Computing power resource session response processing method and system applied to deep learning

Abstract

The invention discloses a computing power resource session response processing method and system applied to deep learning. According to the method and the device, target session interaction knowledge identification and information set construction are carried out on the computing power resource session data, so that key information in the session can be accurately extracted, and interference of irrelevant information is avoided. According to the method, the associated session interaction knowledge information set of the computing power resource business keyword is coded and subjected to cross collision to obtain the initial associated session interaction knowledge vector, so that the knowledge connotation related to the business keyword can be deeply mined, and the understanding ability of knowledge in different business scenes is improved. The target associated session interaction knowledge vector is determined according to the correlation characteristic value, and the most relevant knowledge resources can be accurately matched to improve the response accuracy. The session response result obtained according to the determined session interaction service theme keyword can quickly and accurately respond to the computing power resource session, so that the session processing efficiency is effectively improved, the response quality is improved, and the management and utilization of computing power resources are optimized.

Description

Technical Field

[0001] The present application belongs to the field of data analysis technology, and specifically relates to a method and system for processing session responses of computing resources applied to deep learning. Background Art

[0002] In the current computing resource management field, with the widespread application of computing resources, session processing faces many challenges. Traditional session response methods often cannot effectively extract key information from complex session data, resulting in an inaccurate understanding of computing resource business needs. For example, when an enterprise allocates computing resources, a user's session may contain a variety of demand information, but traditional methods are difficult to accurately identify and associate it with specific business keywords. Moreover, the lack of means to deeply mine knowledge related to business keywords results in the inability to fully utilize existing knowledge resources when answering, which greatly reduces the accuracy and comprehensiveness of the answer. Furthermore, existing technologies are inefficient in associating and matching session knowledge with business keywords, resulting in slow session response speeds and affecting the management efficiency of computing resources. Summary of the invention

[0003] The present application provides a method and system for processing session responses of computing resources applied to deep learning, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0004] The present application provides a computing power resource session response processing method applied to deep learning, which is applied to a session response processing system. The method includes: obtaining computing power resource session data to be processed, identifying target session interaction knowledge in the computing power resource session data to be processed, and obtaining a target session interaction knowledge information set; mining a current session interaction knowledge vector corresponding to the target session interaction knowledge information set, and obtaining an initial associated session interaction knowledge vector corresponding to each computing power resource business keyword, wherein the initial associated session interaction knowledge vector is obtained by respectively encoding each associated session interaction knowledge information set corresponding to the computing power resource business keyword, and obtaining each associated session interaction knowledge code; and The computing power resource session response result corresponding to the computing power resource session data to be processed is determined according to the computing power resource session data subject keyword of the target session interaction knowledge information set, and the computing power resource session response result corresponding to the computing power resource session data to be processed is determined according to the session interaction business subject keyword of the target session interaction knowledge information set.

[0005] The present application provides a conversation response processing system, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.

[0006] The present application provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0007] This application can accurately extract key information in the conversation and avoid interference from irrelevant information by identifying the target conversation interaction knowledge and constructing an information set for the computing resource conversation data. The initial associated conversation interaction knowledge vector is obtained by encoding and cross-colliding the associated conversation interaction knowledge information set of the computing resource business keywords, which can deeply mine the knowledge connotation related to the business keywords and improve the ability to understand knowledge in different business scenarios. By determining the target associated conversation interaction knowledge vector based on the correlation eigenvalue, the most relevant knowledge resources can be accurately matched to improve the accuracy of the response. The conversation response results derived based on the determined conversation interaction business theme keywords can quickly and accurately respond to the computing resource conversation, effectively improving the conversation processing efficiency, while improving the quality of the response, thereby optimizing the management and utilization of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A flowchart of a method for processing computing power resource session responses for deep learning provided in this application.

[0009] Figure 2 A structural diagram of a conversation response processing system provided in this application. DETAILED DESCRIPTION

[0010] The following will be combined with the drawings in this application to clearly and completely describe the technical solutions in this application. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0011] The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are usually a class, and the number of objects is not limited. For example, the first object can be one or more. In addition, in this application, "and / or" represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0012] Figure 1 A method for processing session responses of computing resources applied to deep learning is shown, which is applied to a session response processing system. The method includes the following steps 202-208.

[0013] Step 202: Acquire the computing power resource session data to be processed, identify the target session interaction knowledge in the computing power resource session data to be processed, and obtain the target session interaction knowledge information set.

[0014] Step 204 , mining the current session interaction knowledge vector corresponding to the target session interaction knowledge information set, and obtaining the initial associated session interaction knowledge vector corresponding to each computing resource business keyword.

[0015] In the present application, the initial associated session interaction knowledge vector is obtained by respectively encoding each associated session interaction knowledge information set corresponding to the computing resource business keyword into a knowledge vector to obtain each associated session interaction knowledge encoding, and then cross-colliding each associated session interaction knowledge encoding.

[0016] Step 206: Determine a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector from the associated session interaction knowledge vectors corresponding to the computing power resource business keywords based on the correlation feature values ​​between the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the computing power resource business keywords.

[0017] Step 208: Use the computing power resource business keywords corresponding to the target-associated session interaction knowledge vector as the session interaction business subject keywords of the corresponding target session interaction knowledge information set, and determine the computing power resource session response result corresponding to the computing power resource session data to be processed based on the session interaction business subject keywords of the target session interaction knowledge information set.

[0018] In the present application, an exemplary introduction to the technical solutions described in the above steps 202 to 208 is described as follows.

[0019] In step 202, the source of the computing resource session data to be processed is wide, and may come from the interactive interface between the user and the computing resource service provider, such as the user's consultation session on the cloud computing service platform, or the communication records of internal employees of the enterprise on the allocation and use of computing resources. These data can be in text form, containing various questions, descriptions and requests related to computing resources. For example, the data may be "Our department needs to carry out large-scale data mining tasks. How much computing resources are needed?" or "The computing power of the GPU server suddenly dropped during the deep learning model training. How to find out the cause?" and so on.

[0020] When acquiring data, you need to ensure the integrity and accuracy of the data. For data from different channels, you may need to unify the format and perform preliminary cleaning operations. For example, unify texts in different encoding formats into UTF-8 encoding and remove garbled characters and irrelevant special characters.

[0021] In order to identify the target conversation interaction knowledge, the knowledge extraction method in natural language processing (NLP) technology can be used. For example, a rule-based method can be used to predefine some grammatical rules and keyword patterns related to computing resources. Taking the identification of knowledge about computing resource capacity as an example, the rule can be "if there are keywords such as 'need', 'computing resources', 'how much', followed by numbers or capacity units (such as GB, TFLOPS, etc.), then the knowledge extracted is about the computing resource capacity demand."

[0022] You can also use classification algorithms in machine learning, such as support vector machines (SVMs), to classify and identify knowledge in conversation data. First, you need to build a labeled training set that contains a large number of labeled computing resource conversation data samples, such as labeling statements about computing resource performance as one category and statements about computing resource costs as another category. Then, use this training set to train the SVM model, input the conversation data to be processed into the trained SVM model, and obtain classification results, thereby identifying different types of target conversation interaction knowledge.

[0023] After identification, the relevant knowledge is organized into a target conversation interaction knowledge information set. For example, an information set may contain {“computing resource requirement type: large-scale data mining task”, “involved equipment: GPU server”, “problem type: computing resource requirement assessment”}, etc.

[0024] Next, in step 204, in order to mine the current conversation interaction knowledge vector corresponding to the target conversation interaction knowledge information set, a word vector model, such as Word2Vec or GloVe, can be used. Taking Word2Vec as an example, each word in the target conversation interaction knowledge information set is first mapped to a low-dimensional vector space. For example, if the vocabulary size is (V) and the vector dimension is (d) (e.g., (d=300)), then each word is represented as a (d)-dimensional vector.

[0025] For the target conversational interaction knowledge information set, the vocabulary vectors are combined. An exemplary combination method is to find the average vector. For example, if the information set contains three words (w_1), (w_2), (w_3), and their corresponding word vectors are (vec{v}_1), (vec{v}_2), (vec{v}_3), then the current conversational interaction knowledge vector (vec{v}_{current}) can be expressed as:

[0026] (vec{v}_{current}=frac{vec{v}_1+vec{v}_2+vec{v}_3}{3}).

[0027] For each computing power resource business keyword, we first need to determine its associated conversation interaction knowledge information set. For example, for the business keyword "GPU computing power", its associated conversation interaction knowledge information set may include all knowledge content in historical conversation data related to GPU computing power, such as "the advantages of GPU computing power in image rendering" and "methods to improve GPU computing power".

[0028] Then, these associated conversation interaction knowledge information sets are encoded as knowledge vectors. For example, a neural network-based encoding method, such as a multi-layer perceptron (MLP), is used. For each associated conversation interaction knowledge information set, the vocabulary sequence therein is input into an MLP with multiple hidden layers (e.g., 3 hidden layers, with 128, 64, and 32 neurons in each layer, respectively), to obtain an associated conversation interaction knowledge encoding of a fixed length (e.g., 128 dimensions).

[0029] Finally, each associated conversation interaction knowledge encoding is cross-collided to obtain the initial associated conversation interaction knowledge vector. Cross-collision can be performed using methods such as vector dot product or cosine similarity calculation. For example, for two associated conversation interaction knowledge encodings (vec{e}_1) and (vec{e}_2), their cross-collision results (taking cosine similarity as an example) are: (cosine(vec{e}_1, vec{e}_2)

[0030] =frac{vec{e}_1cdotvec{e}_2}{vertvec{e}_1vertvertvec{e}_2vert}), and the initial associated conversation interaction knowledge vector is obtained by comprehensively processing multiple such results.

[0031] In step 206, it involves calculating the correlation feature values ​​of the current session interaction knowledge vector and the associated session interaction knowledge vector corresponding to each computing resource business keyword. A variety of methods can be used here, such as the Pearson correlation coefficient or a similarity calculation method based on deep learning. Take the Pearson correlation coefficient as an example:

[0032] For the current session interaction knowledge vector (vec{v}_{current}) and the associated session interaction knowledge vector (vec{v}_{associated}) corresponding to a computing resource business keyword, the calculation formula of the Pearson correlation coefficient (r) is: (r=frac{sum_{i=1}^{n}(x_i-bar{x})(y_i-bar{y})}{sqrt{sum_{i=1}^{n}(x_i-bar{x})^2sum_{i=1}^{n}(y_i-bar{y})^2}}), where (x_i) and (y_i) are the (i)th elements of (vec{v}_{current}) and (vec{v}_{associated}), respectively, (bar{x}) and (bar{y}) are their means, respectively, and (n) is the dimension of the vector.

[0033] For example, the current conversation interaction knowledge vector (vec{v}_{current}) is a 100-dimensional vector, and a certain associated conversation interaction knowledge vector (vec{v}_{associated}) is also a 100-dimensional vector. The Pearson correlation coefficient between them calculated by the above formula is 0.8, and this 0.8 is the correlation eigenvalue between them.

[0034] Next, the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector can be determined from the associated session interaction knowledge vectors corresponding to the business keywords of each computing resource according to the calculated correlation feature value. A threshold can be set, for example, an associated session interaction knowledge vector with a correlation feature value greater than 0.7 is determined as the target associated session interaction knowledge vector. If there are multiple associated session interaction knowledge vectors that meet this condition, the vector with the highest correlation feature value may be selected as the target associated session interaction knowledge vector.

[0035] Finally, in step 208, the computing power resource business keyword corresponding to the target-associated session interaction knowledge vector is used as the session interaction business subject keyword of the corresponding target session interaction knowledge information set. For example, if the computing power resource business keyword corresponding to the target-associated session interaction knowledge vector is "GPU computing power optimization", then "GPU computing power optimization" is the session interaction business subject keyword of the target session interaction knowledge information set.

[0036] The computing resource session response result corresponding to the computing resource session data to be processed is determined according to the session interaction service subject keywords of the target session interaction knowledge information set. A knowledge base query or a template-based response generation method can be used.

[0037] For the knowledge base query method, the system maintains a knowledge base containing a large amount of computing power resource-related knowledge, and uses the conversation interaction service subject keyword as an index to search for the corresponding answer content in the knowledge base. For example, if the conversation interaction service subject keyword is "GPU computing power optimization", the knowledge base may store answers such as "GPU computing power can be improved by optimizing algorithms, upgrading drivers, and increasing parallel computing."

[0038] For the template-based response generation method, some response templates corresponding to different conversation interaction business topic keywords are pre-defined. For example, for the topic keyword "GPU computing power optimization", the template may be "For your questions about GPU computing power optimization, here are some suggestions: [List of optimization measures]", and then fill in the [List of optimization measures] part in the template according to the specific knowledge content, such as "For your questions about GPU computing power optimization, here are some suggestions: optimize the algorithm structure, upgrade to the latest GPU driver, reasonably set the number of parallel computing threads, etc.", so as to obtain the computing power resource conversation response result.

[0039] Through the above four main steps, namely, obtaining the computing power resource session data to be processed, mining related knowledge vectors, determining the target-related session interaction knowledge vectors, and finally obtaining the computing power resource session response results, an effective computing power resource session response processing system is realized. This system helps to improve the interaction efficiency in computing power resource management, enhance user experience, and can accurately handle various session requirements related to computing power resources.

[0040] In summary, this application first identifies the target conversation interaction knowledge and constructs an information set for the computing power resource conversation data, so as to accurately extract the key information in the conversation and avoid the interference of irrelevant information. In the process of mining knowledge vectors, the associated conversation interaction knowledge information set of the computing power resource business keywords is encoded and cross-collided to obtain the initial associated conversation interaction knowledge vector. This method can deeply mine the knowledge connotation related to the business keywords and improve the ability to understand knowledge in different business scenarios. By determining the target associated conversation interaction knowledge vector based on the correlation eigenvalue, the most relevant knowledge resources can be accurately matched, thereby improving the accuracy of the response. Finally, the conversation response result is obtained based on the determined conversation interaction business theme keywords, which can quickly and accurately respond to the computing power resource conversation, effectively improve the conversation processing efficiency, reduce processing time, and at the same time improve the quality of the response, thereby optimizing the management and utilization of computing power resources.

[0041] In the actual application process, before obtaining the computing power resource session data to be processed, identifying the target session interaction knowledge in the computing power resource session data to be processed, and obtaining the target session interaction knowledge information set, it also includes: obtaining each associated session interaction knowledge information set corresponding to the set computing power resource business keyword; encoding each associated session interaction knowledge information set by knowledge vector to obtain each associated session interaction knowledge encoding; cross-collision of each associated session interaction knowledge encoding to obtain the initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword.

[0042] On the basis of the above content, the cross-collision of the various associated session interaction knowledge codes to obtain the initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword includes: combining the various associated session interaction knowledge codes to obtain a session interaction knowledge combination code; performing full connection processing based on the session interaction knowledge combination code to obtain the initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword.

[0043] It is understandable that in the entire computing resource management and conversation response system, the computing resource business keywords need to be clearly defined first. These keywords are words or phrases closely related to the computing resource business, such as "GPU computing power optimization", "CPU resource allocation strategy", "data center computing power load balancing", etc. These keywords are the basis for the entire system to classify and process computing resource knowledge.

[0044] For each set computing power resource business keyword, it is necessary to obtain the corresponding associated session interaction knowledge information set. This process involves mining and sorting a large amount of historical session data. For example, for the keyword "GPU computing power optimization", the associated session interaction knowledge information set may include all historical session content related to GPU computing power optimization. These contents may include questions raised by users on how to improve the efficiency of GPU computing power in specific tasks (such as deep learning model training), and the corresponding answers given by technical personnel. These information sets can be obtained from the company's knowledge base, customer service records, or previous technical communication documents. During the acquisition process, it is necessary to ensure the integrity and accuracy of the information, and data cleaning operations may be required to remove erroneous information, incomplete records, or irrelevant information.

[0045] In order to convert the associated conversation interaction knowledge information set into a computable and analyzable form, a knowledge vector encoding method is used. An exemplary method is an encoding method based on a neural network, for example, a multi-layer perceptron (MLP) can be used for encoding.

[0046] For example, the associated conversation interaction knowledge information set can be represented as a sequence of words (S = {w_1, w_2, cdots, w_n}), where (w_i) represents the (i)th word in the information set. First, each word needs to be mapped to a low-dimensional vector space. You can use a pre-trained word vector model such as Word2Vec or GloVe. For example, if the vocabulary size is (V) and the word vector dimension is (d) (for example, (d = 300)), then each word (w_i) is represented as a (d)-dimensional vector (vec{v}_{w_i}).

[0047] Then, these word vectors are input into an MLP with multiple hidden layers in the order of the words in the information set. For example, the MLP has 3 hidden layers, and the number of neurons in each layer is (h_1=128), (h_2=64), and (h_3=32). The number of neurons in the input layer is equal to the word vector dimension (d), and the number of neurons in the output layer is also a fixed value (e) (for example, (e=128)). For the input word vector sequence, a fixed length (e) dimensional vector is obtained through forward propagation calculation as the encoding of the associated conversation interaction knowledge.

[0048] For the calculation from the input layer to the first hidden layer, for example, the input vector is (vec{x}) (here (vec{x}) is a vector sequence composed of vocabulary vectors), the weight matrix of the first hidden layer is (W_1), and the bias vector is (vec{b}_1), then the output of the first hidden layer (vec{h}_1) is: (vec{h}_1=f(W_1vec{x}+vec{b}_1)), where (f) is the activation function, such as the ReLU function (f(x)=max(0,x)).

[0049] Similarly, for the second hidden layer, the input is (vec{h}_1), the weight matrix is ​​(W_2), the bias vector is (vec{b}_2), and the output (vec{h}_2) is: (vec{h}_2=f(W_2vec{h}_1+vec{b}_2)).

[0050] For the third hidden layer, the input is (vec{h}_2), the weight matrix is ​​(W_3), the bias vector is (vec{b}_3), and the output (vec{h}_3) is: (vec{h}_3=f(W_3vec{h}_2+vec{b}_3)).

[0051] Finally, the output of the output layer (vec{y}) (i.e., the associated conversation interaction knowledge encoding) is: (vec{y}=f(W_4vec{h}_3+vec{b}_4)), where (W_4) is the weight matrix of the output layer and (vec{b}_4) is the bias vector.

[0052] Furthermore, each associated session interaction knowledge code is cross-collided to obtain an initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword: for example, for a set computing power resource business keyword, (m) associated session interaction knowledge codes (vec{e}_1, vec{e}_2, cdots, vec{e}_m) are obtained. An exemplary combination method is to splice these vectors to obtain a session interaction knowledge combination code (vec{C}). For example, if each associated session interaction knowledge code is a 128-dimensional vector, then the dimension of the spliced ​​session interaction knowledge combination code (vec{C}) is (128m) dimensions.

[0053] A fully connected neural network is used to process the combined encoding of conversational interaction knowledge. For example, the input of the push discriminant branch is (vec{C}), the weight matrix is ​​(W_f), the bias vector is (vec{b}_f), and the output is the initial associated conversational interaction knowledge vector (vec{v}). The calculation method of the push discriminant branch is: (vec{v}=f(W_fvec{C}+vec{b}_f)), where (f) can still be a ReLU activation function. Here, the dimension of the weight matrix (W_f) depends on the dimension of the input vector (vec{C}) and the dimension requirement of the output vector (vec{v}). For example, if (vec{C}) is (128m) dimensional and the output vector (vec{v}) is required to be 64 dimensional, then the dimension of (W_f) is (64*128m).

[0054] With this design, first of all, by obtaining in advance the various associated conversation interaction knowledge information sets corresponding to the set computing power resource business keywords, a comprehensive knowledge basis is provided for the subsequent conversation response processing. In the process of knowledge vector encoding, the use of neural network methods such as multi-layer perceptrons can effectively convert complex associated conversation interaction knowledge information sets into knowledge encodings with fixed dimensions, which is convenient for subsequent calculations and analysis. The initial associated conversation interaction knowledge vector is obtained by combining and fully connecting the various associated conversation interaction knowledge encodings. This method can fully integrate the information in different associated conversation interaction knowledge encodings, so that the initial associated conversation interaction knowledge vector can more comprehensively reflect the knowledge features related to the set computing power resource business keywords. This helps to more accurately match the target conversation interaction knowledge in subsequent conversation interactions, improve the accuracy and efficiency of computing power resource conversation responses, and optimize the conversation processing flow in the entire computing power resource management.

[0055] On the basis of obtaining each associated session interaction knowledge information set corresponding to the set computing power resource business keyword; performing knowledge vector encoding on the each associated session interaction knowledge information set to obtain each associated session interaction knowledge code; cross-colliding the each associated session interaction knowledge code to obtain the initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword, the method also includes: loading the each associated session interaction knowledge information set into the initial AI knowledge embedding network; performing knowledge vector encoding on the each associated session interaction knowledge information set through the knowledge vector encoding module in the initial AI knowledge embedding network to obtain each associated session interaction knowledge code; cross-colliding the each associated session interaction knowledge code through the knowledge cross-collision module in the initial AI knowledge embedding network to obtain the initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword.

[0056] Based on the technical solution of the above-mentioned initial AI knowledge embedding network, the debugging method of the initial AI knowledge embedding network includes: obtaining each computing power resource session data sample and the corresponding prior conversation response training annotation, and loading the each computing power resource session data sample into the original deep learning network; performing knowledge vector encoding on each computing power resource session data sample through the initial AI knowledge embedding network in the original deep learning network to obtain each computing power resource session knowledge vector sample; performing cross-collision on each computing power resource session knowledge vector sample through the initial knowledge cross-collision module in the original deep learning network to obtain a cross-collision sample vector; outputting the initial conversation response in the original deep learning network The module performs a conversation response output on the cross-collision sample vector to obtain a conversation response prediction result; optimizes the original deep learning network according to the prior conversation response training annotations and the conversation response prediction results to obtain an intermediate deep learning network; uses the intermediate deep learning network as the original deep learning network, and jumps to the steps of obtaining each computing power resource conversation data sample and the corresponding prior conversation response training annotations, and loading each computing power resource conversation data sample into the original deep learning network, and implements the steps in a loop until the debugging process is completed to obtain the target deep learning network; obtains the initial AI knowledge embedding network according to the target knowledge vector encoding module and the target knowledge cross-collision module in the target deep learning network.

[0057] Based on the previously acquired knowledge information sets of each associated session interaction corresponding to the set computing resource business keywords, these information sets are loaded into the initial AI knowledge embedding network. This initial AI knowledge embedding network is a deep learning network structure specially built for processing computing resource session related knowledge.

[0058] The knowledge vector encoding module in the initial AI knowledge embedding network is responsible for encoding the knowledge vectors of each associated conversation interaction knowledge information set. A variety of deep learning algorithms can be used here, for example, an encoding method based on a convolutional neural network (CNN).

[0059] For example, the associated conversation interaction knowledge information set is represented as a two-dimensional matrix (M), where the rows of the matrix represent different vocabulary sequences (for example, each vocabulary sequence corresponds to a sentence), and the columns of the matrix represent the index of the vocabulary in the vocabulary. Let the vocabulary size be (V) and the maximum length of each vocabulary sequence be (L). For CNN encoding, multiple convolution kernels are used for convolution operations. Let the convolution kernel size be (k*d) (for example, (k=3), (d) is the word vector dimension, for example, (d=300)), and the step size be (s) (for example, (s=1)).

[0060] For a convolution kernel (w), the formula for convolution operation on matrix (M) is: (C_{i, j} = (w*M)_{i, j} = sum_{m = 0}^{k-1}sum_{n = 0}^{d-1}w_{m, n}M_{i+m, j+n}), where (C) is the convolution result matrix. After convolution operation with multiple convolution kernels, the data dimension is reduced by pooling operation (such as maximum pooling) to obtain a fixed-length knowledge vector encoding. For example, by using the maximum pooling operation, the maximum value is taken in a local area, and finally the knowledge encoding of each associated session interaction is obtained.

[0061] Next, the knowledge cross collision module in the initial AI knowledge embedding network cross-collides the obtained associated conversation interaction knowledge encodings. For example, if (m) associated conversation interaction knowledge encodings (vec{e}_1, vec{e}_2, cdots, vec{e}_m) are obtained, a method based on matrix multiplication and nonlinear transformation can be used.

[0062] First, construct a matrix (E = [vec{e}_1, vec{e}_2, cdots, vec{e}_m]), and then perform matrix multiplication (E*E^T) to obtain an (m*m) matrix (P), where (P_{ij} = vec{e}_icdotvec{e}_j) (here (cdot) represents the vector dot product). Next, perform a nonlinear transformation on each element of the matrix (P), for example, through the sigmoid function (f(x) = frac{1}{1+e^{-x}}), to obtain the transformed matrix (P'). Finally, combine the elements of the matrix (P') (such as averaging or weighted summing) to obtain the initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword.

[0063] Based on the above content, the debugging method of embedding initial AI knowledge into the network includes the following contents.

[0064] Obtaining data samples and training annotations: First, obtain data samples of each computing resource session and the corresponding prior conversation response training annotations. These data samples are representative samples extracted from the actual computing resource session history data, such as conversation contents such as different types of computing resource demand inquiries and troubleshooting. Prior conversation response training annotations are the correct responses corresponding to these data samples, provided by experts or verified standard answers.

[0065] Knowledge vector encoding to obtain session knowledge vector samples: Each computing resource session data sample is loaded into the initial AI knowledge embedding network in the original deep learning network. These data samples are knowledge vector encoded by the knowledge vector encoding module in the initial AI knowledge embedding network. Similar to the encoding method mentioned above, a specific algorithm is used here to convert each session data sample into a knowledge vector form. For example, for a session data sample sentence containing (n) words, after encoding, a computing resource session knowledge vector sample (vec{v}_{s}) of a fixed length (for example, (d'=128)) is obtained.

[0066] Cross-collision to obtain cross-collision sample vectors: Use the initial knowledge cross-collision module in the original deep learning network to cross-collision each computing resource session knowledge vector sample. For example, (p) computing resource session knowledge vector samples (vec{v}_{s1}, vec{v}_{s2}, cdots, vec{v}_{sp}) are obtained, and the cross-collision sample vector (vec{v}_{c}) is obtained according to the cross-collision method mentioned above (such as based on matrix multiplication and nonlinear transformation).

[0067] Conversation response output to obtain prediction results: The cross-collision sample vector is output through the initial conversation response output module in the original deep learning network. This module can be a classifier or generator based on a multi-layer perceptron (MLP). For example, if it is a classifier, for (q) possible response types (such as response categories for different types of computing resource problems), the input of the MLP is the cross-collision sample vector (vec{v}_{c}), and the output is a (q)-dimensional vector (vec{y}) representing the probability of each response type. The softmax function (y_i=frac{e^{z_i}}{sum_{j=1}^{q}e^{z_j}}) (where (z_i) is the unnormalized value of the output of the last layer of the MLP) can be used to convert the output into a probability form, and the category with the highest probability is the conversation response prediction result.

[0068] Optimize the deep learning network based on annotations: Optimize the original deep learning network based on the prior conversation response training annotations and conversation response prediction results. A loss function can be used to measure the difference between the predicted result and the correct result. For example, for classification problems, an exemplary cross entropy loss function (L = -sum_{i = 1}^{q}y_i^{true}log(y_i^{pred})), where (y_i^{true}) is the true probability of the correct response type in the prior conversation response training annotations (if it is in one-hot encoding form, only the correct type is 1, and the others are 0), and (y_i^{pred}) is the probability in the conversation response prediction results. Through the back-propagation algorithm, the parameters in the original deep learning network are adjusted according to the loss function to obtain an intermediate deep learning network.

[0069] The target deep learning network is obtained through loop debugging: the intermediate deep learning network is used as the original deep learning network, and the previous steps of obtaining data samples, encoding, cross-collision, response output and optimization are repeated until the debugging process is completed. This loop process continuously adjusts the network parameters so that the prediction ability of the network is gradually improved. When certain stopping conditions are met (such as reaching a preset number of iterations or the loss function value is less than a certain threshold), the target deep learning network is obtained. Finally, the optimized initial AI knowledge embedding network is obtained based on the target knowledge vector encoding module and the target knowledge cross-collision module in the target deep learning network.

[0070] It can be seen that by loading the associated conversation interaction knowledge information set into the initial AI knowledge embedding network for knowledge vector encoding and cross-collision, the powerful feature extraction capability of the deep learning network is utilized to process knowledge information more accurately. During the debugging process, based on the computing power resource conversation data samples and the prior conversation response training annotations, the original deep learning network is continuously optimized to improve the accuracy of the network's responses to different types of computing power resource conversations. The knowledge vector encoding module uses specific algorithms such as CNN-based encoding to effectively capture semantic information in conversations. The operation mode of the knowledge cross-collision module further integrates the information of different knowledge encodings. The entire debugging cycle ensures that the network gradually converges to the optimal state, and the target AI knowledge embedding network obtained in the end can operate more efficiently and accurately in the computing power resource conversation response processing, improving the ability of the entire system to handle computing power resource conversations.

[0071] In some preferred embodiments, the identifying of target conversation interaction knowledge in the computing power resource session data to be processed to obtain a target conversation interaction knowledge information set includes: mining the conversation interaction semantic vector of the computing power resource session data to be processed, performing feature preprocessing based on the conversation interaction semantic vector to obtain a conversation interaction semantic preprocessing vector; performing conversation requirement decoding on the conversation interaction semantic preprocessing vector to obtain a conversation requirement decoding vector, and collecting information based on the conversation requirement decoding vector to obtain the target conversation interaction knowledge information set.

[0072] It is understandable that in order to mine the conversation interaction semantic vectors of the computing resource session data to be processed, a method combining word embedding technology and deep learning architecture can be used. For example, a pre-trained Bidirectional Encoder Representation (BERT) model is used. The BERT model is pre-trained on a large-scale text corpus and can capture rich semantic information.

[0073] For example, the session data of computing resources to be processed is represented as a text sequence (T = {t_1, t_2, cdots, t_n}), where (t_i) represents the (i)th word or token in the session data. This text sequence is input into the BERT model. The BERT model contains a Transformer structure with multiple layers. In the Transformer structure, for each input vocabulary token (t_i), the relationship between it and other vocabulary tokens is calculated through the Multi-Head Attention mechanism.

[0074] The calculation formula of the multi-head attention mechanism is as follows: First, for the attention calculation of a single head, (Attention(Q, K, V) = softmax(frac{QK^T}{sqrt{d_k}})V), where (Q), (K), (V) are the query vector, key vector and value vector, respectively, which are obtained by linearly transforming the input vocabulary token vector, and (d_k) is the dimension of the key vector.

[0075] In the multi-head attention mechanism, the attention results of multiple heads are concatenated and then linearly transformed to obtain the final result. For example, using an attention mechanism with (h) heads, for each head (i), (Attention_i(Q, K, V)) is calculated, and the final multi-head attention result is (MultiHead(Q, K, V) = Concat(Attention_1(Q, K, V), cdots, Attention_h(Q, K, V))W^O), where (W^O) is the output linear transformation weight matrix.

[0076] After multiple layers of processing by the BERT model, the vector output by the last layer can be used as the session interaction semantic vector (\(\vec{s}\)). This vector contains the semantic information of the entire session data, and its dimension depends on the settings of the BERT model, for example, it is 768 dimensions.

[0077] Then, after obtaining the session interaction semantic vector (\(\vec{s}\)), perform feature preprocessing operations. This may include normalization, dimensionality reduction, etc.

[0078] For normalization, the min-max normalization method can be used. For example, for the session interaction semantic vector (\(\vec{s}=(s_1, s_2, \cdots, s_d)\)) (\(d\) is the vector dimension, such as \(d = 768\)), the min-max normalization formula is \(s_i'=\frac{s_i - min(s)}{max(s)-min(s)}\), where \(s_i'\) is the \(i\)-th element after normalization, and \(min(s)\) and \(max(s)\) are the minimum and maximum values in the vector \(\vec{s}\) respectively. By normalization, the values of the vector elements are mapped to the interval \([0, 1]\), which helps to improve the stability of subsequent processing.

[0079] For dimensionality reduction, the principal component analysis (PCA) method can be used. For example, the covariance matrix of the original session interaction semantic vector \(\vec{s}\) is \(\Sigma\), calculate the eigenvalues \((\lambda_1, \lambda_2, \cdots, \lambda_d)\) and the corresponding eigenvectors \((\vec{u}_1, \vec{u}_2, \cdots, \vec{u}_d)\) of \(\Sigma\). Sort the eigenvectors according to the eigenvalue magnitudes, and select the first \(k\) eigenvectors (\(k < d\), for example, \(k = 128\)) to construct the projection matrix \(P = [\vec{u}_1, \vec{u}_2, \cdots, \vec{u}_k]\). The session interaction semantic preprocessing vector \(\vec{s}' = P^T\vec{s}\), reducing the original \(d\)-dimensional vector to \(k\) dimensions.

[0080] In this embodiment, the session requirement decoding method can adopt a method based on a recurrent neural network (RNN) for session requirement decoding. Specifically, use a long short-term memory network (LSTM).

[0081] Input the elements of the session interaction semantic preprocessing vector \(\vec{s}'\) into the LSTM network in sequence. The core units of the LSTM network include an input gate (\(i_t\)), a forget gate (\(f_t\)), an output gate (\(o_t\)), and a cell state (\(c_t\)).

[0082] The calculation formula for the input gate is:

[0083] (i_t=sigma(W_{xi}x_t+W_{hi}h_{t-1}+b_i)), where (x_t) is the input vector element (i.e. the element in (vec{s}')), (W_{xi}) and (W_{hi}) are weight matrices, (h_{t-1}) is the hidden state at the previous moment, (b_i) is the bias vector, and (sigma) is the sigmoid function.

[0084] The calculation formula of the forget gate is:

[0085] (f_t=sigma(W_{xf}x_t+W_{hf}h_{t-1}+b_f)).

[0086] The updating formula of cell state is (c_t=f_tc_{t-1}+i_ttanh(W_{xc}x_t+W_{hc}h_{t-1}+b_c)).

[0087] The calculation formula of the output gate is:

[0088] (o_t=sigma(W_{xo}x_t+W_{ho}h_{t-1}+b_o)), the final hidden state (h_t=o_ttanh(c_t)).

[0089] After being processed by the LSTM network, the hidden state (h_T) at the last moment (for example, the input sequence length is (T)) can be used as the session demand decoding vector (vec{d}).

[0090] Finally, information is collected based on the session demand decoding vector to obtain the target session interaction knowledge information set: Information is collected based on the session demand decoding vector (vec{d}). A rule-based information extraction method can be used in combination with a predefined knowledge base.

[0091] For example, the knowledge base contains various computing resource-related knowledge templates, such as computing resource demand templates, computing resource failure templates, etc. For the session demand decoding vector (vec{d}), calculate its similarity with each knowledge template in the knowledge base. The cosine similarity calculation method can be used. For the vector (vec{d}) and the knowledge template vector (vec{k}) in the knowledge base, the cosine similarity (cos(vec{d}, vec{k})

[0092] =frac{vec{d}cdotvec{k}}{vertvec{d}vertvertvec{k}vert}).

[0093] According to the calculated similarity, the knowledge content corresponding to the knowledge template with the highest similarity is selected as the target conversation interaction knowledge information set. For example, if the similarity with the computing resource requirement template is the highest, then the target conversation interaction knowledge information set includes relevant knowledge content about computing resource requirements, such as the type of requirement (computing power requirement, storage requirement, network bandwidth requirement, etc.), the scale of the requirement (specific numerical requirements, etc.).

[0094] With this design, first, by mining the conversation interaction semantic vector, the semantic information of the computing resource conversation data to be processed is effectively captured using the multi-head attention mechanism in the BERT model. The normalization and dimensionality reduction operations in feature preprocessing improve the stability of the data and reduce the amount of calculation. LSTM is used for conversation demand decoding, and its internal gating mechanism can well handle the long-distance dependencies in the sequence data and accurately convert the preprocessed semantic vector into the conversation demand decoding vector. Finally, information is collected based on the decoding vector, and the target conversation interaction knowledge information set can be accurately determined by calculating the similarity with the knowledge template in the knowledge base, which helps to improve the accuracy of the entire computing resource conversation response processing system in understanding the conversation content, thereby improving the system's response quality and efficiency.

[0095] In some other preferred embodiments, the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector is determined from the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords based on the correlation feature values ​​of the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, including: determining feature commonality scores of the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords to obtain respective correlation feature values; determining a target correlation feature value from the respective correlation feature values, and using the associated session interaction knowledge vector corresponding to the target correlation feature value as the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector.

[0096] In the actual application process, the following is a detailed description of the technical solution recorded in this embodiment:

[0097] 1. Determine the feature commonality score of the current conversation interaction knowledge vector and the associated conversation interaction knowledge vector to obtain the correlation feature value

[0098] Selection of calculation method for feature commonality score: In order to determine the feature commonality score (i.e., correlation feature value) of the current session interaction knowledge vector and the associated session interaction knowledge vector corresponding to each computing resource business keyword, multiple methods can be used. One effective method is to combine the cosine similarity calculation based on the vector space model (VSM) with the semantic distance metric.

[0099] Calculation based on cosine similarity: Assume that the current session interaction knowledge vector is (vec{v}_{current}), its dimension is (d) (for example, (d=300)), and the associated session interaction knowledge vector corresponding to a computing resource business keyword is (vec{v}_{associated}).

[0100] The cosine similarity formula is:

[0101] (cosine(vec{v}_{current}, vec{v}_{associated})=

[0102] frac{vec{v}_{current}cdotvec{v}_{associated}}{vertvec{v}_{current}vertvertvec{v}_{associated}vert});

[0103] in:

[0104] (vec{v}_{current}cdotvec{v}_{associated}

[0105] =sum_{i=1}^{d}v_{current, i}v_{associated, i}), (vertvec{v}_{current}vert=sqrt{sum_{i=1}^{d}v_{current, i}^2}), (vertvec{v}_{associated}vert=sqrt{sum_{i=1}^{d}v_{associated, i}^2}). The value of this cosine similarity reflects the degree of similarity between the two vectors in direction, and the value range is between ([-1, 1]). The closer the value is to (1), the more similar they are.

[0106] Supplement of semantic distance metric: Cosine similarity alone may not fully and accurately reflect feature commonality. Therefore, a semantic distance metric is introduced. For example, a semantic knowledge base such as WordNet is used to calculate the semantic distance between words. For example, the words in the current session interaction knowledge vector and the associated session interaction knowledge vector are (w_{1}, w_{2}, cdots) and (w_{1}', w_{2}', cdots). For each pair of corresponding words (if they exist), their semantic distance (d(w_i, w_i')) is determined by querying WordNet.

[0107] An exemplary comprehensive calculation method is to perform a weighted summation of cosine similarity and semantic distance. Assume weights (w_{cos}) and (w_{sem}) (e.g., (w_{cos}=0.7), (w_{sem}=0.3)), then the feature commonality score (correlation feature value) (r=w_{cos}*cosine(vec{v}_{current}, vec{v}_{associated})+w_{sem}*frac{1}{1+sum_{i}d(w_i, w_i')}). In this formula, (frac{1}{1+sum_{i}d(w_i, w_i')}) is a normalization of the semantic distance, so that the influence of the semantic distance is combined at the same level as the cosine similarity.

[0108] 2. Determine the target correlation feature value from each correlation feature value and determine the target associated conversation interaction knowledge vector

[0109] Method for determining target correlation feature value: After obtaining each correlation feature value, it is necessary to determine the target correlation feature value from these values. A common method is to set a threshold (t) (eg (t=0.6)) and filter out correlation feature values ​​greater than or equal to the threshold.

[0110] If there are multiple correlation feature values ​​that meet this condition, they can be further selected according to other rules. For example, the largest correlation feature value can be selected as the target correlation feature value. This is because the largest correlation feature value indicates that the current session interaction knowledge vector is closest to the corresponding associated session interaction knowledge vector in terms of feature commonality and is most likely to contain knowledge content related to the current session.

[0111] Determine the target associated session interaction knowledge vector: Once the target correlation feature value is determined, the associated session interaction knowledge vector corresponding to the target correlation feature value is used as the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector. For example, if the correlation feature value (r^*) corresponding to a certain associated session interaction knowledge vector (vec{v}_{associated}^*) is determined as the target correlation feature value, then (vec{v}_{associated}^*) is the target associated session interaction knowledge vector. This target associated session interaction knowledge vector will be used in subsequent steps to determine the session interaction business topic keywords, and then determine the computing power resource session response results corresponding to the computing power resource session data to be processed.

[0112] The above-mentioned embodiment uses cosine similarity combined with semantic distance measurement to determine the correlation feature value, which can more comprehensively evaluate the feature commonality between the current conversation interaction knowledge vector and the associated conversation interaction knowledge vector. Cosine similarity measures the directional similarity from the perspective of vector space, while the semantic distance measurement supplements the difference considerations at the lexical semantic level. The correlation feature value obtained by weighted summation of the two more accurately reflects the actual correlation. When determining the target correlation feature value, by setting thresholds and further screening rules, the most relevant vector can be accurately selected from multiple associated conversation interaction knowledge vectors as the target associated conversation interaction knowledge vector. This helps to improve the accuracy of the entire computing resource conversation response processing system, so that the final determined conversation response result is more in line with the actual needs of the conversation, thereby improving the performance and effectiveness of the system.

[0113] In some examples, the computing power resource business keywords include computing power scheduling business keywords and cloud computing power mode update keywords; the determining, based on the correlation feature values ​​of the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector from the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords includes: determining a feature commonality score between the current session interaction knowledge vector and the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword to obtain a computing power scheduling correlation feature value; determining a feature commonality score between the current session interaction knowledge vector and the associated session interaction knowledge vector corresponding to the cloud computing power mode update keyword to obtain the mode update correlation feature value; and determining, based on a distinguishing variable between the computing power scheduling correlation feature value and the mode update correlation feature value, a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector from the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword and the associated session interaction knowledge vector corresponding to the cloud computing power mode update keyword.

[0114] Further, the determining, based on the distinguishing variable between the computing power scheduling correlation feature value and the pattern update correlation feature value, a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector from the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword and the associated session interaction knowledge vector corresponding to the cloud computing power pattern update keyword includes: determining the distinguishing variable between the computing power scheduling correlation feature value and the pattern update correlation feature value, and when the distinguishing variable is greater than a set variable value, performing a comparison analysis on the computing power scheduling correlation feature value and the pattern update correlation feature value; when the computing power scheduling correlation feature value is greater than the pattern update correlation feature value, taking the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword as the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector; when the pattern update correlation feature value is greater than the computing power scheduling correlation feature value, taking the associated session interaction knowledge vector corresponding to the cloud computing power pattern update keyword as the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector; when the distinguishing variable is not greater than the set variable value, taking the general business theme keyword as the session interaction business theme keyword of the target session interaction knowledge information set.

[0115] Similar to the previous method of determining the correlation feature value, it is first necessary to determine the feature commonality score of the current session interaction knowledge vector and the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword to obtain the computing power scheduling correlation feature value.

[0116] For example, a method based on a vector space model (VSM) is used in combination with semantic analysis. Let the current session interaction knowledge vector be (vec{v}_{current}), and the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword be (vec{v}_{scheduling}).

[0117] Cosine similarity calculation based on vector space model: (cosine(vec{v}_{current}, vec{v}_{scheduling})

[0118] =frac{vec{v}_{current}cdotvec{v}_{scheduling}}{vertvec{v}_{current}vertvertvec{v}_{scheduling}vert}), where:

[0119] (vec{v}_{current}cdotvec{v}_{scheduling}

[0120] =sum_{i=1}^{d}v_{current, i}v_{scheduling, i}), (vertvec{v}_{current}vert=sqrt{sum_{i=1}^{d}v_{current, i}^2}), (vertvec{v}_{scheduling}vert=sqrt{sum_{i=1}^{d}v_{scheduling, i}^2}), where (d) is the vector dimension, for example (d=300).

[0121] At the same time, semantic analysis is supplemented. For example, the words in the two vectors are (w_{1}, w_{2}, cdots) and (w_{1}', w_{2}', cdots). For each pair of corresponding words (if any), their semantic distance (d(w_i, w_i')) is determined by querying the semantic knowledge base (such as WordNet). A comprehensive calculation method is to perform a weighted summation of the cosine similarity and the semantic distance to obtain the computing power scheduling correlation feature value (r_{scheduling}). Assume weights (w_{cos}=0.7), (w_{sem}=0.3), then (r_{scheduling}=w_{cos}*cosine(vec{v}_{current}, vec{v}_{scheduling})+w_{sem}*frac{1}{1+sum_{i}d(w_i, w_i')}).

[0122] Determine the correlation feature value with the cloud computing power mode update keyword (mode update correlation feature value): Similarly, determine the feature commonality score of the current session interaction knowledge vector and the associated session interaction knowledge vector corresponding to the cloud computing power mode update keyword to obtain the mode update correlation feature value.

[0123] Assume that the current conversation interaction knowledge vector is (vec{v}_{current}), and the associated conversation interaction knowledge vector corresponding to the cloud computing mode update keyword is (vec{v}_{update}).

[0124] First calculate the cosine similarity (cosine(vec{v}_{current}, vec{v}_{update})=frac{vec{v}_{current}cdotvec{v}_{update}}{vertvec{v}_{current}vertvertve c{v}_{update}vert}).

[0125] Then perform semantic analysis to determine the semantic distance between words (d(w_i, w_i')), and obtain the pattern update correlation feature value (r_{update}) according to the above weighted summation method, that is, (r_{update}=w_{cos}*cosine(vec{v}_{current}, vec{v}_{update})+w_{sem}*frac{1}{1+sum_{i}d(w_i, w_i')}).

[0126] 2. Determine the target-related conversation interaction knowledge vector based on the distinguishing variables

[0127] Determine the distinguishing variable: The distinguishing variable reflects the degree of difference between the computing power scheduling correlation characteristic value and the mode update correlation characteristic value. Calculate the distinguishing variable (Delta = vertr_{scheduling}-r_{update}vert).

[0128] Processing when the difference variable is greater than the set variable value: The set variable value is a predetermined threshold, for example, the set variable value is (0.2). When (Delta>0.2), a comparison analysis is performed on (r_{scheduling}) and (r_{update}).

[0129] If (r_{scheduling}>r_{update}), this means that the current session interaction knowledge vector is closer to the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword in terms of feature commonality. Therefore, the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword is used as the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector.

[0130] On the contrary, if (r_{update}>r_{scheduling}), the associated session interaction knowledge vector corresponding to the cloud computing mode update keyword is used as the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector.

[0131] Processing when the distinguishing variable is not greater than the set variable value: When (Deltaleq0.2), it means that the current session interaction knowledge vector and the associated session interaction knowledge vector corresponding to the computing power scheduling business keyword and the associated session interaction knowledge vector corresponding to the cloud computing power mode update keyword have little difference in feature commonality. At this time, the general business theme keyword is used as the session interaction business theme keyword of the target session interaction knowledge information set. The general business theme keyword is a pre-defined keyword with a wide coverage and related to a variety of computing power resource businesses, such as "basic management of computing power resources". This means that the current session interaction knowledge vector may not be clearly attributed to the computing power scheduling business or the cloud computing power mode update business, so the general business theme keyword is used to process subsequent session responses.

[0132] It can be seen that by respectively determining the correlation feature values ​​corresponding to the computing power scheduling business keywords and the cloud computing power mode update keywords, a comprehensive vector space model and semantic analysis method are adopted to accurately evaluate the degree of correlation between the current conversation interaction knowledge vector and the knowledge vectors related to different business keywords. The target associated conversation interaction knowledge vector is determined based on the distinguishing variable between the computing power scheduling correlation feature value and the mode update correlation feature value, and the variable value is set as a distinguishing boundary, so that reasonable decisions can be made under different correlation feature value differences. When the difference is large, the associated conversation interaction knowledge vector corresponding to the most relevant business keyword can be accurately selected according to the specific size relationship; when the difference is small, the general business theme keyword is used to avoid misjudgment, improve the accuracy and flexibility of the entire computing power resource conversation response processing system, and thus improve the system's ability to handle different types of conversations.

[0133] In an optional embodiment, the target session interaction knowledge information set includes at least two; the determining of the computing power resource session response result corresponding to the computing power resource session data to be processed based on the session interaction business subject keywords of the target session interaction knowledge information set includes: determining the thermal evaluation coefficients of the at least two target session interaction knowledge information sets and the computing power resource session data to be processed respectively, and taking the session interaction business subject keywords of the target session interaction knowledge information set with the largest thermal evaluation coefficient as the computing power resource session response result corresponding to the computing power resource session data to be processed.

[0134] In this embodiment, the thermal evaluation coefficient is an indicator used to measure the degree of correlation between the target conversation interaction knowledge information set and the computing resource conversation data to be processed. By calculating this coefficient, it is possible to determine which target conversation interaction knowledge information set is the most critical in the current conversation context, thereby providing a basis for determining the correct computing resource conversation response result.

[0135] For each target conversation interaction knowledge information set, a method based on the term frequency-inverse document frequency (TF-IDF) algorithm combined with semantic similarity is used to calculate its thermal evaluation coefficient with the conversation data of the computing power resources to be processed.

[0136] First, we perform word frequency statistics on the session data of computing power resources to be processed. Let the session data of computing power resources to be processed be (D), and the vocabulary set contained in it is (V = {v_1, v_2, cdots, v_n}). For each vocabulary (v_i), its word frequency (tf(v_i, D)) in (D) is defined as the number of times the vocabulary appears in (D).

[0137] Then, the inverse document frequency (idf(v_i)) is calculated. For example, there is a corpus (C = {D_1, D_2, cdots, D_m}) containing multiple documents (here other historical computing resource session data can be regarded as documents), and the number of documents containing the word (v_i) is (df(v_i)), then (idf(v_i) = logfrac{m}{df(v_i)+1}). The addition (1) here is to avoid the situation where the denominator is (0).

[0138] For the target conversation interaction knowledge information set (K_j) ((j=1, 2, cdots) represents different target conversation interaction knowledge information sets), the vocabulary set contained in it is (V_j={v_{j1}, v_{j2}, cdots, v_{jn_j}}). Calculate the TF-IDF-based similarity between (K_j) and (D) (s_{TF-IDF}(K_j, D)):

[0139] First, calculate the TF-IDF value of each vocabulary (v_{ji}) in (K_j) in (D) (tf-idf(v_{ji}, D) = tf(v_{ji}, D)*idf(v_{ji})).

[0140] Then, (s_{TF-IDF}(K_j, D) = sum_{v_{ji}inV_jcapV}tf-idf(v_{ji}, D)), this value reflects the similarity between (K_j) and (D) from the perspective of term frequency-inverse document frequency.

[0141] At the same time, the semantic similarity is calculated. A pre-trained word vector model (such as Word2Vec) is used to obtain the word vector of the vocabulary. For the vocabulary (v_{ji}) in (K_j) and the vocabulary (v_i) in (D), if there are the same vocabulary, their semantic similarity is (1); if they are different vocabulary, let their word vectors be (vec{w}_{ji}) and (vec{w}_{i}) respectively, then their semantic similarity is (s_{sem}(v_{ji},v_i)=frac{vec{w}_{ji}cdotvec{w}_{i}}{vertvec{w}_{ji}vertvertvec{w}_{i}vert}).

[0142] The thermal evaluation coefficient (h_j) is calculated by combining TF-IDF similarity and semantic similarity. Assume weights (w_{TF-IDF}=0.6), (w_{sem}=0.4), then (h_j=w_{TF-IDF}*s_{TF-IDF}(K_j,D)+w_{sem}*frac{1}{n_j}sum_{v_{ji}inV_jcapV}s_{sem}(v_{ji},v_i)). Here (frac{1}{n_j}sum_{v_{ji}inV_jcapV}s_{sem}(v_{ji},v_i)) is the average semantic similarity of the common words in (K_j) and (D).

[0143] After calculating the thermal evaluation coefficient (h_j) corresponding to each target conversation interaction knowledge information set, compare the magnitudes of these coefficients. For example, there are three target conversation interaction knowledge information sets (K_1), (K_2), and (K_3), and their thermal evaluation coefficients are calculated to be (h_1 = 0.45), (h_2 = 0.52), and (h_3 = 0.48), respectively.

[0144] Since (h_2) is the largest, the session interaction service subject keyword of the target session interaction knowledge information set (K_2) is used as the computing resource session response result corresponding to the computing resource session data to be processed. This session interaction service subject keyword will be used to query the knowledge base or generate response content. If the knowledge base is queried, the corresponding response information is searched in the knowledge base with this subject keyword as the index; if the response content is generated, the knowledge related to this subject keyword is used as the basis to generate appropriate response content according to the predefined template or rule.

[0145] It can be seen that by calculating the thermal evaluation coefficient of the target conversation interaction knowledge information set and the computing resource conversation data to be processed, the word frequency-inverse document frequency algorithm and the semantic similarity method are combined. The TF-IDF algorithm can measure the importance of words from the frequency and rarity of the words in the document, thereby reflecting the degree of association between the target conversation interaction knowledge information set and the conversation data to be processed at the lexical level. Semantic similarity further supplements the evaluation of the degree of association from the perspective of lexical semantics. The thermal evaluation coefficient obtained by weighted summation of the two more accurately reflects the actual degree of association between the target conversation interaction knowledge information set and the conversation data to be processed. The conversation interaction business topic keywords of the target conversation interaction knowledge information set with the largest thermal evaluation coefficient are selected as the response result, which improves the accuracy of determining the result of the computing resource conversation response, enables the system to make appropriate responses to different conversation situations more accurately, and improves the performance of the entire conversation response processing system.

[0146] On the basis of step 202-step 208, the method further includes: loading the target session interaction knowledge information set and the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords into a bidirectional long short-term memory network; mining the current session interaction knowledge vector corresponding to the target session interaction knowledge information set through a knowledge vector encoding module in the bidirectional long short-term memory network, and determining the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector from the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords based on the correlation feature values ​​between the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, and using the computing power resource business keyword corresponding to the target associated session interaction knowledge vector as the session interaction business theme keyword corresponding to the target session interaction knowledge information set.

[0147] Furthermore, the debugging method of the bidirectional long short-term memory network includes: obtaining a sample of a conversation interaction knowledge information set and a corresponding scene conversation response training annotation; loading the sample of the conversation interaction knowledge information set and the associated conversation interaction knowledge vectors corresponding to the business keywords of each computing resource into the original bidirectional long short-term memory network for conversation keyword analysis to obtain the conversation interaction business prediction keywords corresponding to the sample of the conversation interaction knowledge information set; optimizing the original bidirectional long short-term memory network according to the conversation interaction business prediction keywords and the scene conversation response training annotation to obtain an optimized bidirectional long short-term memory network; using the optimized bidirectional long short-term memory network as the original bidirectional long short-term memory network, and jumping to the step of obtaining the sample of the conversation interaction knowledge information set and the corresponding scene conversation response training annotation to implement in a loop until the debugging process is completed to obtain the bidirectional long short-term memory network.

[0148] Based on the previous steps 202-208, the target conversation interaction knowledge information set and the associated conversation interaction knowledge vectors corresponding to the business keywords of each computing resource are loaded into the bidirectional long short-term memory network. The bidirectional long short-term memory network is a special recurrent neural network (RNN) that can process long-distance dependencies in sequence data, and due to its bidirectional structure, it can simultaneously consider the forward and reverse information of the sequence.

[0149] The knowledge vector encoding module in the bidirectional long short-term memory network is responsible for mining the current conversation interaction knowledge vector corresponding to the target conversation interaction knowledge information set.

[0150] Assume that the target conversation interaction knowledge information set is represented as (K = {k_1, k_2, cdots, k_n}), and input it into the bidirectional long short-term memory network in sequence. The basic unit of the bidirectional long short-term memory network is the long short-term memory unit (LSTM unit). For forward propagation, the calculation of the LSTM unit is as follows:

[0151] Forget gate (f_t = sigma(W_{xf}x_t+W_{hf}h_{t-1}+b_f)), where (x_t) is the input element (here is the element in the knowledge information set), (W_{xf}) and (W_{hf}) are weight matrices, (h_{t-1}) is the hidden state at the previous moment, (b_f) is the bias vector, and (sigma) is the sigmoid function.

[0152] Input gate: (i_t=sigma(W_{xi}x_t+W_{hi}h_{t-1}+b_i)).

[0153] Cell Status:

[0154] (c_t=f_tc_{t-1}+i_ttanh(W_{xc}x_t+W_{hc}h_{t-1}+b_c)).

[0155] Output Gate:

[0156] (o_t=sigma(W_{xo}x_t+W_{ho}h_{t-1}+b_o)), the final hidden state (h_t=o_ttanh(c_t)).

[0157] For back propagation, the calculation process is similar, except that the input order is reversed. Finally, the hidden states obtained by forward and backward propagation are combined (e.g., concatenated) to obtain the current session interaction knowledge vector (vec{v}_{current}).

[0158] Then, according to the correlation feature values ​​of the current session interaction knowledge vector (vec{v}_{current}) and the associated session interaction knowledge vectors corresponding to each computing resource business keyword, the target associated session interaction knowledge vector is determined from these associated session interaction knowledge vectors.

[0159] The method for calculating the correlation feature value can be cosine similarity. Assume that the associated conversation interaction knowledge vector corresponding to a computing resource business keyword is (vec{v}_{associated}), then cosine similarity (cosine(vec{v}_{current}, vec{v}_{associated}) = frac{vec{v}_{current}cdotvec{v}_{associated}}{vertvec{v}_{current}vertver tvec{v}_{associated}vert});

[0160] in:

[0161] (vec{v}_{current}cdotvec{v}_{associated}

[0162] =sum_{i=1}^{d}v_{current, i}v_{associated, i}), (vertvec{v}_{current}vert=sqrt{sum_{i=1}^{d}v_{current, i}^2}), (vertvec{v}_{associated}vert=sqrt{sum_{i=1}^{d}v_{associated, i}^2}), where (d) is the vector dimension, for example (d=300). By calculating the cosine similarity with each associated conversation interaction knowledge vector, the associated conversation interaction knowledge vector corresponding to the maximum cosine similarity is found as the target associated conversation interaction knowledge vector, and the computing power resource business keyword corresponding to the target associated conversation interaction knowledge vector is used as the conversation interaction business topic keyword of the corresponding target conversation interaction knowledge information set.

[0163] Furthermore, the debugging method of the bidirectional long short-term memory network is as follows.

[0164] 1) Obtain data samples and training annotations: First, obtain the conversation interaction knowledge information set samples and the corresponding scene conversation response training annotations. The conversation interaction knowledge information set samples are representative samples extracted from a large amount of computing resource conversation data, and the scene conversation response training annotations are the annotation information of the correct response results corresponding to these samples.

[0165] 2) Perform conversation keyword analysis to obtain predicted keywords: Load the conversation interaction knowledge information set samples and the associated conversation interaction knowledge vectors corresponding to each computing resource business keyword into the original bidirectional long short-term memory network for conversation keyword analysis.

[0166] According to the calculation method of the bidirectional long short-term memory network described above, the input conversation interaction knowledge information set samples are processed to finally obtain the conversation interaction business prediction keywords corresponding to the conversation interaction knowledge information set samples.

[0167] 3) Optimize the network based on predicted keywords and training annotations: Optimize the original bidirectional long short-term memory network based on the conversational interaction business prediction keywords and scenario conversation response training annotations.

[0168] The cross entropy loss function can be used to measure the difference between the predicted keywords and the correct keywords. Assume that the probability distribution of the predicted conversational interaction service keywords is (y^{pred}={y_1^{pred},y_2^{pred},cdots,y_m^{pred}}) (where (m) is the number of possible keyword types), and the probability distribution of the correct conversational interaction service keywords is (y^{true}={y_1^{true},y_2^{true},cdots,y_m^{true}}) (usually one-hot encoding is used, that is, the correct type is (1) and the others are (0)), then the cross entropy loss function is (L=-sum_{i=1}^{m}y_i^{true}log(y_i^{pred})).

[0169] Through the back propagation algorithm, the weight parameters in the bidirectional long short-term memory network are updated according to the gradient calculated by the loss function to obtain an optimized bidirectional long short-term memory network.

[0170] 4) Debug the loop to get the final network: Use the optimized bidirectional long short-term memory network as the original bidirectional long short-term memory network, repeat the previous steps of obtaining samples, parsing, and optimization until the debugging process is completed to get the bidirectional long short-term memory network. This loop process continuously adjusts the network parameters, so that the network's ability to predict conversational interaction business keywords gradually improves. The resulting bidirectional long short-term memory network can more accurately process the relationship between the target conversational interaction knowledge information set and the associated conversational interaction knowledge vector, thereby improving the performance of the entire computing resource conversation response processing system.

[0171] With this design, by loading the target conversation interaction knowledge information set and the associated conversation interaction knowledge vector into the bidirectional long short-term memory network, and utilizing its ability to process sequence data bidirectionally, the current conversation interaction knowledge vector can be mined more comprehensively. During the mining process, the calculation method of the bidirectional long short-term memory network can effectively capture the long-distance dependencies in the knowledge information, improving the accuracy of knowledge vector mining. When determining the target associated conversation interaction knowledge vector, based on methods such as cosine similarity, the most relevant vector can be accurately found, and then the accurate conversation interaction business topic keywords can be determined. The debugging method of the bidirectional long short-term memory network is based on the conversation interaction knowledge information set samples and the scene conversation response training annotations. The network is continuously optimized through the cross entropy loss function and the back propagation algorithm, ensuring that the prediction accuracy of the network is gradually improved, thereby improving the efficiency and accuracy of the entire system in processing computing resource conversations.

[0172] It is worth mentioning that the unique application and innovation of the bidirectional long short-term memory network (Bi-LSTM) in the above technical solution are described as follows.

[0173] 1. Unique application of bidirectional long short-term memory network in technical solutions

[0174] 1. Processing knowledge vector relationships

[0175] Mining the current conversation interaction knowledge vector: In the above technical solution, the knowledge vector encoding module of the bidirectional long short-term memory network (Bi-LSTM) is used to mine the current conversation interaction knowledge vector corresponding to the target conversation interaction knowledge information set. The unique structure of Bi-LSTM enables it to consider both the forward and reverse information of the input sequence (target conversation interaction knowledge information set) at the same time.

[0176] For example, when the target conversation interaction knowledge information set contains multiple elements, the traditional unidirectional network can only process the relationship between these elements in one direction (such as from front to back), and some reverse semantic information may be lost. Bi-LSTM can capture the dependencies between elements more comprehensively by processing forward and reverse LSTM units at the same time. Taking an exemplary target conversation interaction knowledge information set containing three elements (k_1, k_2, k_3) as an example, the forward LSTM unit will calculate a hidden state based on (k_1), and then combine (k_2) to calculate the next hidden state, and so on; the reverse LSTM unit starts from (k_3) and calculates the hidden state in reverse. Finally, the forward and reverse hidden states are combined (such as splicing) to obtain the current conversation interaction knowledge vector. This method can better mine the potential semantics and logical relationships in the knowledge information set, thereby obtaining a more accurate current conversation interaction knowledge vector.

[0177] Determine the target associated session interaction knowledge vector: The output of Bi-LSTM (current session interaction knowledge vector) is used to determine the correlation feature value of the associated session interaction knowledge vector corresponding to each computing resource business keyword. Since the knowledge vector mined by Bi-LSTM contains richer information, the relationship between the target session interaction knowledge information set and the associated session interaction knowledge vector can be more accurately reflected when calculating the correlation feature value (such as using cosine similarity calculation).

[0178] For example, if the associated session interaction knowledge vector has a complex semantic and logical relationship with the current session interaction knowledge vector (may be a forward, reverse, or multi-element relationship), the knowledge vector mined by Bi-LSTM can better capture this relationship, thereby improving the accuracy of determining the target associated session interaction knowledge vector. By finding the associated session interaction knowledge vector with the highest correlation eigenvalue, its corresponding computing resource business keyword can be accurately determined as the session interaction business topic keyword, which is crucial for the subsequent accurate determination of the computing resource session response result.

[0179] (II) Network debugging and optimization

[0180] Keyword parsing based on specific samples: In the debugging method of the bidirectional long short-term memory network, it is used to perform conversation keyword parsing on the conversation interaction knowledge information set samples. The conversation interaction knowledge information set samples and the associated conversation interaction knowledge vectors corresponding to the business keywords of each computing resource are loaded into the original bidirectional long short-term memory network.

[0181] Unlike general keyword parsing methods, Bi-LSTM can take into account long-distance dependencies and bidirectional semantic information in samples during parsing. For example, for a long conversational interaction knowledge information set sample, there may be semantic associations between early elements and late elements. Bi-LSTM can effectively capture this association through its bidirectional structure, thereby more accurately parsing conversational interaction business prediction keywords.

[0182] Optimize the network to improve accuracy: Based on the parsed conversation interaction business prediction keywords and scene conversation response training annotations, the bidirectional long short-term memory network is optimized using the cross entropy loss function and back propagation algorithm. Since Bi-LSTM can better process the semantic information in the sample during the initial parsing process, the network parameters can be adjusted more effectively during the optimization process.

[0183] For example, when calculating the cross entropy loss function, if there is a deviation between the predicted keyword and the correct keyword, the Bi-LSTM structure helps to determine which part of the sequence is caused by the incorrect semantic understanding, so as to more specifically adjust the weight parameters during the back propagation process and improve the network's prediction accuracy for conversational interaction business keywords. This optimization process is repeated continuously, so that the performance of the network is gradually improved, and finally a bidirectional long short-term memory network that can accurately handle computing resource conversation-related tasks is obtained.

[0184] 2. The innovation of bidirectional long short-term memory network in the above technical solutions

[0185] 1. Comparison with traditional technical means

[0186] Limitations of traditional networks: In traditional technical means in this field, exemplary unidirectional neural networks or rule-based methods are often used to process knowledge vector relationships related to computing resource sessions. When processing sequence data, unidirectional neural networks (such as ordinary RNNs or LSTMs) can only capture information in one direction and are prone to losing reverse semantic information. For example, when processing computing resource session data containing multiple sub-concepts, if there is a reverse logical relationship between early concepts and late concepts, the unidirectional network may not be able to capture it effectively.

[0187] Although rule-based methods can process knowledge vector relationships according to predefined rules, they lack the ability to adaptively learn data and are difficult to handle complex and changing computing resource conversation scenarios. For example, when new computing resource business keywords or new conversation modes appear, rule-based methods are difficult to effectively expand and adapt.

[0188] The advantages of Bi-LSTM are innovative: the application of bidirectional long short-term memory network in the above technical solution overcomes the limitations of traditional technical means. It processes both forward and reverse information through a bidirectional structure, and can better adapt to the complex semantic relationships in computing resource session data. For example, when determining the target-related session interaction knowledge vector, it can more accurately reflect the correlation between knowledge vectors in different directions, which is difficult to do with traditional unidirectional networks.

[0189] Moreover, the application of Bi-LSTM in the network debugging process can more accurately adjust parameters to improve accuracy when optimizing the network by utilizing its effective ability to capture long-distance dependencies. This method of combining Bi-LSTM with specific computing resource session processing tasks (such as mining knowledge vectors, determining target association vectors, network debugging, etc.) is not a conventional technical means in this field and reflects innovation.

[0190] (II) Innovative combination in specific technical solutions

[0191] Collaboration with other modules: In the above technical solution, Bi-LSTM works with other modules (such as knowledge vector encoding module, module for determining target-related conversation interaction knowledge vector according to correlation feature value, etc.) to form a complete computing resource conversation response processing system. This collaborative approach is unique. For example, the knowledge vector encoding module uses the current conversation interaction knowledge vector mined by Bi-LSTM for subsequent processing. The modules are interdependent and interact with each other, and together realize the entire process from the target conversation interaction knowledge information set to determining the conversation interaction business topic keywords.

[0192] This specific module combination and collaborative working method has not appeared in previous technologies in this field, and is not a common technical means that technicians in this field can easily think of. It reflects innovative thinking in solving the problem of computing resource session response processing, integrating Bi-LSTM into the architecture of the entire technical solution, and improving the overall performance and accuracy of the system.

[0193] Customization for computing resource session processing: The application of Bi-LSTM in this technical solution is customized for the specific field of computing resource session processing. For example, when processing the relationship between computing resource business keywords and the target session interaction knowledge information set, the parameter adjustment and application method of Bi-LSTM are to better adapt to the semantic characteristics and business logic of the computing resource field. This customized application for a specific field is not a conventional technical means in this field, which further reflects the innovation.

[0194] In some scalable embodiments, after taking the computing power resource business keywords corresponding to the target-associated session interaction knowledge vector as the session interaction business subject keywords of the corresponding target session interaction knowledge information set, and determining the computing power resource session response results corresponding to the computing power resource session data to be processed according to the session interaction business subject keywords of the target session interaction knowledge information set, the method further includes: based on the computing power resource session response results, obtaining a consultation response event data sequence, wherein the consultation response event data sequence includes a continuous plurality of groups of consultation response event data; obtaining a directed push event data sequence according to the consultation response event data sequence, wherein the directed push event data sequence includes a continuous plurality of groups of directed push event data; based on the Based on the consultation answer event data sequence, a consultation answer frequent item sequence is obtained through the first residual processing branch included in the resource conversation analysis network, wherein the consultation answer frequent item sequence includes multiple consultation answer frequent items; based on the directional push event data sequence, a directional push frequent item sequence is obtained through the second residual processing branch included in the resource conversation analysis network, wherein the directional push frequent item sequence includes multiple directional push frequent items; based on the consultation answer frequent item sequence and the directional push frequent item sequence, push label information corresponding to the consultation answer event data is obtained through the push discrimination branch included in the resource conversation analysis network; and directional user portrait information of the consultation answer event data sequence is determined according to the push label information.

[0195] The above technical solution first obtains the consultation response event data sequence based on the computing power resource session response result.

[0196] Concept and composition of consultation response event data sequence: After taking the computing power resource business keyword corresponding to the target-associated conversation interaction knowledge vector as the conversation interaction business subject keyword and determining the computing power resource conversation response result corresponding to the computing power resource conversation data to be processed based on it, the consultation response event data sequence must first be obtained based on this response result. The consultation response event data sequence is an ordered collection of a series of data groups related to consultation responses, where each group of consultation response event data contains various information about a specific consultation response interaction.

[0197] For example, for a consultation about computing resource allocation, the response result gives a specific allocation strategy. The related consultation response event data may include the timestamp of the consultation, the identity of the consultant (which can be a unique user number or department number), the specific content of the consultation (such as the details of the computing resource requirements under a specific task), the response content (such as the recommended resource allocation plan, related cost estimates, etc.), and some metadata of the consultation response interaction (such as the interaction channel, whether it is through a web interface, mobile application or internal system interface, etc.).

[0198] Then, a directed push event data sequence is obtained from the consultation response event data sequence.

[0199] Method for obtaining a directional push event data sequence: obtain a directional push event data sequence according to a consultation response event data sequence. A directional push event data sequence is also composed of multiple continuous groups of data, and each group of directional push event data is related to a specific directional push operation.

[0200] In order to obtain a sequence of targeted push event data, a rule-based mapping method can be used. For example, for certain specific elements in the consultation response event data (such as the consultant's identity, specific keywords in the consultation content, etc.), set predefined rules to determine the corresponding targeted push content. For example, if the consultation content contains the keyword "deep learning task computing power requirements", according to the pre-set rules, the targeted push event data may include pushing links to technical documents related to deep learning computing power optimization, preferential package information for deep learning tasks by cloud computing service providers, etc. These information related to targeted pushes form a sequence of targeted push event data in chronological order or in the order of association with the consultation response events.

[0201] Next, the consultation answer frequent item sequence is obtained through the first residual processing branch of the resource conversation analysis network.

[0202] Resource conversation analysis network and the first residual processing branch: The resource conversation analysis network is a deep learning network specially built for analyzing computing resource conversation related data. The first residual processing branch is used to process the consultation response event data sequence to obtain the consultation response frequent item sequence.

[0203] The residual processing branch adopts the idea of ​​the residual network (ResNet) structure and avoids the gradient vanishing or gradient exploding problem of deep neural networks during training by constructing residual blocks. For example, the first residual processing branch consists of multiple residual blocks, each of which contains two paths: one is the identity mapping path, and the other is the path through the convolution layer, batch normalization layer, and activation function layer (such as ReLU).

[0204] Assume that the consultation response event data sequence input to the first residual processing branch is (X = {x_1, x_2, cdots, x_n}), where (x_i) represents the (i)th group of consultation response event data. For a single residual block, for example, the input is (y), the output after the convolution layer is (z = W*y+b) (where (W) is the convolution kernel weight matrix, (b) is the bias vector), and the output after batch normalization is:

[0205] (z'=gamma*frac{z-mu}{sqrt{sigma^2+epsilon}}+beta)(where (gamma) and (beta) are learnable parameters, (mu) is the mean, (sigma^2) is the variance, and (epsilon) is a small constant to avoid the denominator being (0)), and the output after the activation function ReLU is (z"=max(0,z')). Then (z") is added to (y) of the identity mapping path to obtain the output of the residual block, and multiple residual blocks process the input data in sequence.

[0206] Obtaining the consultation answer frequent item sequence: After the multi-layer processing of the first residual processing branch, the output result is used to obtain the consultation answer frequent item sequence. A frequent item mining algorithm, such as the Apriori algorithm, can be used.

[0207] For example, the output data after the first residual processing branch is represented as (Y = {y_1, y_2, cdots, y_m}), which is regarded as a transaction database, where each (y_i) is a transaction (containing multiple items, corresponding to different elements in the consultation response event data). For the Apriori algorithm, first determine the minimum support threshold (s) (for example, (s = 0.3)), that is, the minimum proportion of an item set appearing in all transactions.

[0208] The algorithm first scans all transactions, counts the number of occurrences of a single item, and finds the frequent (1)-item set (L_1) that meets the minimum support threshold. Then, the candidate item set (C_2) is generated through (L_1), and the transaction database is scanned again to count the number of occurrences of the item set in (C_2), and the frequent (2)-item set (L_2) that meets the threshold is found, and so on, until no new frequent item sets can be generated. The final frequent item set sequence is the consultation answer frequent item sequence, in which each frequent item may be a specific consultant identity, a frequently appearing consultation content keyword, or a common answer type, etc.

[0209] The directed push frequent item sequence is further obtained through the second residual processing branch of the resource session analysis network.

[0210] Operation of the second residual processing branch: Similar to the first residual processing branch, the second residual processing branch is used in the resource session analysis network to process the directional push event data sequence to obtain the directional push frequent item sequence. The residual block of the residual network structure is also used for processing.

[0211] Assume that the input directional push event data sequence is (A = {a_1, a_2, cdots, a_p}), which is processed by multiple residual blocks. The calculation method of each residual block is similar to that of the residual block in the first residual processing branch. For example, for the input (b), after the convolution layer (W_1*b+b_1), batch normalization, activation function and other operations, the output is added to the identity mapping path, and multiple residual blocks process the input data in turn.

[0212] Obtaining a frequent item sequence for directional push: using a frequent item mining algorithm similar to that for obtaining a frequent item sequence for consultation responses (such as the Apriori algorithm) to process the output data processed by the second residual processing branch.

[0213] For example, the processed output data is (B = {b_1, b_2, cdots, b_q}), and the minimum support threshold is set (for example, (s' = 0.2)). By scanning the data and counting the number of item sets, the frequent item sets that meet the threshold are found. These frequent item sets form a frequent item sequence for targeted push. The frequent items may be content types that are frequently pushed, pushes that are frequently received by a specific user group, etc.

[0214] Then, the push label information is obtained based on the consultation answer frequent item sequence and the directional push frequent item sequence.

[0215] Working principle of push discrimination branch: The push discrimination branch in the resource conversation analysis network obtains the push label information corresponding to the consultation answer event data based on the consultation answer frequent item sequence and the directional push frequent item sequence. The push discrimination branch can adopt a classification model based on logistic regression.

[0216] Assume that the consultation response frequent item sequence is (F = {f_1, f_2, cdots, f_r}), and the directional push frequent item sequence is (G = {g_1, g_2, cdots, g_s}). The elements in these two sequences are combined and feature engineered to obtain a feature vector (vec{x}). For example, the number of occurrences of frequent items, the combination relationship between different frequent items, etc. can be converted into numerical features.

[0217] For a logistic regression model, for example, the model parameters are (theta = (theta_0, theta_1, cdots, theta_t)), the basic formula of logistic regression is (h_{theta}(vec{x}) = frac{1}{1+e^{-theta^Tvec{x}}});

[0218] Where (theta^Tvec{x}=theta_0+theta_1x_1+cdots+theta_tx_t). The formula (h_{theta}(vec{x})) is a value between (0) and (1), indicating the probability of a certain push label (e.g., different types of push labels such as "high-value push" or "low-value push").

[0219] Determine push label information: According to the probability value calculated by the logistic regression model, set a threshold (for example, (0.5)) to determine the push label information. If (h_{theta}(vec{x})>0.5), the push label is determined as a certain category (such as "high-value push"); if (h_{theta}(vec{x})<0.5), it is determined as another category (such as "low-value push").

[0220] Finally, the targeted user portrait information of the consultation response event data sequence is determined based on the push tag information.

[0221] The basis for constructing targeted user portrait information: Push tag information reflects certain characteristics of the consultation response event data. Based on these push tag information, the targeted user portrait information of the consultation response event data sequence can be determined. Targeted user portrait information is a comprehensive description of the consultant (or user group), including user characteristics related to computing resource consultation responses and targeted push.

[0222] For example, if the push tag information indicates that a user's consultation response events are often marked as "high-value pushes", then the targeted user portrait information may include characteristics such as the user is a high-demand, high-value user, may have in-depth business needs for computing resources, and may be a potential long-term customer.

[0223] A cluster analysis method can be used to cluster consultation and response event data with similar push tag information, with each cluster representing a type of user. For example, using the K-Means clustering algorithm, set the number of cluster centers (K=3). First, randomly initialize (K) cluster centers (mu_1, mu_2, mu_3), and then for each consultation and response event data, calculate its distance to each cluster center (such as the Euclidean distance (d(x, mu)=sqrt{sum_{i=1}^{n}(x_i-mu_i)^2})), and assign it to the cluster with the closest distance. Then recalculate the center of each cluster, and repeat this process until the cluster center no longer changes significantly. The users in each cluster have similar directional user portrait information.

[0224] With this design, first of all, by constructing the consultation response event data sequence and the targeted push event data sequence, various data related to the computing power resource session response can be systematically integrated, providing a comprehensive data foundation for subsequent analysis. The residual processing branch in the resource session analysis network is used to obtain the frequent item sequence. The residual network structure ensures the effectiveness of deep data processing and avoids the problems in traditional deep network training. The frequent item mining algorithms such as Apriori can mine the frequent patterns in the data. These frequent items are very helpful for understanding the common elements in consultation responses and targeted pushes. The push discrimination branch accurately determines the push label information based on the logistic regression model. This classification method helps to classify the consultation response events in a targeted manner. Finally, the targeted user portrait information is constructed based on the push label information. Through methods such as cluster analysis, users can be better classified and understood, thereby providing valuable user insights for computing power resource service providers and helping to formulate more accurate marketing strategies and service optimization plans.

[0225] Further, Figure 2 Schematic diagram of the structure of a conversation response processing system 200 provided in this application. Figure 2 The illustrated conversation response processing system 200 includes a processor 210, which can call and run a computer program from a memory to implement the method of the present application.

[0226] Alternatively, if Figure 2 As shown, the conversation response processing system 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the present application.

[0227] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0228] Alternatively, if Figure 2 As shown, the conversation response processing system 200 may further include a transceiver 220. The processor 210 may control the transceiver 220 to interact with other devices. Specifically, it may send information or data to other devices, or receive information or data sent by other devices.

[0229] Optionally, the session response processing system 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the present application, which will not be described in detail here for the sake of brevity.

[0230] It should be understood that the processor of the present application may be an integrated circuit chip with signal processing capabilities.

[0231] It is understood that the memory in the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0232] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0233] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0234] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0235] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the present application, all of which are within the protection of the present application.

Claims

1. A method for processing computing resource session responses for deep learning, characterized in that: The method is applied to a conversation response processing system, and the method comprises: Acquire the computing power resource session data to be processed, identify the target session interaction knowledge in the computing power resource session data to be processed, and obtain the target session interaction knowledge information set; Mining the current session interaction knowledge vector corresponding to the target session interaction knowledge information set, and obtaining the initial associated session interaction knowledge vector corresponding to each computing power resource business keyword, wherein the initial associated session interaction knowledge vector is obtained by performing knowledge vector encoding on each associated session interaction knowledge information set corresponding to the computing power resource business keyword, and performing cross collision on each associated session interaction knowledge code; Determine, according to correlation feature values ​​of the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the computing power resource business keywords, a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector from the associated session interaction knowledge vectors corresponding to the computing power resource business keywords; The computing power resource business keywords corresponding to the target-associated session interaction knowledge vector are used as the session interaction business subject keywords of the corresponding target session interaction knowledge information set, and the computing power resource session response results corresponding to the computing power resource session data to be processed are determined based on the session interaction business subject keywords of the target session interaction knowledge information set.

2. The method according to claim 1, characterized in that Before acquiring the computing power resource session data to be processed, identifying the target session interaction knowledge in the computing power resource session data to be processed, and obtaining the target session interaction knowledge information set, the method further includes: Obtain each associated session interaction knowledge information set corresponding to the set computing resource business keyword; Performing knowledge vector encoding on each of the associated session interaction knowledge information sets to obtain each of the associated session interaction knowledge encodings; The associated session interaction knowledge codes are cross-collided to obtain an initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword.

3. The method according to claim 2, characterized in that The cross-collision of the associated session interaction knowledge codes to obtain the initial associated session interaction knowledge vector corresponding to the set computing resource business keyword includes: Combining the associated conversation interaction knowledge codes to obtain a conversation interaction knowledge combination code; Full connection processing is performed according to the session interaction knowledge combination coding to obtain the initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword.

4. The method according to claim 2, characterized in that The method further comprises: Loading each of the associated session interaction knowledge information sets into the initial AI knowledge embedding network; Performing knowledge vector encoding on each of the associated session interaction knowledge information sets through a knowledge vector encoding module in the initial AI knowledge embedding network to obtain each associated session interaction knowledge encoding; The knowledge cross-collision module in the initial AI knowledge embedding network cross-collides the respective associated session interaction knowledge codes to obtain an initial associated session interaction knowledge vector corresponding to the set computing power resource business keyword; The debugging method of embedding the initial AI knowledge into the network includes: Obtaining each computing resource session data sample and the corresponding prior session response training annotation, and loading each computing resource session data sample into the original deep learning network; Performing knowledge vector encoding on each computing resource session data sample through the initial AI knowledge embedding network in the original deep learning network to obtain each computing resource session knowledge vector sample; Perform cross-collision on the knowledge vector samples of each computing resource session through the initial knowledge cross-collision module in the original deep learning network to obtain a cross-collision sample vector; Performing a conversation response output on the cross-collision sample vector through an initial conversation response output module in the original deep learning network to obtain a conversation response prediction result; Optimizing the original deep learning network according to the prior conversation response training annotations and the conversation response prediction results to obtain an intermediate deep learning network; The intermediate deep learning network is used as the original deep learning network, and the steps of obtaining the session data samples of each computing resource and the corresponding prior session response training annotations, and loading the session data samples of each computing resource into the original deep learning network are implemented in a loop until the debugging process is completed to obtain the target deep learning network; The initial AI knowledge embedding network is obtained according to the target knowledge vector encoding module and the target knowledge cross-collision module in the target deep learning network.

5. The method according to claim 1, characterized in that The identifying target session interaction knowledge in the to-be-processed computing resource session data to obtain a target session interaction knowledge information set includes: Mining the session interaction semantic vector of the computing resource session data to be processed, performing feature preprocessing according to the session interaction semantic vector, and obtaining a session interaction semantic preprocessing vector; The conversation interaction semantic preprocessing vector is subjected to conversation requirement decoding to obtain a conversation requirement decoding vector, and information is collected based on the conversation requirement decoding vector to obtain the target conversation interaction knowledge information set.

6. The method according to claim 1, characterized in that The determining, based on correlation feature values ​​of the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, from the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector comprises: Determine the feature commonality scores of the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the computing resource business keywords, respectively, to obtain respective correlation feature values; A target correlation feature value is determined from the respective correlation feature values, and an associated session interaction knowledge vector corresponding to the target correlation feature value is used as a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector.

7. The method according to claim 1, characterized in that The computing power resource business keywords include computing power scheduling business keywords and cloud computing power mode update keywords; The determining, based on correlation feature values ​​of the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, from the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector comprises: Determine a feature commonality score between the current session interaction knowledge vector and an associated session interaction knowledge vector corresponding to the computing power scheduling service keyword to obtain a computing power scheduling correlation feature value; Determine a feature commonality score between the current session interaction knowledge vector and an associated session interaction knowledge vector corresponding to the cloud computing power mode update keyword to obtain the mode update correlation feature value; Determine, according to a distinguishing variable between the computing power scheduling correlation feature value and the mode update correlation feature value, a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector from the associated session interaction knowledge vector corresponding to the computing power scheduling service keyword and the associated session interaction knowledge vector corresponding to the cloud computing power mode update keyword; The determining, based on the difference variable between the computing power scheduling correlation feature value and the mode update correlation feature value, from the associated session interaction knowledge vector corresponding to the computing power scheduling service keyword and the associated session interaction knowledge vector corresponding to the cloud computing power mode update keyword, a target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector includes: Determine a distinguishing variable between the computing power scheduling correlation characteristic value and the mode update correlation characteristic value, and when the distinguishing variable is greater than a set variable value, perform a comparative analysis on the computing power scheduling correlation characteristic value and the mode update correlation characteristic value; When the computing power scheduling correlation feature value is greater than the mode update correlation feature value, taking the associated session interaction knowledge vector corresponding to the computing power scheduling service keyword as the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector; When the mode update correlation feature value is greater than the computing power scheduling correlation feature value, taking the associated session interaction knowledge vector corresponding to the cloud computing power mode update keyword as the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector; When the distinguishing variable is not greater than the set variable value, the general business theme keyword is used as the conversation interaction business theme keyword of the target conversation interaction knowledge information set.

8. The method according to claim 1, characterized in that The target session interaction knowledge information set includes at least two; the determining of the computing power resource session response result corresponding to the computing power resource session data to be processed according to the session interaction service subject keyword of the target session interaction knowledge information set includes: Determine the thermal evaluation coefficients of the at least two target session interaction knowledge information sets and the computing power resource session data to be processed, and use the session interaction business subject keywords of the target session interaction knowledge information set with the largest thermal evaluation coefficient as the computing power resource session response result corresponding to the computing power resource session data to be processed.

9. The method according to claim 1, characterized in that The method further comprises: Loading the target session interaction knowledge information set and the associated session interaction knowledge vectors corresponding to the computing power resource business keywords into a bidirectional long short-term memory network; The knowledge vector encoding module in the bidirectional long short-term memory network mines the current session interaction knowledge vector corresponding to the target session interaction knowledge information set, and according to the correlation feature values ​​of the current session interaction knowledge vector and the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, the target associated session interaction knowledge vector corresponding to the current session interaction knowledge vector is determined from the associated session interaction knowledge vectors corresponding to the respective computing power resource business keywords, and the computing power resource business keyword corresponding to the target associated session interaction knowledge vector is used as the session interaction business theme keyword corresponding to the target session interaction knowledge set; The debugging method of the bidirectional long short-term memory network includes: Obtaining conversation interaction knowledge information set samples and corresponding scene conversation response training annotations; Loading the associated session interaction knowledge vectors corresponding to the session interaction knowledge information set samples and the computing power resource business keywords into the original bidirectional long short-term memory network to perform session keyword analysis, and obtaining session interaction business prediction keywords corresponding to the session interaction knowledge information set samples; Optimizing the original bidirectional long short-term memory network according to the conversation interaction service prediction keywords and the scene conversation response training annotations to obtain an optimized bidirectional long short-term memory network; The optimized bidirectional long short-term memory network is used as the original bidirectional long short-term memory network, and the step of jumping to the step of obtaining the conversation interaction knowledge information set sample and the corresponding scene conversation response training annotation is implemented in a loop until the debugging process is completed to obtain the bidirectional long short-term memory network.

10. A conversation response processing system, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images