the multi-layer stacked gated recurrent unit includes a stacked multi-layer gated recurrent unit. The input layer of the first layer of gated recurrent units is used to obtain a weighted word embedding vector sequence, the candidate layer of the previous layer of gated recurrent units is connected to the input layer of the next layer of gated recurrent units, and the output layers of the gated recurrent units of each layer are connected to a fully connected layer, and the fully connected layer is used to fuse the hidden states in the candidate layers of the gated recurrent units of each layer to obtain a comprehensive hidden state representing the consultation intention.

[0057] It should be noted that by stacking multiple layers of GRU units, each layer can further abstract and process data based on the features extracted by the previous layer, thereby capturing more complex temporal dependencies and semantic information. The first layer may focus on local, short-term dependencies, while subsequent layers can capture dependencies with longer time spans and higher-level semantic features. By fusing the hidden states of each layer of GRU in the fully connected layer, the model can integrate features at different levels and avoid excessive influence of features at a single level on the final result. This can enhance the robustness and generalization ability of the model, making the model more stable when processing input sequences of different types and lengths.

[0058] In a possible implementation, S4 specifically includes sub-steps S401 to S403:

[0059] S401: Input the word embedding vector sequence into the first layer gated recurrent unit to obtain the hidden state of the first layer gated recurrent unit:

[0060]

[0061]

[0062]

[0063]

[0064] Among them, z t represents the output vector of the update gate at time t, σ() represents the activation function, and W uz represents the weight matrix between the input layer and the update gate, x trepresents the input data at time t (weighted word embedding vector sequence), W hz represents the self-connection weight matrix between the update gate at time t and time t-1, represents the hidden state of the first layer of gated recurrent unit at time t-1, b z represents the bias term of the update gate, r t represents the output vector of the reset gate at time t, W ur represents the weight matrix between the input layer and the reset gate, W hr represents the self-connection weight matrix between the reset gate at time t and time t-1, b r represents the bias term of the reset gate, c t represents the output vector of the candidate layer at time t, W uc Represents the weight matrix between the input layer and the candidate layer, W hc represents the self-connection weight matrix between the candidate layer at time t and time t-1, b c represents the bias term of the candidate layer, ⊙ represents the element-wise product operation, Represents the hidden state of the first layer of gated recurrent unit at time t.

[0065] It should be noted that by resetting the gate and updating the gate, the model can effectively selectively remember or forget past information, thereby enhancing the model's ability to remember important information.

[0066] S402: Input the hidden state of the i-1th gated recurrent unit into the i-th gated recurrent unit to obtain the hidden state of the i-th gated recurrent unit:

[0067]

[0068]

[0069]

[0070]

[0071] in, represents the hidden state of the i-1th layer gated recurrent unit at time t-1, represents the hidden state of the i-th layer gated recurrent unit at time t-1, represents the hidden state of the i-th layer gated recurrent unit at time t, i ≥ 2.

[0072] It should be noted that in a multi-layer GRU, each layer processes the output of the previous layer, thereby accumulating multi-level information. This approach can better capture short-term and long-term dependencies in time series, enhance the model's expressiveness, and enable the model to better comprehensively consider various information when facing complex input data.

[0073] S403: The hidden states in the candidate layers of the gated recurrent units of each layer are fused through the fully connected layer to obtain a comprehensive hidden state to analyze the user's current consultation intention:

[0074]

[0075] Among them, H represents the comprehensive latent state that represents the current consultation intention, β i Represents the fusion coefficient of the i-th layer gated recurrent unit.

[0076] It should be noted that the GRU hidden state is updated gradually over time and can accumulate and reflect dynamic changes in the input sequence. Therefore, the comprehensive hidden state contains a dynamic reflection of the entire sequence of user input and can capture the changes and trends expressed by the user throughout the conversation, which is crucial for understanding the current consultation intention.

[0077] In the present invention, the features and information captured by the GRU hidden states at different levels may be different. The low-level GRU may capture more basic and local features, while the high-level GRU may capture more abstract and global features. By fusing these information from different levels, the features of different levels can be comprehensively considered, thereby forming a more comprehensive and rich comprehensive hidden state.

[0078] S5: According to the current consultation intention, multiple production portal information matching the current consultation data are determined through a multi-layer perceptron.

[0079] Among them, Multi-Layer Perceptron (MLP) is a basic and widely used feedforward neural network model, which consists of multiple layers of neurons (or nodes), usually including input layer, hidden layer and output layer. The neurons in each layer are fully connected with the neurons in the next layer, thus forming a nonlinear mapping that can be used for classification, regression and other prediction tasks.

[0080] In a possible implementation, S5 specifically includes sub-steps S501 to S506:

[0081] S501: Encode each piece of production portal information using a one-hot encoding method to form a production portal information feature vector.

[0082] Among them, One-Hot Encoding is a commonly used encoding technology, which is mainly used to convert categorical data into numerical data that can be processed by machine learning models.

[0083] It should be noted that one-hot encoding generates a unique feature vector for each piece of production portal information. This encoding method is simple and effective, and can ensure that all production portal information can be uniquely identified, especially when there are many categories of data, it can avoid interference between features.

[0084] S502: Map the production portal information feature vector to a latent factor vector through a latent factor matrix:

[0085]

[0086] Among them, q j represents the potential factor vector corresponding to the j-th production portal information, Q j represents the potential factor matrix corresponding to the j-th production portal information, T represents the transpose operation of the matrix, v j Represents the j-th production portal information feature vector.

[0087] It should be noted that mapping the high-dimensional one-hot encoding vector to the latent factor space through the latent factor matrix can effectively reduce the dimension of the data and capture the latent semantic relationship of the production portal information. The latent factor vector retains more useful information in a lower dimension, which helps to better characterize the characteristics of the production portal information.

[0088] S503: The latent factor vector corresponding to each piece of production portal information is concatenated with the comprehensive latent state used to characterize the current consultation intention to obtain a concatenated vector:

[0089] p j =H,q j

[0090] Among them, p j represents the concatenated vector corresponding to the j-th production portal information, and H represents the comprehensive latent state that represents the current consultation intention.

[0091] It should be noted that concatenating the comprehensive latent state of the current consultation intention with the latent factor vectors of each production portal information helps to combine the semantic information of user needs with the feature information of production portal information. This concatenated vector can more comprehensively express the correlation between user needs and portal information, providing a stronger foundation for subsequent matching calculations.

[0092] S504: Determine the interaction vector between each production portal information and the current consultation intention through a multi-layer perceptron:

[0093] z 1,j =ReLU(W1p j +b1)

[0094] z 2,j=ReLU(W2z 1,j +b2)

[0095]

[0096] z l,j =ReLU(W l z l-1,j +b l )

[0097] Among them, z l,j represents the interaction vector between the j-th production portal information and the current consultation intention, z l-1,j represents the interaction vector between the j-th production portal information and the current consultation intention, ReLU represents the ReLU activation function, and W l represents the weight matrix of the lth layer, b l Represents the bias term of the lth layer, l = 1, 2,, L, and L represents the total number of layers of the multilayer perceptron.

[0098] It should be noted that a multi-layer perceptron (MLP) is used to process the concatenated vector, and the complex relationship between the user's consulting intention and the production portal information is extracted layer by layer through nonlinear transformations of multiple hidden layers. MLP can learn deeper feature interactions, especially when dealing with nonlinear relationships. At the same time, the ReLU activation function can effectively prevent the gradient vanishing problem, while providing the model with nonlinear capabilities, allowing the model to fit more complex data patterns. Ultimately, the output vector (interaction vector) of the MLP can accurately reflect the degree of match between the user's intention and the portal information.

[0099] S505: Calculate the matching degree between each production portal information and the current consultation intention according to the interaction vector between each production portal information and the current consultation intention:

[0100] r j =Sigmoid(W r z l,j +b r )

[0101] Among them, r j represents the matching degree between the j-th production portal information and the current consultation intention, Sigmoid represents the Sigmoid activation function, and W r represents the matching layer weight matrix, b r represents the matching layer bias term.

[0102] It should be noted that the matching degree is obtained by inputting the interaction vector into the matching layer and mapping the output value to [0,1] using the Sigmoid function. This step compresses the high-dimensional interaction information into a matching degree score, which helps to directly compare the degree of fit between different production portal information and user intentions.

[0103] S506: Sort each piece of production portal information in descending order of matching degree, and select a preset number of production portal information with top rankings.

[0104] Among them, those skilled in the art can set the size of the preset number according to actual conditions, and the present invention does not limit it. Optionally, the preset number is specifically 10.

[0105] In the present invention, not only can the user's current consultation intention be effectively combined with the characteristics of the production portal information, but also the complex feature interaction can be captured through the deep learning model, and finally accurate personalized recommendations can be achieved. This design can significantly improve the accuracy of the recommendation system and user experience.

[0106] S6: Obtain a summary of the user's consultation intentions formed during the historical consultation process.

[0107] Among them, the consultation intention summary contains multiple typical consultation intentions of users in the historical consultation process.

[0108] It should be noted that the user's historical consulting behavior can reflect their long-term interests and preferences. By summarizing these behaviors as typical consulting intentions, the system can more accurately understand the user's long-term interests and use them as the basis for recommendations and matching. This method allows the recommendation system to not only be based on the user's current behavior, but also make full use of the user's past behavior data to provide more targeted recommendations.

[0109] S7: Update the user's consulting intention summary according to the similarity between the current consulting intention and each typical consulting intention in the consulting intention summary.

[0110] In a possible implementation, S7 specifically includes sub-steps S701 and S702:

[0111] S701: Calculate the cosine similarity between the current consultation intention and each typical consultation intention in the consultation intention summary:

[0112]

[0113] Among them, sim k represents the cosine similarity between the kth typical consulting intention in the consulting intention summary and the current consulting intention, H represents the comprehensive latent state that characterizes the current consulting intention, and H kRepresents the comprehensive hidden state that characterizes the k-th typical consultation intention and represents the vector modulus operation.

[0114] S702: When the cosine similarity between a certain target typical consulting intention and the current consulting intention is lower than a preset similarity, the target typical consulting intention is replaced with the current consulting intention, and the consulting intention summary is updated.

[0115] When the cosine similarities between multiple typical consulting intentions and the current consulting intention are lower than the preset similarity, the current consulting intention is used to replace the typical consulting intention with the lowest similarity, and the consulting intention summary is updated.

[0116] It should be noted that replacement is only made when the current inquiry intent has a low similarity to some old intents. This approach ensures that the intent retained in the summary is always the most representative of the user's interests. When the user's interests diversify or change, this strategy enables the system to respond in a timely manner to avoid old, irrelevant intents taking up space, thereby maintaining the representativeness of the summary. Through this update mechanism, the system can gradually remove content that the user is no longer interested in, reduce the risk of recommending irrelevant content, and thus improve the accuracy of the recommendation results. Doing so can enable the recommendation system to better match the user's actual needs and enhance the user experience.

[0117] In the present invention, by calculating the similarity between the current consultation intention and the historical typical consultation intention, and replacing the old intention with lower similarity with the current intention, the system can dynamically update the user's interest model. This means that the user's historical interests will not remain unchanged, but will be constantly adjusted with new behaviors and consultations, so that the system can more accurately reflect the user's current needs and preferences. This enhances the accuracy and relevance of recommendations, while controlling the size of the summary to maintain computational efficiency.

[0118] S8: According to the updated consultation intent summary, the multiple successfully matched production portal information are sorted.

[0119] In a possible implementation, S8 specifically includes sub-steps S801 to S803:

[0120] S801: Calculate the matching degree between the successfully matched production portal information and each updated typical consulting intention through a multi-layer perceptron.

[0121] Specifically, how to calculate the matching degree through the multi-layer perceptron can refer to the contents of S503 to S505 above, and to avoid repetition, the present invention will not go into details.

[0122] S802: Calculate the average matching degree between the successfully matched production portal information and the updated consultation intent summary:

[0123]

[0124] Among them, R j represents the average matching degree between the jth production portal information that is successfully matched and the updated consultation intent summary, r kj It represents the matching degree between the successfully matched j-th production portal information and the updated k-th typical consulting intent, and K represents the total number of typical consulting intents contained in the consulting intent summary.

[0125] S803: Sort each production portal information in descending order according to the average matching degree.

[0126] In the present invention, the portal information is sorted in descending order according to the average matching degree, ensuring that the recommended content seen by the user is the most relevant and best meets their current needs. This sorting method significantly improves user satisfaction and the effectiveness of the recommendation system by giving priority to the most matching content.

[0127] S9: Recommending multiple sorted production portal information to the user.

[0128] It should be noted that by sorting by average matching degree, the system can ensure that users first see the production portal information that is most relevant to their current consultation intentions and historical interests. This approach can maximize the satisfaction of user needs and improve the accuracy and effectiveness of recommendations.

[0129] In a possible implementation, the training method of a large model based on multi-layer stacked gated recurrent units specifically includes:

[0130] Construct the mean square error loss function:

[0131]

[0132] Among them, f represents the mean square error loss function, r j represents the matching degree between the j-th production portal information calculated by the model and the current consultation intention, It represents the actual matching degree between the j-th production portal information and the current consultation intention, and m represents the total number of production portal information samples.

[0133] According to the mean square error loss function, the Adam optimizer is used to adaptively adjust the learning rate and acceleration parameters to train a large model based on multi-layer stacked gated recurrent units:

[0134]

[0135]

[0136]

[0137]

[0138]

[0139] m t =β1m t-1 +(1-β1)g t

[0140] Among them, θ t represents the model parameters at the tth iteration, θ t-1 represents the model parameters at the t-1th iteration, α represents the basic learning rate, represents the second-order moment estimate after the correction of the deviation at the t-th iteration, ε represents the hyperparameter for numerical stability, represents the first-order moment estimate after the correction of the deviation at the tth iteration, v t represents the second-order moment estimate at the tth iteration, β2 represents the second-order moment estimate attenuation coefficient, and v t-1 represents the second-order moment estimate at the t-1th iteration, g t represents the gradient at the tth iteration, ▽ represents the gradient operation, f represents the loss function, ▽ θ f t (θ t-1 ) indicates that the loss function performs gradient calculation on the model parameters at the tth iteration, m t represents the first-order moment estimate at the t-th iteration, β1 represents the first-order moment estimate attenuation coefficient, m t-1 Represents the first-order moment estimate at the t-1th iteration.

[0141] It should be noted that the Adam optimizer adaptively adjusts the learning rate of each parameter based on the historical gradient information of each parameter. For parameters with drastic gradient changes, Adam will reduce the learning rate, and for parameters with gentle gradient changes, it will increase the learning rate. This adaptability allows the model to converge faster while reducing training problems caused by improper learning rate selection.

[0142] In the present invention, by using the Adam optimizer, the model training process can adaptively adjust the learning rate and acceleration parameters, effectively handle sparse gradients, improve numerical stability, accelerate convergence, and reduce sensitivity to hyperparameters. This optimization method is particularly suitable for complex multi-layer stacked GRUs and multi-layer perceptrons, which can improve the training efficiency, stability and generalization ability of the model, and ultimately achieve better prediction performance.

[0143] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0144] In the present invention, a large model based on multi-layer stacked gated recurrent units is used to effectively capture the user's contextual information and analyze the user's deep consulting intention. At the same time, the user's consulting intention summary is continuously updated to dynamically capture changes in user interests. A multi-layer perceptron is used to perform multi-level feature extraction and matching processing on the current consulting intention, which helps to find the production portal information that best meets the user's needs, improve the accuracy and quality of the recommended content, and make the recommendation results more in line with the user's long-term preferences and have a higher degree of personalization.

[0145] Reference Manual Attached Figure 3 , showing a structural schematic diagram of a production portal information recommendation system based on large model drive provided by the present invention.

[0146] The present invention further provides a large model driven production portal information recommendation system 20, which is applied to the above-mentioned large model driven production portal information recommendation method, comprising:

[0147] Processor 201.

[0148] The memory 202 stores computer-readable instructions. When the computer-readable instructions are executed by the processor 201, the production portal information recommendation method based on large model driving as in the method embodiment is implemented.

[0149] The large model-driven production portal information recommendation system 20 provided in the present invention can execute the above-mentioned large model-driven production portal information recommendation method and achieve the same or similar technical effects. To avoid repetition, the present invention will not go into details.

[0150] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0151] In the present invention, a large model based on multi-layer stacked gated recurrent units is used to effectively capture the user's contextual information and analyze the user's deep consulting intention. At the same time, the user's consulting intention summary is continuously updated to dynamically capture changes in user interests. A multi-layer perceptron is used to perform multi-level feature extraction and matching processing on the current consulting intention, which helps to find the production portal information that best meets the user's needs, improve the accuracy and quality of the recommended content, and make the recommendation results more in line with the user's long-term preferences and have a higher degree of personalization.

[0152] It should be understood that the processor in the embodiment of the present invention may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0153] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0154] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0155] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0156] In the present invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can be represented by: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0157] It should be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0158] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0159] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0160] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0161] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0162] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0163] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0164] An embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the method for recommending production portal information based on large model driving as described in the method embodiment is implemented.

[0165] A computer-readable storage medium provided by the present invention can implement the steps and effects of the large model-driven production portal information recommendation method of the above method embodiment. To avoid repetition, the present invention will not go into details.

[0166] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0167] In the present invention, a large model based on multi-layer stacked gated recurrent units is used to effectively capture the user's contextual information and analyze the user's deep consulting intention. At the same time, the user's consulting intention summary is continuously updated to dynamically capture changes in user interests. A multi-layer perceptron is used to perform multi-level feature extraction and matching processing on the current consulting intention, which helps to find the production portal information that best meets the user's needs, improve the accuracy and quality of the recommended content, and make the recommendation results more in line with the user's long-term preferences and have a higher degree of personalization.

[0168] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

[0169] There are a few points to note:

[0170] (1) The drawings of the embodiments of the present invention only relate to the structures related to the embodiments of the present invention, and other structures may refer to the general design.

[0171] (2) For the sake of clarity, in the drawings used to describe the embodiments of the present invention, the thickness of the layers or regions is exaggerated or reduced, that is, these drawings are not drawn according to the actual scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, the element may be "directly" "on" or "under" the other element or there may be intermediate elements.

[0172] (3) In the absence of conflict, the embodiments of the present invention and the features therein may be combined with each other to obtain new embodiments.

[0173] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A production portal information recommendation method based on large model drive, characterized in that: include: S1: Get the user's current consultation data; S2: performing word segmentation processing on the current consultation data; S3: construct word embedding vectors for each word segment and form a word embedding vector sequence; S4: Analyze the user's current consultation intention according to the word embedding vector sequence through a large model based on multi-layer stacked gated recurrent units; S5: According to the current consultation intention, determine multiple production portal information matching the current consultation data through a multi-layer perceptron; S6: Obtain a consultation intention summary formed by the user in the historical consultation process, where the consultation intention summary includes multiple typical consultation intentions of the user in the historical consultation process; S7: updating the user's consultation intention summary according to the similarity between the current consultation intention and each typical consultation intention in the consultation intention summary; S8: sorting and processing the multiple successfully matched production portal information according to the updated consultation intention summary; S9: Recommending multiple sorted production portal information to the user.

2. The method for recommending production portal information based on large model drive according to claim 1, characterized in that: The S2 is specifically: The current consultation data is segmented by regular expression, maximum matching method, bidirectional maximum matching method, hidden Markov model and / or conditional random field.

3. The method for recommending production portal information based on large model drive according to claim 1, characterized in that: The S3 is specifically: Build word embedding vectors for each word segment through Word2Vec, GloVe and / or FastText, and form a word embedding vector sequence.

4. The method for recommending production portal information based on large model drive according to claim 1, characterized in that: The multi-layer stacked gated recurrent unit includes a multi-layer gated recurrent unit in a stacked arrangement; the input layer of the first layer of gated recurrent units is used to obtain a weighted word embedding vector sequence, the candidate layer of the previous layer of gated recurrent units is connected to the input layer of the next layer of gated recurrent units, and the output layers of each layer of gated recurrent units are connected to a fully connected layer, and the fully connected layer is used to fuse the hidden states in the candidate layers of each layer of gated recurrent units to obtain a comprehensive hidden state that represents the consulting intention.

5. The method for recommending production portal information based on large model drive according to claim 4 is characterized in that: The S4 specifically includes: S401: Input the word embedding vector sequence into the first-layer gated recurrent unit to obtain the hidden state of the first-layer gated recurrent unit: Among them, z t represents the output vector of the update gate at time t, σ() represents the activation function, and W uz represents the weight matrix between the input layer and the update gate, x t represents the input data at time t (weighted word embedding vector sequence), W hz represents the self-connection weight matrix between the update gate at time t and time t-1, represents the hidden state of the first layer of gated recurrent unit at time t-1, b z represents the bias term of the update gate, r t represents the output vector of the reset gate at time t, W ur represents the weight matrix between the input layer and the reset gate, W hr represents the self-connection weight matrix between the reset gate at time t and time t-1, b r represents the bias term of the reset gate, c t represents the output vector of the candidate layer at time t, W uc Represents the weight matrix between the input layer and the candidate layer, W hc represents the self-connection weight matrix between the candidate layer at time t and time t-1, b c represents the bias term of the candidate layer, ⊙ represents the element-wise product operation, Represents the hidden state of the first layer of gated recurrent unit at time t; S402: Input the hidden state of the i-1th gated recurrent unit into the i-th gated recurrent unit to obtain the hidden state of the i-th gated recurrent unit: in, represents the hidden state of the i-1th layer gated recurrent unit at time t-1, represents the hidden state of the i-th layer gated recurrent unit at time t-1, represents the hidden state of the i-th layer gated recurrent unit at time t, i ≥ 2; S403: The hidden states in the candidate layers of the gated recurrent units of each layer are fused through the fully connected layer to obtain a comprehensive hidden state to analyze the user's current consultation intention: Among them, H represents the comprehensive latent state that represents the current consultation intention, β i Represents the fusion coefficient of the i-th layer gated recurrent unit.

6. The method for recommending production portal information based on large model drive according to claim 5 is characterized in that: The S5 specifically includes: S501: Encode each piece of production portal information using a one-hot encoding method to form a production portal information feature vector; S502: Mapping the production portal information feature vector to a latent factor vector through a latent factor matrix: Among them, q j represents the potential factor vector corresponding to the j-th production portal information, Q j represents the potential factor matrix corresponding to the j-th production portal information, T represents the transpose operation of the matrix, v j represents the feature vector of the j-th production portal information; S503: The latent factor vector corresponding to each piece of production portal information is concatenated with the comprehensive latent state used to characterize the current consultation intention to obtain a concatenated vector: p j =<H,q j Among them, p j represents the concatenated vector corresponding to the j-th production portal information, and H represents the comprehensive latent state representing the current consultation intention; S504: Determine the interaction vector between each production portal information and the current consultation intention through a multi-layer perceptron: Among them, z l,j represents the interaction vector between the j-th production portal information and the current consultation intention, z l-1,j represents the interaction vector between the j-th production portal information and the current consultation intention, ReLU represents the ReLU activation function, and W l represents the weight matrix of the lth layer, b l represents the bias term of the lth layer, l = 1, 2, ..., L, L represents the total number of layers of the multilayer perceptron; S505: Calculate the matching degree between each production portal information and the current consultation intention according to the interaction vector between each production portal information and the current consultation intention: r j =Sigmoid(W r z l,j +b r ) Among them, r j represents the matching degree between the j-th production portal information and the current consultation intention, Sigmoid represents the Sigmoid activation function, and W r represents the matching layer weight matrix, b r represents the matching layer bias term; S506: Sort each piece of production portal information in descending order of matching degree, and select a preset number of production portal information with top rankings.

7. The method for recommending production portal information based on large model drive according to claim 1, characterized in that: The S7 specifically includes: S701: Calculate the cosine similarity between the current consultation intention and each typical consultation intention in the consultation intention summary: Among them, sim k represents the cosine similarity between the kth typical consulting intention in the consulting intention summary and the current consulting intention, H represents the comprehensive latent state that characterizes the current consulting intention, and H k represents the comprehensive hidden state representing the kth typical consultation intention, || || represents the vector modulus operation; S702: When the cosine similarity between a certain target typical consulting intention and the current consulting intention is lower than a preset similarity, the target typical consulting intention is replaced with the current consulting intention, and the consulting intention summary is updated; When the cosine similarities between multiple typical consulting intentions and the current consulting intention are lower than a preset similarity, the current consulting intention is used to replace the typical consulting intention with the lowest similarity, and the consulting intention summary is updated.

8. The method for recommending production portal information based on large model drive according to claim 1, characterized in that: The S8 specifically includes: S801: Calculate the matching degree between the successfully matched production portal information and each updated typical consultation intention through a multi-layer perceptron; S802: Calculate the average matching degree between the successfully matched production portal information and the updated consultation intent summary: Among them, R j represents the average matching degree between the jth production portal information that is successfully matched and the updated consultation intent summary, r kj represents the matching degree between the j-th production portal information that has been successfully matched and the k-th typical consulting intent after update, and K represents the total number of typical consulting intents contained in the consulting intent summary; S803: Sort each production portal information in descending order according to the average matching degree.

9. The method for recommending production portal information based on large model drive according to claim 1, characterized in that: The training method of the large model based on multi-layer stacked gated recurrent units specifically includes: Construct mean square error loss function; According to the mean square error loss function, the Adam optimizer is used to adaptively adjust the learning rate and acceleration parameters to train the large model based on multi-layer stacked gated recurrent units.

10. A production portal information recommendation system based on large model drive, characterized in that: include: processor; A memory having computer-readable instructions stored thereon, wherein when the computer-readable instructions are executed by the processor, the method for recommending production portal information based on large model driving according to any one of claims 1 to 6 is implemented.

Images