Multi-scene text intelligent recommendation and automatic generation method and system based on AI

Abstract

The invention provides an AI-based multi-scene text intelligent recommendation and automatic generation method and system, and relates to the technical field of text generation, and the method comprises the steps: carrying out the entity extraction of recruitment demand information to form a structured feature, constructing a hierarchical semantic vector through a capsule dynamic routing network, and carrying out the automatic generation of a multi-scene text. Key features are selected in combination with a multi-layer attention mechanism and an interactive fusion mechanism, texts are generated through a reserve pool network and a layered decoder, and meanwhile user feedback is introduced for dynamic adjustment. According to the method, the text generation precision and the individuation degree are improved, and the cross-scene adaptive capacity is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of text generation technology, and in particular to an AI-based multi-scenario text intelligent recommendation and automatic generation method and system. Background Art

[0002] With the advancement of informatization in human resource management, intelligent recruitment text generation and recommendation technologies are gaining increasing attention. In traditional recruitment processes, the writing of recruitment texts often relies on manual experience, which is time-consuming, labor-intensive, and of varying quality. In recent years, with the development of artificial intelligence (AI), AI-based text generation and recommendation systems have been gradually applied to the recruitment field. By learning and analyzing historical recruitment texts, they assist HR professionals in efficiently generating recruitment texts that meet specific requirements. Currently, mainstream technologies are primarily based on methods such as natural language processing and machine learning. By understanding and analyzing recruitment requirements, they retrieve similar cases from historical data or directly generate new recruitment texts.

[0003] However, existing technologies still have obvious shortcomings in the intelligent recommendation and generation of recruitment texts. Existing methods have a rather one-sided understanding of recruitment needs, often only focusing on surface keyword matching and lacking understanding of deep semantics and industry-specific knowledge, resulting in a large deviation between recommendation results and actual needs; existing recommendation systems usually use simple feature extraction and matching algorithms, which make it difficult to capture the hierarchical structure and multi-dimensional features in recruitment texts, making the generated content lack professionalism and pertinence; existing technologies lack effective quality assessment and optimization mechanisms, and cannot be dynamically adjusted based on user feedback, resulting in the system being difficult to continuously improve recommendation quality and unable to adapt to the specific requirements of different industries and positions. Summary of the Invention

[0004] The embodiments of the present invention provide an AI-based multi-scenario text intelligent recommendation and automatic generation method and system, which can solve the problems in the prior art.

[0005] A first aspect of an embodiment of the present invention provides an AI-based multi-scenario text intelligent recommendation and automatic generation method, comprising:

[0006] Perform word segmentation and named entity recognition on the recruitment demand information to extract job description entity information and form structured demand features; perform similarity matching on the structured demand features in the historical recruitment text database to screen and obtain an initial candidate text set;

[0007] A capsule dynamic routing network is used to build a multi-level feature extraction structure, which performs hierarchical feature extraction and dynamic combination on the initial candidate text set to generate a hierarchical semantic vector.

[0008] Based on hierarchical semantic vectors, feature importance is calculated by variance contribution and feature selection is performed in combination with a multi-layer attention mechanism. An interactive fusion mechanism is established between correlation features and quality assessment features to obtain a fused feature vector.

[0009] The fused feature vector is used to construct a reserve pool network for state iteration and sampling processing, and the layered decoder is used to achieve the conversion from feature space to text space to obtain the initial recommended text;

[0010] The initial recommendation text is normatively scored and optimized and rewritten to generate a final recommendation text; and the network is dynamically adjusted based on user feedback information of the final recommendation text to achieve iterative optimization of model parameters.

[0011] In an optional embodiment, a capsule dynamic routing network is used to construct a multi-level feature extraction structure, hierarchical feature extraction and dynamic combination are performed on the initial candidate text set, and the hierarchical semantic vector is generated, including:

[0012] Inputting the initial candidate text set into the underlying capsule network, the underlying capsule network uses a bidirectional GRU structure to map the initial candidate text set into an initial capsule vector group, each capsule vector in the initial capsule vector group contains position encoding information and semantic encoding information;

[0013] The information entropy eigenvalue is obtained by calculating the inner product distribution of the initial capsule vector group and the capsule vector mean center, and an adaptive weight function is constructed based on the information entropy eigenvalue to dynamically adjust the coupling coefficient between adjacent capsule layers to obtain a dynamic coupling coefficient;

[0014] Mapping the initial capsule vector group to a high-dimensional feature space using a transformation matrix, and performing feature reorganization in combination with the dynamic coupling coefficient to obtain a high-level capsule vector, while establishing a residual transmission path between the high-level capsule vector and the initial capsule vector group;

[0015] Nonlinear activation fusion is performed on the high-level capsule vector and the initial capsule vector transmitted through the residual transmission path to obtain a fused capsule vector, and the fused capsule vector is input into the next layer of capsule network. After recursive iterative processing, a hierarchical semantic vector is output.

[0016] In an optional embodiment, an information entropy eigenvalue is obtained by calculating the inner product distribution of the initial capsule vector group and the capsule vector mean center, and an adaptive weight function is constructed based on the information entropy eigenvalue to dynamically adjust the coupling coefficient between adjacent capsule layers. The dynamic coupling coefficient includes:

[0017] Perform an inner product operation on each capsule vector in the initial capsule vector group and the mean center of the capsule vector of the current batch to obtain the capsule vector inner product. The capsule vector inner product is converted into a conditional probability distribution through a probability normalization function. The information entropy eigenvalue of the initial capsule vector group is obtained based on the cross entropy calculation of the conditional probability distribution.

[0018] Constructing an adaptive weight function based on the information entropy eigenvalue, performing a negative exponential transformation on the information entropy eigenvalue through a learnable scaling factor and adjusting it through a normalization coefficient to obtain an adaptive weight coefficient;

[0019] Performing a dot product operation on the current layer capsule vector of the initial capsule vector group and the next layer capsule vector to obtain a capsule vector routing probability, performing accumulation and normalization operations on the capsule vector routing probability to obtain an initial routing coefficient;

[0020] Multiplying the adaptive weight coefficient by the initial routing coefficient to obtain a dynamic coupling coefficient, and performing weighted combination on the capsule vector of the current layer to obtain a routing fusion feature;

[0021] The ratio of the square of the modulus length of the routing fusion feature to the unit value is calculated to obtain a modulus length ratio, the modulus length ratio is multiplied by the unit vector of the routing fusion feature to obtain a normalized feature, and the normalized feature is processed by a nonlinear compression function to obtain a capsule vector of the next layer.

[0022] In an optional embodiment, based on the hierarchical semantic vector, feature importance is calculated by variance contribution and feature selection is performed in combination with a multi-layer attention mechanism. An interactive fusion mechanism is established between the correlation feature and the quality assessment feature. The fused feature vector obtained includes:

[0023] Calculate the variance contribution of each dimension feature of the hierarchical semantic vector, determine the feature importance coefficient based on the variance contribution, and generate a feature importance matrix;

[0024] Performing a three-layer attention calculation on the feature importance matrix, generating an attention weight distribution by multiplying a query matrix, a key value matrix, and a value matrix, and weightedly recombining the features according to the attention weight distribution to obtain a correlation feature vector;

[0025] Selectively transfer features of the correlation feature vector, determine the propagation probability through the temporal dependency between features, calculate the feature importance score based on the propagation probability, and dynamically adjust the control parameters. After multi-level attention calculation, a quality assessment feature vector is obtained.

[0026] Mapping the correlation feature vector and the quality assessment feature vector to a unified feature space through a feature transformation matrix, calculating the cosine distance between the feature vectors to obtain a cross-task attention coefficient, and performing feature fusion based on the cross-task attention coefficient to obtain an interaction feature vector;

[0027] The interactive feature vector is optimized by using an adversarial training method to obtain a fused feature vector.

[0028] In an optional embodiment, selective feature transfer is performed on the correlation feature vector, the propagation probability is determined by the temporal dependency between the features, the feature importance score is calculated based on the propagation probability, and the control parameters are dynamically adjusted. The quality assessment feature vector obtained through multi-level attention calculation includes:

[0029] Calculate the conditional entropy between the features of each dimension for the correlation feature vector, determine the feature dependency strength, establish a connection relationship between the feature nodes whose feature dependency strength is greater than the preset strength threshold, and generate a feature dependency adjacency matrix;

[0030] Based on the feature dependency adjacency matrix, combined with the temporal arrangement order of historical features, the feature propagation probability is calculated by feature dependency strength and temporal distance to obtain a temporal propagation matrix;

[0031] Using the time series propagation matrix, the propagation probability of the new input feature and the historical feature is weighted to obtain the feature importance score;

[0032] According to the feature importance scores, the input control parameters are dynamically adjusted through a forward mapping mechanism, and the forgetting control parameters are dynamically adjusted through a reverse mapping mechanism to form a dynamic selection sequence;

[0033] The dynamic selection sequence is input into a multi-layer adaptive weight adjustment unit, and the feature importance score and the propagation probability are weightedly fused to generate a quality assessment feature vector.

[0034] In an optional embodiment, a reservoir network is constructed for the fused feature vectors, and state iteration and sampling processing are performed. The conversion from feature space to text space is achieved through a layered decoder, and the initial recommended text is obtained, including:

[0035] Constructing a local density distribution matrix of the fused feature vector, establishing a reservoir network using the local density distribution matrix, and performing neuron activation processing on the reservoir network to obtain state weights;

[0036] Inputting the fused feature vector into the reservoir network to form an initial state, constructing a historical state record matrix through iterative calculation and extracting a stable state interval to obtain a reservoir state matrix; performing state node screening and connectivity verification on the reservoir state matrix to obtain a sampling trajectory;

[0037] Constructing a text feature space and establishing a feature-text mapping dictionary, inputting the sampling trajectory into a hierarchical decoder, and obtaining a feature mapping result through semantic dependency extraction and adversarial optimization;

[0038] The feature mapping results are screened to obtain candidate texts, and echo sampling parameters are adjusted according to the semantic relevance between the candidate texts and the fused feature vector to obtain an initial recommended text.

[0039] In an optional embodiment, the fused feature vector is input into the reservoir network to form an initial state, a historical state record matrix is ​​constructed through iterative calculation, and a stable state interval is extracted to obtain a reservoir state matrix; state node screening and connectivity verification are performed on the reservoir state matrix to obtain a sampling trajectory including:

[0040] Input the fused feature vector into the reservoir network, establish the corresponding mapping relationship, calculate the activation state value of each network node, and generate the initial network state;

[0041] Constructing a state history record matrix, storing the initial network state in the first row of the state history record matrix, iteratively calculating the initial network state based on the recursive connection weight matrix, and sequentially storing the network state obtained in each iteration in the state history record matrix until a preset number of iterations are completed;

[0042] Calculating state difference values ​​between adjacent state rows in the state history matrix, determining a stable state interval based on the state difference values, and extracting network states corresponding to the stable state intervals to construct a reserve pool state matrix;

[0043] Selecting a state node whose response intensity exceeds a preset intensity threshold in the reserve pool state matrix, generating a state node index sequence, and extracting corresponding state values ​​from the reserve pool state matrix according to the state node index sequence to construct a sampling sequence;

[0044] The state connectivity of adjacent sampling points in the sampling sequence is verified, and the sampling points that do not meet the state transition conditions are eliminated, and the sampling points are arranged in time sequence to form a sampling trajectory.

[0045] A second aspect of an embodiment of the present invention provides an AI-based multi-scenario text intelligent recommendation and automatic generation system, comprising:

[0046] The first unit is used to perform word segmentation and named entity recognition on the recruitment demand information, extract the job description entity information, and form structured demand features; perform similarity matching on the structured demand features in the historical recruitment text database to screen and obtain an initial candidate text set;

[0047] The second unit is used to build a multi-level feature extraction structure using a capsule dynamic routing network, perform hierarchical feature extraction and dynamic combination on the initial candidate text set, and generate a hierarchical semantic vector;

[0048] The third unit is used to calculate feature importance based on hierarchical semantic vectors through variance contribution and combine it with a multi-layer attention mechanism for feature selection. It also establishes an interactive fusion mechanism between correlation features and quality assessment features to obtain a fused feature vector.

[0049] The fourth unit is used to perform state iteration and sampling processing on the reservoir network constructed by the fused feature vector, and realize the conversion from feature space to text space through the layered decoder to obtain the initial recommended text;

[0050] The fifth unit is used to perform normative scoring and optimized rewriting on the initial recommendation text to generate a final recommendation text; dynamically adjust the network according to user feedback information on the final recommendation text to achieve iterative optimization of model parameters.

[0051] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0056] In an embodiment of the present invention, by performing word segmentation and named entity recognition processing on recruitment demand information, combined with a multi-level feature extraction structure constructed by a capsule dynamic routing network, a deep understanding and precise expression of text semantics is achieved, the relevance and accuracy of the recommended text are effectively improved, and the problem of shallow semantic understanding of traditional text recommendation methods is solved; a feature selection method combining variance contribution calculation with a multi-layer attention mechanism is adopted to establish an interactive fusion mechanism between correlation features and quality assessment features, which significantly enhances the quality and pertinence of the recommended text and makes the generated text more in line with the needs of specific scenarios; through the state iteration and hierarchical decoder technology of the reserve pool network, combined with the normative scoring and optimized rewriting mechanism, and the dynamic adjustment mechanism based on user feedback, the adaptive optimization of the text generation process is achieved, the practicality and acceptability of the recommended text are guaranteed, and the overall performance of the system and user satisfaction are greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flowchart of an AI-based multi-scenario text intelligent recommendation and automatic generation method according to an embodiment of the present invention;

[0058] Figure 2 Comparison chart of the impact of different scaling factors on model performance;

[0059] Figure 3 This is a visualization diagram of the time series propagation matrix. DETAILED DESCRIPTION

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0061] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 Schematic diagram of the process of the AI-based multi-scenario text intelligent recommendation and automatic generation method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] Perform word segmentation and named entity recognition on the recruitment demand information to extract job description entity information and form structured demand features; perform similarity matching on the structured demand features in the historical recruitment text database to screen and obtain an initial candidate text set;

[0064] A capsule dynamic routing network is used to build a multi-level feature extraction structure, which performs hierarchical feature extraction and dynamic combination on the initial candidate text set to generate a hierarchical semantic vector.

[0065] Based on hierarchical semantic vectors, feature importance is calculated by variance contribution and feature selection is performed in combination with a multi-layer attention mechanism. An interactive fusion mechanism is established between correlation features and quality assessment features to obtain a fused feature vector.

[0066] The fused feature vector is used to construct a reserve pool network for state iteration and sampling processing, and the layered decoder is used to achieve the conversion from feature space to text space to obtain the initial recommended text;

[0067] The initial recommendation text is normatively scored and optimized and rewritten to generate a final recommendation text; and the network is dynamically adjusted based on user feedback information of the final recommendation text to achieve iterative optimization of model parameters.

[0068] In an optional embodiment, a capsule dynamic routing network is used to construct a multi-level feature extraction structure, hierarchical feature extraction and dynamic combination are performed on the initial candidate text set, and the generation of a hierarchical semantic vector includes:

[0069] Inputting the initial candidate text set into the underlying capsule network, the underlying capsule network uses a bidirectional GRU structure to map the initial candidate text set into an initial capsule vector group, each capsule vector in the initial capsule vector group contains position encoding information and semantic encoding information;

[0070] The information entropy eigenvalue is obtained by calculating the inner product distribution of the initial capsule vector group and the capsule vector mean center, and an adaptive weight function is constructed based on the information entropy eigenvalue to dynamically adjust the coupling coefficient between adjacent capsule layers to obtain a dynamic coupling coefficient;

[0071] Mapping the initial capsule vector group to a high-dimensional feature space using a transformation matrix, and performing feature reorganization in combination with the dynamic coupling coefficient to obtain a high-level capsule vector, while establishing a residual transmission path between the high-level capsule vector and the initial capsule vector group;

[0072] Nonlinear activation fusion is performed on the high-level capsule vector and the initial capsule vector transmitted through the residual transmission path to obtain a fused capsule vector, and the fused capsule vector is input into the next layer of capsule network. After recursive iterative processing, a hierarchical semantic vector is output.

[0073] In one specific embodiment, the initial candidate text set is input into an underlying capsule network for processing. This underlying capsule network utilizes a bidirectional GRU structure, comprising two information processing channels: a forward GRU and a backward GRU. Each text sequence in the input initial candidate text set is converted into a sequence of word vectors. For example, if the text contains n words, and each word has a vector dimension of d, this results in an n×d-dimensional matrix. The bidirectional GRU structure processes this matrix in both the forward and backward directions. The forward GRU begins with the first word, sequentially reading the vector representation of each word and updating its hidden state; the backward GRU begins with the last word and processes the sequence in reverse order. For the i-th word in the text, its forward hidden state is h_forward_i, with dimensions h; its backward hidden state is h_backward_i, also with dimensions h. By concatenating these two hidden states, the complete representation of the word is obtained, [h_forward_i; h_backward_i], with dimensions 2h. Thus, for a text sequence of length n, n 2h-dimensional vectors are obtained, forming the initial capsule vector set. Each capsule vector encodes not only the semantic meaning of the word but also its position in the sequence. For example, for a text containing the phrase "artificial intelligence technology is widely used," after processing it with a bidirectional GRU, the capsule vector for the word "artificial" encodes both its semantic meaning and its position at the beginning of the sentence.

[0074] Next, we calculate the inner product distribution of the initial capsule vector group and the capsule vector mean center to obtain the information entropy eigenvalue. Specifically, we first calculate the mean vector of the initial capsule vector group, that is, the arithmetic mean of all capsule vectors. Assuming there are n initial capsule vectors v_1, v_2, …, v_n, each with dimension 2h, then the mean vector v_mean = (v_1 + v_2 + … + v_n) / n. We then calculate the inner product of each initial capsule vector with the mean vector, obtaining a set of scalar values ​​s_1, s_2, …, s_n, where s_i = v_i · v_mean. These inner product values ​​are normalized to obtain the probability distribution p_1, p_2, …, p_n, where p_i = exp(s_i) / (exp(s_1) + exp(s_2) + … + exp(s_n)). Based on this probability distribution, the information entropy H is calculated as -(p_1×log(p_1) + p_2×log(p_2) + … + p_n×log(p_n)). This information entropy eigenvalue H reflects the degree of dispersion of the initial capsule vector group. Larger values ​​indicate a more dispersed distribution, while smaller values ​​indicate a more concentrated distribution. For example, for a text containing five capsule vectors, if the calculated normalized inner product probability distribution is [0.1, 0.2, 0.4, 0.2, 0.1], its information entropy eigenvalue is approximately 1.361.

[0075] An adaptive weight function is constructed based on the information entropy eigenvalues ​​to dynamically adjust the coupling coefficients between adjacent capsule layers. A nonlinear function f(H) = 1 / (1 + exp(-a × (Hb))) is designed, where a and b are adjustable parameters. In practice, a = 2.0 and b = 1.5 are set. This function maps the information entropy to the interval (0, 1) and serves as a tuning factor. For each pair of vectors v_i and v_j in the initial capsule vector set, the initial coupling coefficient c_ij = cos(v_i, v_j) between them is calculated, which is the cosine similarity between the two vectors. The adaptive weight function is then used to adjust the dynamic coupling coefficient c'_ij = c_ij × (1 -f(H)) + f(H) × (v_i·v_j) / (||v_i|| × ||v_j||). When the information entropy is high, f(H) approaches 1, relying more on the dot product relationship between the vectors. When the information entropy is low, f(H) approaches 0, preserving more of the original cosine similarity information. For example, when the information entropy H=1.361 is calculated, substituting it into the function f(H) yields an adjustment factor of approximately 0.37. If the initial coupling coefficient of the two capsule vectors is 0.8, the adjusted dynamic coupling coefficient is approximately 0.71.

[0076] The initial capsule vector group is mapped to a high-dimensional feature space using a transformation matrix. For each initial capsule vector v_i, a transformation matrix W of dimension 2h × d' is used to map it, resulting in v'_i = v_i × W, where d' is the dimension of the high-dimensional feature space, which can be set to 128 in practice. The dynamic coupling coefficient is combined to perform feature reorganization: the high-level capsule vector u_j = sum(c'_ij × v'_i), which is the weighted sum of all transformed initial capsule vectors, with the weight being the dynamic coupling coefficient. A residual transmission path is also established between the high-level capsule vector and the initial capsule vector group. The initial capsule vector v_i is mapped to the same dimensional space as the high-level capsule vector using a linear transformation matrix W_res, resulting in a residual vector r_i = v_i × W_res.

[0077] Nonlinear activation fusion is performed on the high-level capsule vectors and the initial capsule vector transmitted through the residual transmission path. The high-level capsule vector u_j is added to the residual vector ri_i and then processed through a nonlinear activation function to obtain the fused capsule vector f_j = ReLU(u_j + ri), where ReLU is the rectified linear unit activation function, defined as max(0, x). The fused capsule vector is input into the next layer of the capsule network. After recursive iterative processing, typically with a capsule network depth of 3-5 layers, a hierarchical semantic vector is finally output. In practical applications, for a document containing 100 words, the initial vector dimension of each word is 300. After this processing, a hierarchical semantic vector with a dimension of 512 is obtained, which effectively captures the multi-level semantic structure of the text.

[0078] In an optional embodiment, the information entropy eigenvalue is obtained by calculating the inner product distribution of the initial capsule vector group and the capsule vector mean center, and an adaptive weight function is constructed based on the information entropy eigenvalue to dynamically adjust the coupling coefficient between adjacent capsule layers. The dynamic coupling coefficient includes:

[0079] Perform an inner product operation on each capsule vector in the initial capsule vector group and the mean center of the capsule vector of the current batch to obtain the capsule vector inner product. The capsule vector inner product is converted into a conditional probability distribution through a probability normalization function. The information entropy eigenvalue of the initial capsule vector group is obtained based on the cross entropy calculation of the conditional probability distribution.

[0080] Constructing an adaptive weight function based on the information entropy eigenvalue, performing a negative exponential transformation on the information entropy eigenvalue through a learnable scaling factor and adjusting it through a normalization coefficient to obtain an adaptive weight coefficient;

[0081] Performing a dot product operation on the current layer capsule vector of the initial capsule vector group and the next layer capsule vector to obtain a capsule vector routing probability, performing accumulation and normalization operations on the capsule vector routing probability to obtain an initial routing coefficient;

[0082] Multiplying the adaptive weight coefficient by the initial routing coefficient to obtain a dynamic coupling coefficient, and performing weighted combination on the capsule vector of the current layer to obtain a routing fusion feature;

[0083] The ratio of the square of the modulus length of the routing fusion feature to the unit value is calculated to obtain a modulus length ratio, the modulus length ratio is multiplied by the unit vector of the routing fusion feature to obtain a normalized feature, and the normalized feature is processed by a nonlinear compression function to obtain a capsule vector of the next layer.

[0084] In one embodiment, an initial capsule vector set is obtained. Assume that the initial capsule vector set contains n capsule vectors of dimension d, denoted as V = {v1, v2, …, vn}, where each capsule vector vi is a d-dimensional vector. The mean center μ of all capsule vectors in the current batch is calculated by summing all capsule vectors and dividing by the number of capsule vectors n, to obtain μ = (v1 + v2 + … + vn) / n.

[0085] Calculate the inner product of each capsule vector vi with the mean center μ, obtaining the inner product value si = vi·μ, forming the inner product set S = {s1, s2, …, sn}. As a specific example, suppose there are three 4-dimensional capsule vectors: v1 = [0.2, 0.3, 0.1, 0.4], v2 = [0.1, 0.5, 0.2, 0.2], v3 = [0.3, 0.1, 0.4, 0.2], calculate the mean center μ = [0.2, 0.3, 0.233, 0.267], and then calculate the inner product values ​​s1 = 0.28, s2 = 0.25, and s3 = 0.27.

[0086] The inner product set S is converted to a conditional probability distribution P = {p1, p2, …, pn} using a probability normalization function. Specifically, the softmax function is applied to each inner product value si, that is, pi = exp(si) / Σexp(sj), where j ranges from 1 to n. In the above example, after applying the softmax function, p1 = 0.36, p2 = 0.31, and p3 = 0.33.

[0087] The information entropy eigenvalue H is calculated based on the conditional probability distribution P. Specifically, H = -Σ(pi × log(pi)), where i ranges from 1 to n. In the above example, the calculated information entropy eigenvalue H = 1.097.

[0088] Next, we construct an adaptive weight function based on the information entropy eigenvalue H. We introduce a learnable scaling factor α and a normalization coefficient β, and perform a negative exponential transformation on the information entropy eigenvalue to obtain the adaptive weight coefficient w = β × exp(-α × H). Assuming α = 0.5 and β = 2, in the above example, the adaptive weight coefficient w = 2 × exp(-0.5 × 1.097) = 0.6702.

[0089] Calculate the dot product of the capsule vector of the current layer and the capsule vector of the next layer in the initial capsule vector group to obtain the capsule vector routing probability. Assuming that the current layer has n capsule vectors and the next layer has m capsule vectors, for the i-th capsule vector vi of the current layer and the j-th capsule vector uj of the next layer, calculate the dot product pij = vi·uj to form the routing probability matrix P = {pij}, where i ranges from 1 to n and j ranges from 1 to m.

[0090] The routing probability matrix P is accumulated and normalized to obtain the initial routing coefficient matrix C = {cij}. Specifically, the softmax function is applied to each row i, that is, cij = exp(pij) / Σexp(pik), where k ranges from 1 to m. As an example, assuming n = 3 and m = 2, and the routing probability matrix P = [[0.6, 0.3], [0.4, 0.7], [0.5, 0.5]], the initial routing coefficient matrix C = [[0.58, 0.42], [0.42, 0.58], [0.5, 0.5]] is obtained after applying the softmax function.

[0091] Multiply the adaptive weight coefficient w by the initial routing coefficient matrix C to obtain the dynamic coupling coefficient matrix D = {dij}, where dij = w × cij. In the above example, w = 0.6702, and the dynamic coupling coefficient matrix D = [[0.389, 0.281], [0.281, 0.389], [0.335, 0.335]].

[0092] Perform a weighted combination of the capsule vectors in the current layer to obtain the route fusion feature. For the j-th capsule vector uj in the next layer, calculate uj = Σ(dij × vi), where i ranges from 1 to n. Assuming the capsule vectors v1 = [0.2, 0.3, 0.1, 0.4], v2 = [0.1, 0.5, 0.2, 0.2], and v3 = [0.3, 0.1, 0.4, 0.2] in the current layer, use the dynamic coupling coefficient D to calculate the two capsule vectors u1 = [0.199, 0.283, 0.231, 0.271] and u2 = [0.201, 0.317, 0.235, 0.263] in the next layer.

[0093] Calculate the ratio of the square of the modulus length of the routing fusion feature to the unit value to obtain the modulus length ratio. For the fusion feature uj, calculate its square of the modulus length ||uj|| 2 , the unit value is set to 1, and the module length ratio rj = ||uj|| 2 / 1. Taking u1 as an example, ||u1|| 2 =0.199 2 + 0.283 2 + 0.2312 +0.271 2 = 0.245, module length ratio r1 = 0.245.

[0094] Multiply the modulus ratio rj by the unit vector uj / ||uj|| of the route fusion feature uj to obtain the normalized feature u'j = rj × (uj / ||uj||). Taking u1 as an example, the unit vector u1 / ||u1|| = [0.402, 0.572, 0.467, 0.548], and the normalized feature u'1 = 0.245 × [0.402, 0.572, 0.467, 0.548] = [0.098, 0.140, 0.114, 0.134].

[0095] The normalized feature u'j is processed by a nonlinear compression function to obtain the capsule vector of the next layer. The nonlinear compression function can be a squash function, that is, f(x) = (||x|| 2 / (1 + ||x||^2)) × (x / ||x||). Applying it to u'1, we get the next layer capsule vector f(u'1) = (0.245 / (1 + 0.245)) × [0.402,0.572,0.467,0.548] = 0.197 × [0.402,0.572,0.467,0.548] = [0.079,0.113,0.092,0.108].

[0096] Existing implementations primarily involve calculating the dot product of the current layer's capsule vector and the next layer's capsule vector as the routing probability, applying a softmax function to the routing probability to obtain an initial coupling coefficient, updating the coupling coefficient through multiple rounds of iteration, generating the next layer's capsule vector using a weighted sum, and applying a squash function for nonlinear transformation. The main drawback of this approach is that the routing process lacks the ability to perceive the distribution characteristics of capsule vectors. A uniform routing strategy is applied to all capsule vectors, making it impossible to adaptively adjust routing weights based on the characteristics of different input data. Furthermore, multiple rounds of iterative calculations increase computational complexity.

[0097] The method of this embodiment introduces the information entropy eigenvalue as a measurement indicator of the distribution characteristics of capsule vectors. By calculating the information entropy of the inner product distribution of capsule vectors and the mean center, the aggregation degree and distribution pattern of capsule vectors are effectively captured. An adaptive weight function based on information entropy is designed, and a learnable scaling factor and normalization coefficient are introduced to enable the routing process to dynamically adjust the weight according to the distribution characteristics of capsule vectors. The adaptive weight is combined with the initial routing coefficient to form a dynamic coupling coefficient to achieve adaptive routing for different characteristic patterns. The traditional multi-round iterative routing process is simplified and the computational complexity is reduced.

[0098] Traditional routing algorithms use the same routing strategy for different input data and are unable to adaptively adjust based on data characteristics. In actual applications, different feature patterns may require different routing strategies. By introducing information entropy eigenvalues, the method of this embodiment can perceive the distribution characteristics of capsule vectors, thereby achieving adaptive adjustment of the routing process; improve the adaptability of the capsule network to different feature patterns, enabling the network to dynamically adjust the routing strategy based on the characteristics of the input data; enhance the representation ability of the model, especially when processing complex scenarios and diverse data; improve the generalization performance of the model, showing better recognition accuracy and stability on test data; reduce computational complexity, and compared with traditional multi-round iterative routing algorithms, the method of this embodiment is more efficient; and achieve automatic optimization of the routing process through learnable parameters, reducing the need for manual parameter adjustment.

[0099] like Figure 2 As shown in the figure, the impact of the scaling factor on the model performance in this technical solution is demonstrated. The scaling factor is a key parameter for converting information entropy eigenvalues ​​into adaptive weights. The chart shows the comprehensive impact of the scaling factor on the accuracy, information entropy weight coefficient, and training time. As can be seen from the data, for all three datasets (MNIST, CIFAR-10, and Fashion-MNIST), the accuracy first increases and then decreases with the increase of the scaling factor, and all reach the best value when the scaling factor is 2.0. Specifically, on the MNIST dataset, when the scaling factor is 2.0, the accuracy reaches a maximum value of 97.82%; on the CIFAR-10 dataset, it reaches 94.65%; and on the Fashion-MNIST dataset, it reaches 92.18%. At the same time, the information entropy weight coefficient increases monotonically with the increase of the scaling factor, from 0.283 at a scaling factor of 0.5 to 0.768 at a scaling factor of 5.0, indicating that a larger scaling factor will enhance the influence of the information entropy eigenvalue on the routing process. Training time also increases with the scaling factor, from 39.7 milliseconds per batch at a scaling factor of 0.5 to 60.6 milliseconds per batch at a scaling factor of 5.0, a 52.6% increase. This phenomenon indicates that excessively large scaling factors increase computational complexity and may also lead to excessive concentration of routing weights on certain capsules, reducing the model's generalization ability. Conversely, too small a scaling factor fails to fully utilize the information provided by the information entropy eigenvalues, resulting in a weak adaptive weighting effect. Experimental results clearly demonstrate that a scaling factor of 2.0 is an ideal balance point, achieving the optimal balance between accuracy and computational efficiency. This finding provides an important reference for parameter configuration in practical applications and also reveals the inherent correlation between the information entropy eigenvalues ​​and routing performance.

[0100] In an optional embodiment, based on the hierarchical semantic vector, feature importance is calculated by variance contribution and feature selection is performed in combination with a multi-layer attention mechanism. An interactive fusion mechanism is established between the correlation feature and the quality assessment feature. The fused feature vector obtained includes:

[0101] Calculate the variance contribution of each dimension feature of the hierarchical semantic vector, determine the feature importance coefficient based on the variance contribution, and generate a feature importance matrix;

[0102] Performing a three-layer attention calculation on the feature importance matrix, generating an attention weight distribution by multiplying a query matrix, a key value matrix, and a value matrix, and weightedly recombining the features according to the attention weight distribution to obtain a correlation feature vector;

[0103] Selectively transfer features of the correlation feature vector, determine the propagation probability through the temporal dependency between features, calculate the feature importance score based on the propagation probability, and dynamically adjust the control parameters. After multi-level attention calculation, a quality assessment feature vector is obtained.

[0104] Mapping the correlation feature vector and the quality assessment feature vector to a unified feature space through a feature transformation matrix, calculating the cosine distance between the feature vectors to obtain a cross-task attention coefficient, and performing feature fusion based on the cross-task attention coefficient to obtain an interaction feature vector;

[0105] The interactive feature vector is optimized by using an adversarial training method to obtain a fused feature vector.

[0106] In a specific embodiment, the input hierarchical semantic vector is processed and the variance contribution of each dimensional feature is calculated. Assuming that the input hierarchical semantic vector is a 128-dimensional vector set, the variance of each dimensional feature in the overall data distribution is calculated. For example, for the i-th dimension feature, its variance is 0.0025, and the sum of the variances of all dimensional features is 0.05, then the variance contribution of this dimension is 0.0025 / 0.05=0.05, indicating that the contribution of this dimension to the overall data distribution is 5%. Similarly, the variance contributions of all dimensions are calculated to form a contribution vector. The feature importance coefficient is set according to the variance contribution. For example, when the variance contribution is greater than 0.01, the feature is assigned an importance coefficient of 1.2; when the variance contribution is between 0.005 and 0.01, it is assigned an importance coefficient of 1.0; when the variance contribution is less than 0.005, it is assigned an importance coefficient of 0.8. In this way, a feature importance matrix of size 128×128 is generated.

[0107] A three-layer attention calculation is performed on the feature importance matrix. In the first layer of attention calculation, the feature importance matrix is ​​converted into a query matrix, a key-value matrix, and a numerical matrix respectively. Specifically, these three matrices are obtained by multiplying the feature importance matrix by three different weight matrices. Assuming that the weight matrix size is 128×64, the generated query matrix, key-value matrix, and numerical matrix are all 128×64 matrices. The query matrix is ​​multiplied by the transpose of the key-value matrix to obtain a 128×128 attention score matrix, where each element represents the strength of association between different features. The matrix is ​​normalized and a softmax function can be used to make the sum of each row 1 to obtain the attention weight distribution. The attention weight distribution is multiplied by the numerical matrix to achieve weighted reorganization of the features and obtain a correlation feature vector of size 128×64.

[0108] Selective feature transfer is performed on correlation feature vectors to construct temporal dependencies between features. Assume that feature states from the past five time steps are selected and the temporal correlation between features is calculated. When the feature similarity between two time steps exceeds 0.85, the propagation probability is set to 0.9; when the similarity is between 0.7 and 0.85, the propagation probability is set to 0.7; and when the similarity is below 0.7, the propagation probability is set to 0.5. Based on these propagation probabilities, an importance score is calculated for each feature. For example, if the importance scores of a feature over five time steps are [0.8, 0.75, 0.82, 0.79, 0.77], the weighted average of these scores, 0.786, is taken as the dynamic importance score for the feature. The control parameter is adjusted based on the dynamic importance score: when the score is above 0.8, the control parameter is set to 1.2; when the score is between 0.6 and 0.8, the control parameter is set to 1.0; and when the score is below 0.6, the control parameter is set to 0.8. The adjusted features are processed through a multi-level attention network, which consists of 3 attention layers and 8 attention heads in each layer to obtain the quality assessment feature vector.

[0109] The correlation feature vector and the quality assessment feature vector are mapped to a unified feature space using a feature transformation matrix. Assuming the feature transformation matrix is ​​64×32, the two feature vectors are mapped into 32-dimensional vectors using this matrix. The cosine distance between these two 32-dimensional vectors is calculated. If the value is 0.92, it indicates that the two feature vectors have high similarity. The cross-task attention coefficient is determined based on the cosine distance value. For example, when the cosine distance is greater than 0.9, the attention coefficient is 0.6; when it is between 0.7 and 0.9, the attention coefficient is 0.5; and when it is less than 0.7, the attention coefficient is 0.4. Using the cross-task attention coefficient α = 0.6, the fusion feature is calculated: 0.6 × correlation feature vector + 0.4 × quality assessment feature vector, resulting in the interaction feature vector.

[0110] The interaction feature vector is optimized using adversarial training. A discriminator network is constructed, consisting of three fully connected layers with hidden layer sizes of 64, 32, and 16, respectively, to distinguish between real and generated features. The interaction feature vector is iteratively optimized through an adversarial process that minimizes the generator loss and maximizes the discriminator loss. During training, the discriminator is updated five times and the generator once per iteration. The learning rate is set to 0.001. After 200 iterations, the discriminator accuracy stabilizes at around 52%, indicating that the generated features are difficult to distinguish from real features. The optimized interaction feature vector obtained at this point is the final fused feature vector.

[0111] In an optional embodiment, selective feature transfer is performed on the correlation feature vector, the propagation probability is determined by the temporal dependency between the features, the feature importance score is calculated based on the propagation probability, and the control parameters are dynamically adjusted. The quality assessment feature vector obtained through multi-level attention calculation includes:

[0112] Calculate the conditional entropy between the features of each dimension for the correlation feature vector, determine the feature dependency strength, establish a connection relationship between the feature nodes whose feature dependency strength is greater than the preset strength threshold, and generate a feature dependency adjacency matrix;

[0113] Based on the feature dependency adjacency matrix, combined with the temporal arrangement order of historical features, the feature propagation probability is calculated by feature dependency strength and temporal distance to obtain a temporal propagation matrix;

[0114] Using the time series propagation matrix, the propagation probability of the new input feature and the historical feature is weighted to obtain the feature importance score;

[0115] According to the feature importance scores, the input control parameters are dynamically adjusted through a forward mapping mechanism, and the forgetting control parameters are dynamically adjusted through a reverse mapping mechanism to form a dynamic selection sequence;

[0116] The dynamic selection sequence is input into a multi-layer adaptive weight adjustment unit, and the feature importance score and the propagation probability are weightedly fused to generate a quality assessment feature vector.

[0117] In a specific embodiment, the conditional entropy between the dimensions of the correlation feature vector is calculated to determine the feature dependency strength. For each pair of features in the correlation feature vector, the degree of mutual dependence between features X and Y is evaluated by calculating the conditional entropy H(X|Y) and H(Y|X). The conditional entropy H(X|Y) represents the uncertainty of feature X when feature Y is known. During the calculation, a joint probability distribution table can be constructed using historical sample data, and the conditional entropy value can be calculated based on the probability distribution. For example, for word frequency features and semantic relevance features in text features, the conditional entropy can be calculated by statistically analyzing the distribution of these two features in historical text samples. When the conditional entropy values ​​of two features are low, it indicates that there is a strong dependency relationship between them.

[0118] After determining the dependency strength between each feature pair, a dependency strength threshold (such as 0.7) is set, and feature nodes with dependency strengths greater than the threshold are connected to generate a feature dependency adjacency matrix A. The element aij in matrix A represents the dependency strength between feature i and feature j. If the dependency strength exceeds the threshold, aij is the actual dependency strength value; otherwise, aij is 0. Taking a text recommendation system as an example, if the dependency strength between the word frequency feature and the user click-through rate feature is 0.85, exceeding the preset threshold of 0.7, then 0.85 is recorded at the corresponding position in the adjacency matrix; while the dependency strength between the word frequency feature and the text length feature is 0.6, which is below the threshold, it is recorded as 0.

[0119] Based on the feature dependency adjacency matrix and the temporal order of historical features, the feature propagation probability is calculated and the temporal propagation matrix P is constructed. Temporal distance is defined as the relative position difference between features in the time series, such as the temporal distance between adjacent features is 1. The feature propagation probability is determined by the functional relationship between feature dependency strength and temporal distance and can be calculated using a decay function: the propagation probability pij of feature i to feature j is equal to the feature dependency strength aij multiplied by the temporal decay factor, which can be set as an exponential function of the product of the negative temporal distance of e and the decay coefficient λ. λ can be adjusted according to the actual application scenario and is typically between 0.1 and 0.5. In a text recommendation system, if the dependency strength between title features and content features is 0.8, the temporal distance is 2, and λ is 0.2, the calculated propagation probability is approximately 0.67.

[0120] Using the time series propagation matrix P, the propagation probabilities of new input features and historical features are weighted to obtain the feature importance score S. For a newly input feature vector x, the importance score Si of feature i is calculated as the weighted sum of the new feature value xi and the propagation influence of historical features. Specifically, Si equals xi plus the sum of the influences of all historical features j on feature i, where each influence term is the historical feature value xj multiplied by the propagation probability pji. In practical applications, the time window size can be set (e.g., the last 30 historical records), and only historical features within the window are considered. Taking intelligent text generation as an example, if the feature value of a newly input keyword is 0.9, the propagation probability of the historical context feature is 0.6, and the historical context feature value is 0.8, then the importance score of the keyword is 0.9 + (0.8 × 0.6) = 1.38.

[0121] Based on the feature importance score S, the input control parameter α is dynamically adjusted through a forward mapping mechanism, while the forgetting control parameter β is dynamically adjusted through a reverse mapping mechanism, forming a dynamic selection sequence. The forward mapping mechanism uses a nonlinear function to map the feature importance score to the input control parameter α in the range [0, 1]. For example, the Sigmoid function is used: α equals 1 divided by (1 plus the exponential product of the negative feature importance score and the adjustment coefficient e). The reverse mapping mechanism calculates the forgetting control parameter β based on the inverse relationship between the feature importance scores. This can be achieved using a function β equal to 1 minus α or other customized functions. In a text intelligent recommendation system, if the importance score of a user's interest feature is 1.5, using the Sigmoid function (with an adjustment coefficient of 1) yields an α of approximately 0.82, and a corresponding β of 0.18, indicating that 82% of the feature information will be retained and 18% will be forgotten.

[0122] The dynamically selected sequence is input to a multi-layer adaptive weight adjustment unit, which performs a weighted fusion of feature importance scores and propagation probabilities to generate a quality assessment feature vector. The multi-layer adaptive weight adjustment unit comprises multiple processing layers, each responsible for handling different types of feature dependencies. In the first layer, the input control parameter α is applied to weight the raw features, while the forgetting control parameter β is used to suppress redundant features. In subsequent layers, an adaptive weight adjustment mechanism is applied based on the feature dependencies at different layers, optimizing the feature representation through residual connections and gating mechanisms. The final output quality assessment feature vector contains the most valuable feature information retained after multiple layers of processing. For example, in a text generation task, if the topic relevance feature has an importance score of 1.8 and is processed by the first layer with a weight of 0.85, its weight in the quality assessment feature vector might be adjusted to 0.92 after residual connections and a second layer.

[0123] Existing technologies usually use fixed-weight feature selection methods or simple time decay models, which are difficult to adapt to complex and changing scenario requirements. Existing methods ignore the dependencies between features, do not consider the impact of temporal information on feature importance, and the feature selection strategy lacks dynamic adaptability. The method of this embodiment introduces conditional entropy to calculate feature dependencies, combines temporal information to construct a propagation matrix, and realizes dynamic selection of features through forward and reverse mapping mechanisms. The method of this embodiment enhances the context-awareness and temporal adaptability of feature selection, and improves the accuracy of feature selection by considering the dependencies between features and the temporal influence of historical information.

[0124] like Figure 3 The figure shows the temporal propagation matrix generated by this technical solution, which intuitively presents the propagation probability relationship between features in the form of a heat map. The color depth in the matrix represents the strength of the propagation probability between features, ranging from 0.03 to 0.92. The horizontal axis represents the target feature, the vertical axis represents the source feature, and the value of each cell represents the propagation probability from the source feature to the target feature. The matrix clearly shows several high propagation probability paths: the propagation probability of F2→F3 is as high as 0.92, F5→F8 reaches 0.85, F4→F5 is 0.75, and F4→F8 is 0.71, indicating that there is a strong temporal dependency between these feature pairs. At the same time, the matrix also reveals some low propagation probability paths, such as F1→F8, which is only 0.15, and F8→F1, which is only 0.03, indicating that there is almost no dependency between these features. In addition, the values ​​on the diagonal are all 0, which is consistent with the logic that a feature cannot propagate to itself. Of particular note, when F2 is used as a source feature, it has high propagation probabilities for multiple target features (F3: 0.92, F4: 0.71, F6: 0.62), indicating that F2 is a key information propagation node. Similarly, the high propagation probabilities of F5 and F4 for F8 (0.85 and 0.71, respectively) demonstrate the compound influence of multiple features on a single feature. This precise propagation probability matrix provides a solid foundation for subsequent feature importance calculation and dynamic selection sequence generation.

[0125] In an optional embodiment, the fused feature vector is used to construct a reservoir network for state iteration and sampling processing, and a layered decoder is used to implement the conversion from feature space to text space. The initial recommended text includes:

[0126] Constructing a local density distribution matrix of the fused feature vector, establishing a reservoir network using the local density distribution matrix, and performing neuron activation processing on the reservoir network to obtain state weights;

[0127] Inputting the fused feature vector into the reservoir network to form an initial state, constructing a historical state record matrix through iterative calculation and extracting a stable state interval to obtain a reservoir state matrix; performing state node screening and connectivity verification on the reservoir state matrix to obtain a sampling trajectory;

[0128] Constructing a text feature space and establishing a feature-text mapping dictionary, inputting the sampling trajectory into a hierarchical decoder, and obtaining a feature mapping result through semantic dependency extraction and adversarial optimization;

[0129] The feature mapping results are screened to obtain candidate texts, and echo sampling parameters are adjusted according to the semantic relevance between the candidate texts and the fused feature vector to obtain an initial recommended text.

[0130] In a specific embodiment, when constructing the local density distribution matrix of the fused feature vector, user historical behavior data and content feature data are collected, user preference feature vectors and content semantic feature vectors are extracted, and the two types of feature vectors are fused through the attention mechanism to obtain a fused feature vector. The local density calculation is performed on the fused feature vector, 500 reference points in the feature space are selected, the Euclidean distance between each reference point and each dimension of the fused feature vector is calculated, and a 500×128-dimensional local density distribution matrix is ​​constructed. Taking an e-commerce user as an example, the preference feature vector generated by his historical purchasing behavior is 64-dimensional, and the product content feature vector is 64-dimensional. After fusion, a 128-dimensional fused feature vector is obtained. The 500 reference points selected for the local density calculation are distributed around this vector in the feature space.

[0131] When building a reservoir network using the local density distribution matrix, a reservoir containing 1,000 neurons was designed. The connection weights between neurons were constructed based on the local density distribution, with a connection sparsity of 0.2, meaning that each neuron connected to an average of 200 other neurons. A nonlinear activation function, the hyperbolic tangent function, was applied to the reservoir network, with an activation threshold of 0.3. When the neuron input exceeded the threshold, it activated and generated a state weight. In practice, neurons corresponding to feature dimensions with strong user preferences (such as interest in technology products) received higher activation weights.

[0132] After inputting the fused feature vector into the reservoir network to form the initial state, the number of iterations was set to 100, and the learning rate was 0.05. The neuron state was updated with each iteration and recorded in the historical state record matrix. The state change rate of the historical state record matrix was analyzed. When the state change rate for 10 consecutive iterations was less than 0.01, the network was considered to have entered a stable state interval. The stable state interval was formed from iterations 60 to 90. The state values ​​of these 31 iterations were extracted to form a 1000×31 reservoir state matrix.

[0133] The state matrix of the reservoir pool was screened for state nodes. The activation frequency and activation strength of each neuron in the stable state interval were calculated. Nodes with an activation frequency greater than 0.6 and an activation strength greater than 0.4 were selected as key nodes. The connectivity between key nodes was verified, and a connection graph was constructed to ensure that all key nodes were interconnected through paths of no more than three hops. 120 key nodes were selected, and sampling trajectories of length 20 were constructed between these nodes based on state transition probabilities. The node number sequence was recorded as [45, 78, 102, 56, 89, 201, 245, 67, 88, 156, 237, 312, 401, 188, 267, 345, 389, 412, 489, 512].

[0134] To construct the text feature space, feature representations for 10,000 common words were created based on the corpus, with each word represented by a 128-dimensional vector. A feature-text mapping dictionary was established to map the state space of the reservoir to the text feature space. Each node in the sampled trajectory was associated with a region in the text feature space using the mapping dictionary, with an association strength threshold set to 0.6. The sampled trajectory was input into a three-layer hierarchical decoder, with each layer containing 256 units. The decoder extracts semantic dependencies based on an attention mechanism. Adversarial optimization was applied to the decoded results, using a discriminator to assess text naturalness. The optimization parameters were adjusted within the range of [-0.1, 0.1]. The feature mapping results were iterated 50 times.

[0135] The feature mapping results were screened, with a semantic coherence threshold of 0.7 and a grammatical correctness threshold of 0.85 set. Texts that met both thresholds were retained as candidate texts. The semantic relevance between the candidate texts and the fused feature vector was calculated, with the relevance score ranging from [0, 1]. When the relevance was greater than 0.8, the echo sampling parameters remained unchanged. When the relevance was between [0.5, 0.8], the sampling temperature parameters were linearly increased. When the relevance was less than 0.5, resampling was performed. An example of the resulting initial recommendation was "This ultra-thin and lightweight laptop features a high-performance processor and a long-lasting battery, making it ideal for frequent business travelers. It supports fast charging and fingerprint recognition, and is available in a variety of colors."

[0136] In an optional embodiment, the fused feature vector is input into the reservoir network to form an initial state, a historical state record matrix is ​​constructed through iterative calculation, and a stable state interval is extracted to obtain a reservoir state matrix; state node screening and connectivity verification are performed on the reservoir state matrix to obtain a sampling trajectory, including:

[0137] Input the fused feature vector into the reservoir network, establish the corresponding mapping relationship, calculate the activation state value of each network node, and generate the initial network state;

[0138] Constructing a state history record matrix, storing the initial network state in the first row of the state history record matrix, iteratively calculating the initial network state based on the recursive connection weight matrix, and sequentially storing the network state obtained in each iteration in the state history record matrix until a preset number of iterations are completed;

[0139] Calculating state difference values ​​between adjacent state rows in the state history matrix, determining a stable state interval based on the state difference values, and extracting network states corresponding to the stable state intervals to construct a reserve pool state matrix;

[0140] Selecting a state node whose response intensity exceeds a preset intensity threshold in the reserve pool state matrix, generating a state node index sequence, and extracting corresponding state values ​​from the reserve pool state matrix according to the state node index sequence to construct a sampling sequence;

[0141] The state connectivity of adjacent sampling points in the sampling sequence is verified, and the sampling points that do not meet the state transition conditions are eliminated, and the sampling points are arranged in time sequence to form a sampling trajectory.

[0142] In a specific embodiment, the fused feature vector is input into the reservoir network to establish a corresponding mapping relationship. Specifically, for a given fused feature vector X(t), the vector is a multi-dimensional fused feature obtained by the previous processing, with a dimension of n, where n can be 128, 256, or 512, etc. The feature vector is multiplied by the input weight matrix Win of the reservoir network. Win has a dimension of n×m, where m is the number of nodes in the reservoir network and can be set to vary from 500 to 2000. The input signal of the reservoir network is obtained by the matrix multiplication operation of Win and X(t). For each network node i, its activation state value S(i, 0) is calculated. The activation function can use the hyperbolic tangent function so that the state value is limited to the range of [-1, 1]. Taking an actual case as an example, when the input fused feature vector X(t) is 256-dimensional and the number of reservoir nodes is 1000, the initial network state S(0) is obtained by calculation. This state is a 1000-dimensional vector, and each element represents the activation state of the corresponding node.

[0143] A state history matrix H is constructed to store the evolution of the network state. The initial network state S(0) is stored in the first row of the state history matrix H. The initial network state is then iteratively calculated based on the recursive connection weight matrix W. The dimension of the recursive connection weight matrix W is m×m, that is, 1000×1000. This matrix controls the connection strength between nodes within the reserve pool. The sparsity of the matrix W can be set to 0.1, that is, about 90% of the elements are 0, and the values ​​of the remaining non-zero elements are randomly distributed in the range of [-0.5, 0.5]. The spectral radius is controlled between 0.9 and 1.1 to ensure the dynamic characteristics of the network. For each iteration t, the calculation formula can be expressed as: multiply the current state S(t-1) by the recursive connection weight matrix W, add the input signal, and then obtain the new state S(t) through the activation function. The network state S(t) obtained in each iteration is stored in the t+1th row of the state history matrix H in turn. The iterative process continues until the preset number of iterations T is completed. T can be set between 100 and 200. Taking T=150 as an example, the final state history record matrix H is a 151×1000 matrix.

[0144] The state difference values ​​between adjacent state rows in the state history matrix are calculated to determine the stable state interval of the reservoir network. For each pair of adjacent state rows S(t) and S(t+1), the Euclidean distance or cosine similarity between them is calculated as the state difference value D(t). When multiple consecutive difference values ​​are less than a preset threshold, the network is considered to have entered a stable state. The preset threshold can be set to 0.01, and the number of consecutive states considered stable can be set to 10. By analyzing the sequence of state difference values, the starting time point ts and the ending time point te of the stable state are determined. In a real case, it may be found that the network state tends to be stable starting from the 60th iteration and continues until the 120th iteration. At this point, the stable state interval can be determined to be [60, 120]. The network states corresponding to this interval are extracted to construct the reservoir state matrix R, with dimensions of 61×1000.

[0145] In the reservoir state matrix R, state nodes whose response strength exceeds a preset strength threshold are selected to generate a state node index sequence. Response strength can be measured by calculating the root mean square (RMS) or maximum absolute value of each node's state value within the stable interval. The preset strength threshold is set to 0.5. Each column of the reservoir state matrix is ​​traversed. If the response strength of the corresponding node in that column exceeds the threshold, the index of that node is added to the state node index sequence I. Assuming that the response strength of 200 nodes exceeds the threshold, the state node index sequence I contains the index values ​​of these 200 nodes. Based on the state node index sequence I, the corresponding state values ​​are extracted from the reservoir state matrix R to construct a sampling sequence S_sample with a dimension of 61×200.

[0146] The state connectivity of adjacent sampling points in the sampling sequence is verified, and sampling points that do not meet the state transition conditions are removed. The state transition condition can be defined as the transition probability between two adjacent state points being greater than a threshold, or the state change meeting a predefined smoothness condition. The state transition condition is set to the Euclidean distance between adjacent state points being less than 0.2. Starting from the first time point in the sampling sequence, each pair of adjacent sampling points is checked to see if they meet the transition condition. If not, the next sampling point is removed. After state connectivity verification, the retained sampling points are arranged in chronological order to form the final sampling trajectory T_final. In a real-world case, 15 points that do not meet the conditions may be removed from the original 61 sampling points, resulting in a sampling trajectory consisting of 46 sampling points.

[0147] Traditional reserve pool calculation models often directly use the entire network state for subsequent processing, without considering the differences in state stability and node contribution, resulting in wasted computing resources and unstable results. In the prior art, a fixed number of iterations or a simple threshold judgment is usually used to determine the calculation termination point, lacking in-depth analysis of the dynamic characteristics of the network. The method of this embodiment accurately identifies the stable state interval by constructing a state history record matrix and analyzing the state difference value; screens valid nodes by setting a response strength threshold to reduce redundant calculations; and ensures the consistency and validity of the sampling trajectory through state connectivity verification. This enables the system to adaptively determine the optimal calculation interval, improves the accuracy and robustness of the feature mapping, and reduces computing resource consumption.

[0148] The AI-based multi-scenario text intelligent recommendation and automatic generation system of the embodiment of the present invention includes:

[0149] The first unit is used to perform word segmentation and named entity recognition on the recruitment demand information, extract the job description entity information, and form structured demand features; perform similarity matching on the structured demand features in the historical recruitment text database to screen and obtain an initial candidate text set;

[0150] The second unit is used to build a multi-level feature extraction structure using a capsule dynamic routing network, perform hierarchical feature extraction and dynamic combination on the initial candidate text set, and generate a hierarchical semantic vector;

[0151] The third unit is used to calculate feature importance based on hierarchical semantic vectors through variance contribution and combine it with a multi-layer attention mechanism for feature selection. It also establishes an interactive fusion mechanism between correlation features and quality assessment features to obtain a fused feature vector.

[0152] The fourth unit is used to perform state iteration and sampling processing on the reservoir network constructed by the fused feature vector, and realize the conversion from feature space to text space through the layered decoder to obtain the initial recommended text;

[0153] The fifth unit is used to perform normative scoring and optimized rewriting on the initial recommendation text to generate a final recommendation text; dynamically adjust the network according to user feedback information on the final recommendation text to achieve iterative optimization of model parameters.

[0154] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0155] processor;

[0156] a memory for storing processor-executable instructions;

[0157] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0158] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0159] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. AI-based multi-scenario text intelligent recommendation and automatic generation method, characterized by: include: Perform word segmentation and named entity recognition on recruitment demand information, extract job description entity information, and form structured demand features; Performing similarity matching on the structured demand features in a historical recruitment text database to obtain an initial candidate text set; A capsule dynamic routing network is used to build a multi-level feature extraction structure, which performs hierarchical feature extraction and dynamic combination on the initial candidate text set to generate a hierarchical semantic vector. Based on hierarchical semantic vectors, feature importance is calculated by variance contribution and feature selection is performed in combination with a multi-layer attention mechanism. An interactive fusion mechanism is established between correlation features and quality assessment features to obtain a fused feature vector. The fused feature vector is used to construct a reserve pool network for state iteration and sampling processing, and the layered decoder is used to achieve the conversion from feature space to text space to obtain the initial recommended text; The initial recommendation text is normatively scored and optimized and rewritten to generate a final recommendation text; and the network is dynamically adjusted based on user feedback information of the final recommendation text to achieve iterative optimization of model parameters.

2. The method according to claim 1, characterized in that A capsule dynamic routing network is used to build a multi-level feature extraction structure, which performs hierarchical feature extraction and dynamic combination on the initial candidate text set to generate a hierarchical semantic vector including: Inputting the initial candidate text set into the underlying capsule network, the underlying capsule network uses a bidirectional GRU structure to map the initial candidate text set into an initial capsule vector group, each capsule vector in the initial capsule vector group contains position encoding information and semantic encoding information; The information entropy eigenvalue is obtained by calculating the inner product distribution of the initial capsule vector group and the capsule vector mean center, and an adaptive weight function is constructed based on the information entropy eigenvalue to dynamically adjust the coupling coefficient between adjacent capsule layers to obtain a dynamic coupling coefficient; Mapping the initial capsule vector group to a high-dimensional feature space using a transformation matrix, and performing feature reorganization in combination with the dynamic coupling coefficient to obtain a high-level capsule vector, while establishing a residual transmission path between the high-level capsule vector and the initial capsule vector group; Nonlinear activation fusion is performed on the high-level capsule vector and the initial capsule vector transmitted through the residual transmission path to obtain a fused capsule vector, and the fused capsule vector is input into the next layer of capsule network. After recursive iterative processing, a hierarchical semantic vector is output.

3. The method according to claim 2, characterized in that The information entropy eigenvalue is obtained by calculating the inner product distribution of the initial capsule vector group and the capsule vector mean center. Based on the information entropy eigenvalue, an adaptive weight function is constructed to dynamically adjust the coupling coefficient between adjacent capsule layers. The dynamic coupling coefficient is obtained as follows: Perform an inner product operation on each capsule vector in the initial capsule vector group and the mean center of the capsule vector of the current batch to obtain the capsule vector inner product. The capsule vector inner product is converted into a conditional probability distribution through a probability normalization function. The information entropy eigenvalue of the initial capsule vector group is obtained based on the cross entropy calculation of the conditional probability distribution. Constructing an adaptive weight function based on the information entropy eigenvalue, performing a negative exponential transformation on the information entropy eigenvalue through a learnable scaling factor and adjusting it through a normalization coefficient to obtain an adaptive weight coefficient; Performing a dot product operation on the current layer capsule vector of the initial capsule vector group and the next layer capsule vector to obtain a capsule vector routing probability, performing accumulation and normalization operations on the capsule vector routing probability to obtain an initial routing coefficient; Multiplying the adaptive weight coefficient by the initial routing coefficient to obtain a dynamic coupling coefficient, and performing weighted combination on the capsule vector of the current layer to obtain a routing fusion feature; The ratio of the square of the modulus length of the routing fusion feature to the unit value is calculated to obtain a modulus length ratio, the modulus length ratio is multiplied by the unit vector of the routing fusion feature to obtain a normalized feature, and the normalized feature is processed by a nonlinear compression function to obtain a capsule vector of the next layer.

4. The method according to claim 1, wherein Based on the hierarchical semantic vector, the feature importance is calculated by variance contribution and the multi-layer attention mechanism is combined for feature selection. An interactive fusion mechanism is established between the correlation features and the quality assessment features. The fused feature vector obtained includes: Calculate the variance contribution of each dimension feature of the hierarchical semantic vector, determine the feature importance coefficient based on the variance contribution, and generate a feature importance matrix; Performing a three-layer attention calculation on the feature importance matrix, generating an attention weight distribution by multiplying a query matrix, a key value matrix, and a value matrix, and weightedly recombining the features according to the attention weight distribution to obtain a correlation feature vector; Selectively transfer features of the correlation feature vector, determine the propagation probability through the temporal dependency between features, calculate the feature importance score based on the propagation probability, and dynamically adjust the control parameters. After multi-level attention calculation, a quality assessment feature vector is obtained. Mapping the correlation feature vector and the quality assessment feature vector to a unified feature space through a feature transformation matrix, calculating the cosine distance between the feature vectors to obtain a cross-task attention coefficient, and performing feature fusion based on the cross-task attention coefficient to obtain an interaction feature vector; The interactive feature vector is optimized by using an adversarial training method to obtain a fused feature vector.

5. The method according to claim 4, characterized in that Selective feature transfer is performed on the correlation feature vector, the propagation probability is determined by the temporal dependency between the features, the feature importance score is calculated based on the propagation probability, and the control parameters are dynamically adjusted. The quality assessment feature vector obtained through multi-level attention calculation includes: Calculate the conditional entropy between the features of each dimension for the correlation feature vector, determine the feature dependency strength, establish a connection relationship between the feature nodes whose feature dependency strength is greater than the preset strength threshold, and generate a feature dependency adjacency matrix; Based on the feature dependency adjacency matrix, combined with the temporal arrangement order of historical features, the feature propagation probability is calculated by feature dependency strength and temporal distance to obtain a temporal propagation matrix; Using the time series propagation matrix, the propagation probability of the new input feature and the historical feature is weighted to obtain the feature importance score; According to the feature importance scores, the input control parameters are dynamically adjusted through a forward mapping mechanism, and the forgetting control parameters are dynamically adjusted through a reverse mapping mechanism to form a dynamic selection sequence; The dynamic selection sequence is input into a multi-layer adaptive weight adjustment unit, and the feature importance score and the propagation probability are weightedly fused to generate a quality assessment feature vector.

6. The method according to claim 1, wherein The fused feature vector is used to construct a reserve pool network for state iteration and sampling processing. The layered decoder is used to achieve the conversion from feature space to text space. The initial recommended texts include: Constructing a local density distribution matrix of the fused feature vector, establishing a reservoir network using the local density distribution matrix, and performing neuron activation processing on the reservoir network to obtain state weights; Inputting the fused feature vector into the reservoir network to form an initial state, constructing a historical state record matrix through iterative calculation and extracting a stable state interval to obtain a reservoir state matrix; performing state node screening and connectivity verification on the reservoir state matrix to obtain a sampling trajectory; Constructing a text feature space and establishing a feature-text mapping dictionary, inputting the sampling trajectory into a hierarchical decoder, and obtaining a feature mapping result through semantic dependency extraction and adversarial optimization; The feature mapping results are screened to obtain candidate texts, and echo sampling parameters are adjusted according to the semantic relevance between the candidate texts and the fused feature vector to obtain an initial recommended text.

7. The method according to claim 6, characterized in that The fused feature vector is input into the reservoir network to form an initial state. A historical state record matrix is ​​constructed through iterative calculation and a stable state interval is extracted to obtain a reservoir state matrix. The reservoir state matrix is ​​subjected to state node screening and connectivity verification to obtain a sampling trajectory including: Input the fused feature vector into the reservoir network, establish the corresponding mapping relationship, calculate the activation state value of each network node, and generate the initial network state; Constructing a state history record matrix, storing the initial network state in the first row of the state history record matrix, iteratively calculating the initial network state based on the recursive connection weight matrix, and sequentially storing the network state obtained in each iteration in the state history record matrix until a preset number of iterations are completed; Calculating state difference values ​​between adjacent state rows in the state history matrix, determining a stable state interval based on the state difference values, and extracting network states corresponding to the stable state intervals to construct a reserve pool state matrix; Selecting a state node whose response intensity exceeds a preset intensity threshold in the reserve pool state matrix, generating a state node index sequence, and extracting corresponding state values ​​from the reserve pool state matrix according to the state node index sequence to construct a sampling sequence; The state connectivity of adjacent sampling points in the sampling sequence is verified, and the sampling points that do not meet the state transition conditions are eliminated, and the sampling points are arranged in time sequence to form a sampling trajectory.

8. An AI-based multi-scenario text intelligent recommendation and automatic generation system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to perform word segmentation and named entity recognition on recruitment demand information, extract job description entity information, and form structured demand features; Performing similarity matching on the structured demand features in a historical recruitment text database to obtain an initial candidate text set; The second unit is used to build a multi-level feature extraction structure using a capsule dynamic routing network, perform hierarchical feature extraction and dynamic combination on the initial candidate text set, and generate a hierarchical semantic vector; The third unit is used to calculate feature importance based on hierarchical semantic vectors through variance contribution and combine it with a multi-layer attention mechanism for feature selection. It also establishes an interactive fusion mechanism between correlation features and quality assessment features to obtain a fused feature vector. The fourth unit is used to perform state iteration and sampling processing on the reservoir network constructed by the fused feature vector, and realize the conversion from feature space to text space through the layered decoder to obtain the initial recommended text; The fifth unit is used to perform normative scoring and optimized rewriting on the initial recommendation text to generate a final recommendation text; dynamically adjust the network according to user feedback information on the final recommendation text to achieve iterative optimization of model parameters.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

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

Images