Innovation and entrepreneurship coaching question-answer matching method and system based on semantic understanding

Abstract

The present invention provides a method and system for matching innovation and entrepreneurship counseling questions and answers based on semantic understanding, which relates to the technical field of natural language processing. First, the semantic understanding features of the question sentences input by the user are extracted, and the corresponding knowledge node set is matched based on the preset innovation and entrepreneurship counseling knowledge base. Then, a semantic association path is generated according to the hierarchical relationship and semantic association of the knowledge node set, and the target question and answer strategy is determined and optimized counseling content is generated accordingly. Finally, according to the user's feedback data on the optimized counseling content, the node connection weights and association relationships in the innovation and entrepreneurship counseling knowledge base are updated. Thus, through deep semantic understanding and dynamic optimization of the knowledge base, accurate matching of innovation and entrepreneurship counseling questions and answers and personalized content generation are achieved, providing efficient and high-quality counseling support for innovative entrepreneurs.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to a method and system for matching questions and answers in innovation and entrepreneurship coaching based on semantic understanding. Background Art

[0002] In the field of innovation and entrepreneurship education and coaching, with the rise of the innovation and entrepreneurship boom, a large number of entrepreneurs and potential entrepreneurs are increasingly demanding personalized and precise coaching. Traditional innovation and entrepreneurship coaching often relies on manual consultation, standardized courses, or simple keyword search question-and-answer systems. While manual consultation offers a certain degree of specificity, it is limited by the varying energy and experience of consultants, making it difficult to ensure standardized and timely service. Standardized courses struggle to adapt to the individual needs and questions of different entrepreneurs. Simple keyword search question-and-answer systems can only match literal content and fail to understand the deeper meaning of user questions. This can lead to irrelevant answers, redundant or missing information, and ultimately fail to provide truly effective coaching support for innovative entrepreneurs. Therefore, there is an urgent need for a method that can deeply understand user semantics, accurately match knowledge, generate adaptive coaching content, and continuously optimize the knowledge system based on user feedback to improve the efficiency and quality of innovation and entrepreneurship coaching. Summary of the Invention

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for matching innovation and entrepreneurship coaching questions and answers based on semantic understanding, the method comprising:

[0004] Extract semantic understanding features of the question sentence input by the user;

[0005] Matching a set of knowledge nodes corresponding to the semantic understanding features based on a preset innovation and entrepreneurship coaching knowledge base;

[0006] Generating a semantic association path according to the hierarchical relationship and semantic association of the knowledge node set;

[0007] Determine the target question-answering strategy based on the semantic association path and generate optimized tutoring content;

[0008] The node connection weights and association relationships in the innovation and entrepreneurship coaching knowledge base are updated according to the user's feedback data on the optimized coaching content.

[0009] In a possible implementation of the first aspect, extracting semantic understanding features of the question sentence input by the user includes:

[0010] Performing word segmentation on the question sentence, dividing it into multiple semantic word units and removing stop words from the multiple semantic word units;

[0011] Calling a pre-trained semantic encoding model to perform context encoding on the semantic word unit to generate a context vector for each semantic word unit;

[0012] Input the context vector into the multi-head attention layer to calculate the semantic association weight between each semantic word unit;

[0013] Performing weighted fusion on the context vectors based on the semantic association weights to generate a global semantic feature of the question sentence;

[0014] Input the global semantic features into the feature dimensionality reduction layer to extract core intent features and auxiliary logic features as the semantic understanding features; wherein the core intent features are used to represent the core objectives of the user's question, and the auxiliary logic features are used to describe the dependency relationship and semantic hierarchy between multiple semantic word units in the question sentence;

[0015] The semantic coding model is trained by the following steps:

[0016] Collect sample question sentences and their corresponding annotated intent labels from historical innovation and entrepreneurship coaching scenarios to construct a training dataset;

[0017] The parameters of the semantic encoding model are optimized with the goal of minimizing the cross entropy loss between the predicted intent label and the annotated intent label of the sample question sentence.

[0018] In a possible implementation of the first aspect, calling a pre-trained semantic encoding model to perform context encoding on the semantic word units to generate a context vector for each semantic word unit includes:

[0019] Inputting the semantic word unit into a bidirectional long short-term memory network to obtain a forward hidden state sequence and a backward hidden state sequence of each semantic word unit;

[0020] Concatenate the forward hidden state sequence and the backward hidden state sequence of the same semantic word unit to generate the initial context vector;

[0021] Performing layer normalization on the initial context vector, and fusing the normalized context vector with the initial context vector through a residual connection to obtain an enhanced context vector;

[0022] The enhanced context vector is input into a nonlinear transformation layer to generate a context vector set with uniform dimension.

[0023] In a possible implementation of the first aspect, the matching of the knowledge node set corresponding to the semantic understanding feature based on the preset innovation and entrepreneurship coaching knowledge base includes:

[0024] Traversing the candidate knowledge nodes in the innovation and entrepreneurship coaching knowledge base, calling a node matching model to calculate a first similarity score between a node feature of each candidate knowledge node and the core intent feature;

[0025] Filtering candidate knowledge nodes whose first similarity scores are higher than a first preset threshold to form an initial matching set;

[0026] Calling a logic verification model to analyze whether the node description texts of the candidate knowledge nodes in the initial matching set satisfy the dependency relationship and semantic hierarchy corresponding to the auxiliary logical features; if not, determining them as conflicting nodes and eliminating the conflicting nodes to obtain candidate knowledge nodes after logic verification;

[0027] The candidate knowledge nodes after logical verification are arranged in descending order according to the first similarity score, and the top N nodes are selected as the optimized knowledge node set.

[0028] In a possible implementation of the first aspect, the node matching model training method includes:

[0029] Extract positive sample knowledge nodes from historical successfully matched question and answer data, and record the user question semantic understanding features corresponding to the positive sample knowledge nodes;

[0030] Processing the positive sample knowledge nodes to generate negative sample knowledge nodes, specifically including: retaining the hierarchical attributes and associated labels of the positive sample knowledge nodes, replacing the core keywords of the positive sample knowledge nodes to generate semantically contradictory node description texts, or selecting nodes from the same tutoring field that are semantically irrelevant to the user's question;

[0031] Calling the initialized node matching model to respectively calculate the positive sample similarity between the node feature of the positive sample knowledge node and the corresponding semantic understanding feature, and the negative sample similarity between the node feature of the negative sample knowledge node and the semantic understanding feature;

[0032] A contrast loss function is constructed to maximize the difference between the similarity of positive samples and the similarity of negative samples, and the parameters of the initialized node matching model are updated through a back propagation algorithm until the initialized node matching model converges to obtain a trained node matching model.

[0033] In a possible implementation of the first aspect, generating a semantic association path according to the hierarchical relationship and semantic association of the knowledge node set includes:

[0034] Obtaining hierarchical attributes, associated tags, and historical user interaction data of each node in the knowledge node set;

[0035] Constructing a vertical hierarchical connection relationship between nodes according to the hierarchical attributes;

[0036] Expanding the horizontal semantic association relationship between nodes according to the association label;

[0037] Generate an initial semantic network based on the vertical hierarchical connection relationship and the horizontal semantic association relationship;

[0038] The path generation algorithm is called to traverse the initial semantic network to generate multiple candidate semantic paths. Each candidate semantic path is scored based on the path length, the jump probability between nodes, the node semantic weight and the user feedback index, combined with the historical user interaction data, and the candidate semantic path with the highest score is selected as the target semantic association path.

[0039] In a possible implementation of the first aspect, the call path generation algorithm traverses the initial semantic network to generate multiple candidate semantic paths, scores each candidate semantic path based on path length, inter-node jump probability, node semantic weight, and user feedback index, combined with the historical user interaction data, and selects the candidate semantic path with the highest score as the target semantic association path, including:

[0040] Starting from the core intent node in the knowledge node set, traverse the child nodes downward according to the vertical hierarchical connection relationship, and expand the sibling nodes according to the horizontal semantic association relationship;

[0041] During the traversal process, the visited node sequence and jump relationship are recorded to generate the initial path set;

[0042] Performing redundant path filtering on the initial path set to obtain multiple candidate semantic paths, specifically including: merging paths containing the same node sequence but different jump orders, and eliminating paths containing loop nodes or repeated jumps;

[0043] Obtaining a path length of each candidate semantic path, where the path length is obtained by counting the total number of all nodes in the candidate semantic path;

[0044] Calculate the inter-node jump probability of the candidate semantic path, which is determined based on the ratio of the number of jumps between adjacent nodes in the historical user interaction data to the total number of jumps to the starting node.

[0045] Extracting the node semantic weights of the candidate semantic paths, wherein the node semantic weights read the preset weight values of each node from the innovation and entrepreneurship counseling knowledge base, and the weight values are calculated and generated based on the node citation frequency and the semantic relevance annotated by the user;

[0046] Obtaining a user feedback index, wherein the user feedback index is obtained by counting the average satisfaction score generated by all nodes in the candidate semantic path after being clicked by users in historical sessions and the matching degree of associated subsequent questions;

[0047] The path length, inter-node jump probability, node semantic weight and user feedback index of each candidate semantic path are normalized and then weighted summed to generate a comprehensive score for each candidate semantic path;

[0048] The candidate semantic path with the highest comprehensive score is selected as the target semantic association path.

[0049] In a possible implementation of the first aspect, determining a target question-answering strategy based on the semantic association path and generating optimized tutoring content includes:

[0050] Parsing the node sequence in the semantic association path, and identifying the tutoring topic evolution logic and key knowledge point distribution corresponding to the node sequence;

[0051] Matching a basic question-and-answer strategy template from a preset question-and-answer strategy library according to the tutoring topic evolution logic, wherein the basic question-and-answer strategy template includes content generation rules, interaction process design, and resource reference method;

[0052] Dynamically adjusting the basic question-answering strategy template based on the distribution of key knowledge points to obtain an adjusted basic question-answering strategy template;

[0053] Generating the optimized coaching content based on the adjusted basic question-and-answer strategy template; the optimized coaching content includes step-by-step guidance text, reference cases adapted to user needs, and extended resource links for related knowledge points;

[0054] For example, in a possible implementation of the first aspect, parsing the node sequence in the semantic association path and identifying the tutoring topic evolution logic and key knowledge point distribution corresponding to the node sequence includes:

[0055] Obtaining node attributes of each knowledge node in the semantic association path, wherein the node attributes include node type label, historical interaction frequency, and preset semantic weight;

[0056] Dividing the node sequence into start nodes, intermediate nodes, and end nodes according to the node type labels, and extracting the hierarchical jump directions and semantic association types between adjacent nodes;

[0057] Constructing a vertical evolution branch based on the hierarchical jump direction, generating a horizontal expansion branch according to the semantic association type, and topologically merging the vertical evolution branch and the horizontal expansion branch to generate an initial topic network;

[0058] Traversing the node connection edges in the initial topic network, calculating the comprehensive strength value of each connection edge, wherein the comprehensive strength value is obtained by weighted summing the normalized value of the historical interaction frequency, the logarithmic transformation value of the preset semantic weight, and the priority coefficient of the association type;

[0059] According to the comprehensive strength value, connection edges with a strength above a threshold are selected to form a core connection edge set, and a node subsequence corresponding to the core connection edge set is extracted as a key knowledge point distribution;

[0060] Based on the node subsequence of the core connection edge set, the temporal change pattern of the node type label and the combination rule of the semantic association type are extracted to generate the tutoring topic evolution logic. The temporal change pattern is determined by analyzing the appearance order and frequency distribution of the node type labels in the node subsequence, and the combination rule is obtained by statistically analyzing the co-occurrence frequency and conditional probability of the semantic association type in the node subsequence.

[0061] In a possible implementation of the first aspect, dynamically adjusting the basic question-answering strategy template based on the key knowledge point distribution to obtain the adjusted basic question-answering strategy template includes:

[0062] Identifying a priority parameter for each knowledge node in the key knowledge point distribution, wherein the priority parameter is calculated based on the semantic weight, historical user interaction frequency, and hierarchical attributes of the knowledge node;

[0063] Dividing the key knowledge points into a core knowledge point set and an auxiliary knowledge point set according to the priority parameter, wherein the priority parameter of the core knowledge point set is higher than a preset priority threshold;

[0064] Extracting content generation rules from the basic question-answering strategy template, setting the content generation rules corresponding to the core knowledge point set as a default activation module, and generating an expansion trigger condition based on the semantic relevance of the auxiliary knowledge point set;

[0065] Analyze the unvisited knowledge point sequence in the user's historical interaction data, calculate the matching degree between the unvisited knowledge point sequence and the auxiliary knowledge point set, and select the unvisited knowledge points with a matching degree higher than a preset matching threshold as insertion nodes for supplementary cases;

[0066] Determine, based on the auxiliary logic feature in the semantic understanding feature, a logical embedding position of the insertion node of the supplementary case in the basic question-answering strategy template, wherein the logical embedding position is obtained by matching the dependency relationship in the auxiliary logic feature with the interaction process node sequence in the basic question-answering strategy template;

[0067] Integrate the parameters of the default activation module, extended trigger conditions, and logic embedding positions to generate an adjusted basic question-answering strategy template;

[0068] For example, in a possible implementation of the first aspect, matching a basic question-and-answer strategy template from a preset question-and-answer strategy library according to the tutoring topic evolution logic includes:

[0069] Extracting path attributes in the tutoring topic evolution logic, wherein the path attributes include the number of path stages, node type distribution, and semantic association combination pattern;

[0070] Converting the number of path stages into a stage division vector, converting the node type distribution into a type density matrix, and converting the semantic association combination pattern into an association coding sequence;

[0071] The strategy matching model is called to respectively calculate a first matching degree between the stage division vector and the stage label of each strategy template in the preset question-answering strategy library, a second matching degree between the type density matrix and the node compatibility matrix of each strategy template, and a third matching degree between the association coding sequence and the association rule library of each strategy template;

[0072] Normalizing the first matching degree, the second matching degree, and the third matching degree, and performing weighted fusion based on a preset weight coefficient to generate a comprehensive adaptation score for each policy template;

[0073] Filtering policy templates with comprehensive adaptation scores higher than the adaptation threshold to form a candidate policy set, and performing sorting optimization on the candidate policy set according to the temporal change pattern in the tutoring topic evolution logic, wherein the sorting optimization is achieved by matching the temporal change pattern with the stage evolution consistency in the historical application scenarios of the candidate policy templates;

[0074] The strategy template with the highest ranking in the sorted and optimized candidate strategy set is selected as the basic question-answering strategy template.

[0075] For example, in a possible implementation of the first aspect, updating the node connection weights and association relationships in the innovation and entrepreneurship coaching knowledge base based on user feedback data on the optimized coaching content includes:

[0076] Collecting user interaction behavior data with the optimization coaching content, including content click location, case review times, resource link opening rate, and subsequent question statements;

[0077] Cleaning and structuring the interaction behavior data to generate a feedback data set;

[0078] Calculating a contribution score of each knowledge node in the semantic association path according to the feedback data set;

[0079] For nodes whose contribution scores are higher than a second preset threshold, increase the connection weight between the node and the node corresponding to the core intent feature;

[0080] For nodes whose contribution scores are lower than a third preset threshold, reduce the connection weight of the node or remove invalid connections;

[0081] Based on the newly added semantic understanding features extracted from the user's subsequent question sentences, a new node is created in the innovation and entrepreneurship guidance knowledge base and an association relationship is established with the existing node;

[0082] For example, in a possible implementation of the first aspect, calculating the contribution score of each knowledge node in the semantic association path based on the feedback dataset includes:

[0083] Extracting an interaction indicator set for each knowledge node from the feedback data set, the interaction indicator set including click rate increment, case review time, resource link open rate change value and subsequent question matching degree;

[0084] Performing a time-decay weighted process on the click rate increment to generate a time-corrected click rate, wherein the time-decay weighted process is calculated based on the interval between the time when the click behavior occurred and the current time and a preset decay factor;

[0085] The case review duration is converted into a duration distribution percentile, the resource link opening rate change value is converted into a change intensity coefficient, and the subsequent question matching degree is calculated using a semantic similarity model to generate a matching score;

[0086] Normalizing the time-corrected click-through rate, duration distribution percentile, change intensity coefficient, and matching score to obtain a standardized indicator set;

[0087] Allocate the weight of each indicator in the standardized indicator set according to the preset contribution weight, and generate an initial contribution score through linear weighting;

[0088] Performing path correction on the initial contribution score based on the position attribute of the knowledge node in the semantic association path, wherein the path correction is achieved by calculating the weight coefficient of the path level where the node is located and the jump distance between the node and the core intention node;

[0089] The score after path correction is used as the final contribution score of the knowledge node.

[0090] On the other hand, an embodiment of the present invention also provides an innovation and entrepreneurship coaching question and answer matching system based on semantic understanding, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0091] Based on the above aspects, the embodiment of the present invention breaks through the limitations of traditional simple keyword matching by deeply extracting the semantic understanding features of user question sentences, and accurately grasps the core needs of users. Based on the preset innovation and entrepreneurship counseling knowledge base, the corresponding knowledge node set is matched, and the semantic association path is generated by combining the node hierarchical relationship and semantic relevance, which realizes efficient screening and accurate connection of related content from massive knowledge, constructs a knowledge context that conforms to the user's question logic, determines the target question and answer strategy based on this semantic association path, and generates optimized counseling content, ensuring that the output content is highly adapted to user needs and improving the pertinence and effectiveness of counseling. In addition, based on the user's feedback data on the optimized counseling content, the node connection weights and association relationships in the innovation and entrepreneurship counseling knowledge base are dynamically updated, so that the knowledge system continues to evolve, and the question and answer matching effect is continuously optimized, providing an intelligent and adaptive solution for innovation and entrepreneurship counseling. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 It is a schematic diagram of the execution flow of the innovation and entrepreneurship coaching question-answer matching method based on semantic understanding provided by an embodiment of the present invention.

[0093] Figure 2 Schematic diagram of exemplary hardware and software components of the innovation and entrepreneurship coaching question-and-answer matching system based on semantic understanding provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0094] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an innovation and entrepreneurship coaching question-and-answer matching method based on semantic understanding provided by an embodiment of the present invention. The innovation and entrepreneurship coaching question-and-answer matching method based on semantic understanding is introduced in detail below.

[0095] Step S110: extracting semantic understanding features of the question sentence input by the user.

[0096] In this embodiment, in the scenario of innovation and entrepreneurship counseling, users may input various question statements. For example, "In the current market background where competition is fierce and the policy environment is constantly changing, for a startup at the seed-round financing stage that mainly focuses on the innovative application of natural language processing technology in the field of artificial intelligence, how to formulate an efficient and sustainable financing and market expansion strategy while controlling capital costs and risks, combining the enterprise's own technological advantages and market positioning, and also considering the communication skills with potential investors and the possibility of subsequent enterprise strategic transformation?" In order to effectively process such question statements, it is necessary to extract their semantic understanding features.

[0097] Step S111: Perform word segmentation on the question statement, divide it into multiple semantic word units, and remove the stop words in the multiple semantic word units.

[0098] First, use a word segmentation tool to process the above question statement. The word segmentation tool can be based on various algorithms, such as rule-based word segmentation algorithms, statistics-based word segmentation algorithms, or a hybrid algorithm that combines the two. Taking the hybrid algorithm as an example, it first uses rules to perform a preliminary segmentation of the statement, and then optimizes the segmentation result through a statistical model. After word segmentation, the statement is divided into numerous semantic word units, such as "current", "fierce competition", "policy environment", "constantly changing", "market background", "seed-round financing stage", "field of artificial intelligence", "natural language processing technology", "innovative application", "startup", "control", "capital cost", "risk", "enterprise itself", "technological advantages", "market positioning", "formulate", "efficient", "sustainable", "financing", "market expansion strategy", "potential investor", "communication skills", "subsequent", "possibility of enterprise strategic transformation", etc. Then, remove the stop words in it. Stop words are usually some words that frequently appear in language expressions but contribute little to semantic understanding, such as "of", "in", "for", "而言", "how", "and", "at the same time", "also", etc. After the operation of removing stop words, a set of semantic word units with actual semantic value is obtained, denoted as M = {m1, m2, m3,..., mn}, where mi represents the i-th semantic word unit.

[0099] Step S112: Call a pre-trained semantic encoding model to perform context encoding on the semantic word units and generate the context vector of each semantic word unit.

[0100] Step S1121: Input the semantic word units into a bidirectional long short-term memory network to obtain the forward hidden state sequence and the backward hidden state sequence of each semantic word unit respectively.

[0101] The bidirectional long short-term memory (Bi-LSTM) network consists of a forward LSTM and a backward LSTM. A set of semantic word units M is sequentially input into the Bi-LSTM. The forward LSTM processes semantic word units from left to right. When processing the i-th semantic word unit mi, it updates its state based on the information from previously processed semantic word units m1 to mi-1, generating a forward hidden state ht_f. The backward LSTM processes semantic word units from right to left. When processing the i-th semantic word unit mi, it updates its state based on the information from subsequent semantic word units mi+1 to mn, generating a backward hidden state ht_b. Specifically, the forward LSTM update process can be described as follows: First, based on the current input semantic word unit mi and the previous forward hidden state ht-1_f, the input gate it_f, forget gate ft_f, cell state ct_f, and output gate ot_f are calculated. The input gate it_f determines how much information about the current semantic word unit is added to the cell state; the forget gate ft_f determines how much information from the previous cell state is retained; the cell state ct_f is updated based on the results of the input and forget gates; and the output gate ot_f determines the current hidden state ht_f. The update process of the backward LSTM is similar to that of the forward LSTM, except that the processing order is reversed. Ultimately, for each semantic word unit mi, the corresponding forward hidden state ht_f and backward hidden state ht_b are obtained, forming the forward hidden state sequence H_f = {h1_f, h2_f, …, hn_f} and the backward hidden state sequence H_b = {h1_b, h2_b, …, hn_b}.

[0102] Step S1122: Concatenate the forward hidden state sequence and the backward hidden state sequence of the same semantic word unit to generate an initial context vector.

[0103] Concatenate the hidden states corresponding to the same semantic word unit in the forward hidden state sequence H_f and the backward hidden state sequence H_b. For example, for the i-th semantic word unit mi, concatenate its forward hidden state ht_f and backward hidden state ht_b to obtain the initial context vector vt_0 = [ht_f; ht_b], where the ";" represents the concatenation operation. In this way, for all semantic word units, we can obtain the initial context vector set V_0 = {v1_0, v2_0, ..., vn_0}.

[0104] Step S1123: perform layer normalization on the initial context vector, and fuse the normalized context vector with the initial context vector through a residual connection to obtain an enhanced context vector.

[0105] Layer normalization normalizes the feature dimensions of each sample. For each vector vt_0 in the initial context vector set V_0, its mean μt and variance σt are first calculated. Then, normalization is performed using the formula vt_n = (vt_0 - μt) / σt, resulting in the normalized context vector vt_n. Next, the normalized context vector vt_n is fused with the initial context vector vt_0 via a residual connection to create an enhanced context vector vt_e = vt_0 + vt_n. Residual connections mitigate the vanishing gradient problem, enabling the model to better learn deep-level features. After this processing, the set of enhanced context vectors V_e = {v1_e, v2_e, …, vn_e} is obtained.

[0106] Step S1124: Input the enhanced context vector into the nonlinear transformation layer to generate a set of context vectors with unified dimensions.

[0107] The nonlinear transformation layer can use an activation function, such as the ReLU function. Each vector vt_e in the enhanced context vector set V_e is input into the ReLU function, whose expression is f(x) = max(0, x). Through the action of the ReLU function, the negative part of the input vector is set to 0, and the positive part is retained, thereby introducing nonlinear features. At the same time, to ensure the uniformity of the dimension of the context vector, a linear transformation can be performed after the nonlinear transformation layer. The vector is transformed through a weight matrix W and a bias vector b to obtain the final context vector vt_c = W*vt_e+b. Finally, a set of context vectors with uniform dimensions V_c = {v1_c, v2_c, …, vn_c} is obtained.

[0108] Step S113: Input the context vector into the multi-head attention layer to calculate the semantic association weights between each semantic word unit.

[0109] The multi-head attention layer consists of multiple attention heads. For each attention head, the context vector set V_c is first transformed through three linear transformation matrices W_q, W_k, and W_v, respectively, to obtain the query vector set Q, the key vector set K, and the value vector set V. For each query vector qi, the similarity score with all key vectors kj is calculated. The similarity score can be obtained through the dot product operation, that is, sij = qi * kj. The similarity score is then scaled and smoothed, divided by a scaling factor √d_k (where d_k is the dimension of the key vector), and then passed through the softmax function to obtain the attention weight αij, αij = softmax(sij / √d_k). Finally, the value vector is weighted and summed according to the attention weight to obtain the output vector oi of each attention head. The output vectors of all attention heads are concatenated and then transformed through a linear transformation matrix W_o to obtain the final output vector of the multi-head attention layer. In this process, the attention weight αij represents the semantic association weight between each semantic word unit.

[0110] Step S114: performing weighted fusion on the context vectors based on the semantic association weights to generate global semantic features of the question sentence.

[0111] According to the calculated semantic association weight αij, the context vector set V_c is weighted and fused. For each context vector vi_c, it is weighted according to the corresponding attention weight αij. The weighted fusion here is to splice the weighted context vectors according to the set rules. Specifically, each weighted context vector vi_c*αij is connected end to end in sequence to form a new vector. This new vector contains the information of all semantic word units and takes into account the semantic associations between them, and finally obtains the global semantic feature vector v_g. Through this splicing method, the feature information of each semantic word unit can be better retained, avoiding the problems of information loss or feature ambiguity that may occur during the addition process, thereby generating global semantic features that can accurately reflect the semantics of the question statement.

[0112] Step S115: Input the global semantic features into the feature dimensionality reduction layer, extract the core intent features and auxiliary logic features as semantic understanding features; among them, the core intent features are used to represent the core goals of the user's questions, and the auxiliary logic features are used to describe the dependency relationship and semantic hierarchy between multiple semantic word units in the question sentence.

[0113] The feature dimensionality reduction layer can employ methods such as principal component analysis (PCA) or autoencoders. Taking PCA as an example, the covariance matrix C of the global semantic feature vector v_g is first calculated. Eigenvalue decomposition is then performed on the covariance matrix to obtain eigenvalues and eigenvectors. The eigenvectors corresponding to the top k largest eigenvalues are selected, and the global semantic feature vector is projected into the subspace formed by these eigenvectors to obtain the reduced eigenvector. In this reduced eigenvector, some features can be identified as core intent features, while others can be identified as auxiliary logical features.

[0114] The core intent feature reflects the core objective of the user's question. For example, in the above question, the core objective might be to develop financing and market expansion strategies. The auxiliary logic feature describes the dependencies and semantic hierarchy between semantic word units. For example, there is a conditional dependency between "under the premise of controlling capital costs and risks" and "developing financing and market expansion strategies." Finally, the core intent feature is denoted as f_c, the auxiliary logic feature as f_l, and the semantic understanding feature F = {f_c, f_l}.

[0115] Step S120: Matching a knowledge node set corresponding to the semantic understanding feature based on a preset innovation and entrepreneurship coaching knowledge base.

[0116] Step S121: traverse the candidate knowledge nodes in the innovation and entrepreneurship coaching knowledge base, and call the node matching model to calculate the first similarity score between the node feature and the core intention feature of each candidate knowledge node.

[0117] The innovation and entrepreneurship guidance knowledge base contains a large number of knowledge nodes, and each knowledge node has its corresponding node feature. For each candidate knowledge node in the knowledge base, its node feature and the core intent feature f_c extracted in step S115 are input into the node matching model. The node matching model can be a neural network model based on deep learning, which learns the mapping relationship between node features and core intent features and outputs a similarity score. Assuming that the node feature of the candidate knowledge node is nk, and the output function of the node matching model is S(nk, f_c), then the first similarity score s1=S(nk, f_c).

[0118] Step S122: Filter candidate knowledge nodes whose first similarity scores are higher than a first preset threshold to form an initial matching set.

[0119] Set a first preset threshold t1, traverse the first similarity scores s1 of all candidate knowledge nodes, filter out candidate knowledge nodes with scores higher than t1, and form an initial matching set I={nk1, nk2,…, nkm}, where nki represents the i-th candidate knowledge node that meets the conditions.

[0120] Step S123: Call the logic verification model to analyze whether the node description text of the candidate knowledge node in the initial matching set satisfies the dependency relationship and semantic hierarchy corresponding to the auxiliary logical features. If not, it is determined to be a conflicting node, and the conflicting node is eliminated to obtain the candidate knowledge node after logic verification.

[0121] The logic verification model analyzes the node description text of each candidate knowledge node in the initial matching set I. The auxiliary logic feature f_l contains the dependency relationship and semantic hierarchy information between the semantic word units in the question sentence. The logic verification model checks whether the description text of the candidate knowledge node conforms to these dependency relationships and semantic hierarchy. For example, if the auxiliary logic feature indicates that there is a correlation between "financing strategy" and "market expansion strategy", and a candidate knowledge node only mentions "financing strategy" and does not reflect the correlation with "market expansion strategy", then the node may be judged as a conflict node. The conflict nodes that do not meet the conditions are removed from the initial matching set I to obtain the candidate knowledge node set I_l={nk1_l, nk2_l,…,nkp_l} after logic verification.

[0122] Step S124: Arrange the candidate knowledge nodes after the logic verification in descending order according to the first similarity score, and select the first N nodes as the optimized knowledge node set.

[0123] After logical verification, each node in the candidate knowledge node set I_1 is sorted in descending order according to its corresponding first similarity score s1. Then, the top N nodes are selected to form the optimized knowledge node set K = {nk1_k, nk2_k, ..., nkN_k}.

[0124] Step S130: Generate a semantic association path according to the hierarchical relationship and semantic association of the knowledge node set.

[0125] Step S131: Obtain the hierarchical attributes, associated tags, and historical user interaction data of each node in the knowledge node set.

[0126] For each node nki_k in the optimized knowledge node set K, its hierarchical attributes are obtained from the innovation and entrepreneurship guidance knowledge base. Hierarchical attributes can indicate the node's hierarchical position in the knowledge base, such as first-level node, second-level node, etc. At the same time, the node's associated label is obtained. The associated label describes the semantic relationship between the node and other nodes, such as labels such as "financing," "market expansion," and "risk control." Furthermore, historical user interaction data for the node is required. This historical user interaction data includes information such as the number of clicks on the node, the duration of the user's stay, and subsequent questions.

[0127] Step S132: Constructing a vertical hierarchical connection relationship between nodes according to the hierarchical attributes.

[0128] Based on the hierarchical attributes of knowledge nodes, vertical hierarchical connections are constructed between nodes. If one node is the parent node of another node, then there is a vertical connection between them. For example, in the innovation and entrepreneurship guidance knowledge base, the "financing strategy" node may be the parent node of the "seed round financing strategy" node, and a vertical hierarchical connection exists between them. In this way, a vertical hierarchical connection graph G_l is constructed.

[0129] Step S133: Expand the horizontal semantic association relationship between nodes according to the association tags.

[0130] Based on the association labels of knowledge nodes, we expand the horizontal semantic association relationships between nodes. If two nodes have the same or related association labels, then there is a horizontal semantic association between them. For example, if the "Financing Strategy" node and the "Investment Analysis" node both have association labels related to "finance," a horizontal semantic association can be established between them. In this way, a horizontal semantic association graph G_s is constructed.

[0131] Step S134: generating an initial semantic network based on the vertical hierarchical connection relationship and the horizontal semantic association relationship.

[0132] The vertical hierarchical connection graph G_l and the horizontal semantic association graph G_s are merged to generate the initial semantic network G. In the initial semantic network G, there are both vertical hierarchical connection relationships and horizontal semantic association relationships between nodes.

[0133] Step S135: Call the path generation algorithm to traverse the initial semantic network, generate multiple candidate semantic paths, and score each candidate semantic path based on path length, node jump probability, node semantic weight and user feedback indicators, combined with historical user interaction data, and select the candidate semantic path with the highest score as the target semantic association path.

[0134] Step S1351: Starting from the core intention node in the knowledge node set, traverse the child nodes downward according to the vertical hierarchical connection relationship, and expand the brother nodes according to the horizontal semantic association relationship.

[0135] Identify the core intent node within the knowledge node set K. For example, in the question above, the core intent node might be the "Financing and Market Expansion Strategy" node. Starting from this node, traverse downward through all its child nodes based on vertical hierarchical connections, while also expanding its sibling nodes based on horizontal semantic associations. During this traversal, record the sequence of nodes visited and the jump relationships, generating an initial path set P_0.

[0136] Step S1352: Filter redundant paths on the initial path set to obtain multiple candidate semantic paths.

[0137] The initial path set P_0 is filtered for redundant paths. First, paths containing the same node sequence but different jump orders are merged, as these paths essentially express the same semantic information. Second, paths containing loop nodes or repeated jumps are removed, as these paths do not conform to normal semantic logic. After filtering, a set of candidate semantic paths P = {p1, p2, …, pq} is obtained.

[0138] Step S1353: Obtain the path length of each candidate semantic path. The path length is obtained by counting the total number of all nodes in the candidate semantic path.

[0139] For each path pi in the candidate semantic path set P, the total number of nodes contained in it is counted and used as the path length li of the path.

[0140] Step S1354: Calculate the inter-node jump probability of the candidate semantic path. The inter-node jump probability is determined based on the ratio of the number of jumps between adjacent nodes in the historical user interaction data to the total number of jumps to the starting node.

[0141] Based on historical user interaction data, the number of jumps between adjacent nodes and the total number of jumps to the starting node are counted. For each pair of adjacent nodes (nj, nk) in a candidate semantic path pi, the jump probability pjk between them is calculated, where pjk = the number of jumps between adjacent nodes (nj, nk) / the total number of jumps to the starting node nj. For the entire candidate semantic path pi, the jump probabilities of all adjacent node pairs can be comprehensively considered, for example, by taking an average or weighted average to obtain the inter-node jump probability pi_p for the path.

[0142] Step S1355: Extract the node semantic weights of the candidate semantic paths. The node semantic weights read the preset weight values of each node from the innovation and entrepreneurship guidance knowledge base. The weight values are calculated based on the node citation frequency and the semantic relevance annotated by the user.

[0143] The preset weight value for each node in the candidate semantic path pi is read from the innovation and entrepreneurship guidance knowledge base. Node weights are calculated based on the node's citation frequency and the semantic relevance annotated by the user. For example, a node that is frequently cited and highly semantically relevant to the user's question will have a higher weight. The weight values of all nodes in the path are comprehensively considered, for example, by taking an average or weighted average, to obtain the node semantic weight pi_w for the path.

[0144] Step S1356: Obtain user feedback indicators. The user feedback indicators are obtained by counting the average satisfaction scores of all nodes in the candidate semantic path after being clicked by users in historical sessions and the matching degrees of associated subsequent questions.

[0145] Calculate the average satisfaction score and the matching degree of associated subsequent questions for all nodes in the candidate semantic path pi after being clicked by users in historical sessions. The average satisfaction score can be calculated based on user ratings of the node content, and the matching degree of associated subsequent questions can be calculated using a semantic similarity model. Combined with the average satisfaction score and the matching degree of associated subsequent questions, the user feedback indicator pi_f for the path is obtained.

[0146] Step S1357: The path length, inter-node jump probability, node semantic weight, and user feedback index of each candidate semantic path are normalized and then weighted summed to generate a comprehensive score for each candidate semantic path.

[0147] The path length li, inter-node jump probability pi_p, node semantic weight pi_w, and user feedback index pi_f are normalized so that they are within the same dimensional range. This normalization can be performed using common methods, such as Z-score normalization, which subtracts the mean from each feature value and divides it by its standard deviation. Let the normalized path length be li_s, the inter-node jump probability be pi_p_s, the node semantic weight be pi_w_s, and the user feedback index be pi_f_s. Then, each normalized feature is assigned a weight, denoted as wl, wp, ww, and wf, respectively. These weights can be adjusted based on actual conditions to reflect the importance of each feature in the comprehensive score. A weighted summation method is used to generate the comprehensive score si for each candidate semantic path: si = wl*li_s+wp*pi_p_s+ww*pi_w_s+wf*pi_f_s.

[0148] Step S1358: Select the candidate semantic path with the highest comprehensive score as the target semantic association path.

[0149] The comprehensive scores si of all candidate semantic paths are compared, and the candidate semantic path with the highest score is selected as the target semantic association path p_t. This target semantic association path can best reflect the hierarchical relationship and semantic association between knowledge nodes.

[0150] Step S140: Determine the target question-answering strategy based on the semantic association path and generate optimized tutoring content.

[0151] Step S141: parse the node sequence in the semantic association path, and identify the tutoring topic evolution logic and key knowledge point distribution corresponding to the node sequence.

[0152] Step S1411: Obtain the node attributes of each knowledge node in the semantic association path, where the node attributes include node type label, historical interaction frequency, and preset semantic weight.

[0153] For each knowledge node in the target semantic association path p_t, its node attributes are retrieved from the innovation and entrepreneurship guidance knowledge base. Node type labels can be categorized as "concept explanation," "strategy advice," "case analysis," and so on, describing the primary nature of the node's content. The historical interaction frequency reflects the number of times a node has been visited by users in past sessions, while the preset semantic weight reflects the node's importance within the knowledge base.

[0154] Step S1412: Divide the node sequence into start nodes, intermediate nodes, and end nodes according to the node type labels, and extract the hierarchical jump directions and semantic association types between adjacent nodes.

[0155] According to the node type label, the node sequence in the target semantic association path p_t is divided into a starting node, an intermediate node, and an ending node. The starting node is usually the node most relevant to the core intention of the user's question, the intermediate node is used to further elaborate and expand the relevant content, and the ending node provides summary or conclusive information. At the same time, the hierarchical jump direction between adjacent nodes is extracted. For example, the jump from the upper node to the lower node indicates in-depth refinement of the content, and the jump from the same-level nodes indicates horizontal association expansion. Semantic association types can include "causal association", "parallel association", "progressive association", etc., which reflect the semantic logical relationship between nodes.

[0156] Step S1413: constructing a vertical evolution branch based on the hierarchical jump direction, and generating a horizontal expansion branch based on the semantic association type, topologically merging the vertical evolution branch and the horizontal expansion branch to generate an initial topic network.

[0157] Vertical evolution branches are constructed based on the hierarchical jump directions between adjacent nodes. For example, if there is a jump from the "Financing Strategy" node to the "Seed Round Financing Strategy" node, a vertical evolution branch can be constructed. Simultaneously, horizontal expansion branches are generated based on the semantic association type. For example, if there is a parallel association between the "Financing Strategy" node and the "Market Expansion Strategy" node, a horizontal expansion branch can be constructed. The vertical evolution branches and horizontal expansion branches are topologically merged to form an initial topic network N_0 consisting of nodes and connecting edges.

[0158] Step S1414: traverse the node connection edges in the initial topic network and calculate the comprehensive strength value of each connection edge. The comprehensive strength value is obtained by weighted summing the normalized value of the historical interaction frequency, the logarithmic transformation value of the preset semantic weight, and the priority coefficient of the association type.

[0159] For each node connection edge in the initial topic network N_0, its comprehensive strength value is calculated. First, the historical interaction frequency is normalized and converted to a uniform range. Then, the preset semantic weight is logarithmically transformed to reduce the difference between the weight values. The priority coefficient of the association type is pre-set according to the importance of different semantic association types. The normalized historical interaction frequency, the preset semantic weight after logarithmic transformation, and the priority coefficient of the association type are weighted and summed to obtain the comprehensive strength value of each connection edge. Let the historical interaction frequency be f, the preset semantic weight be w, the priority coefficient of the association type be c, and the weighting coefficients be wf, ww, and wc respectively. Then the comprehensive strength value of the connection edge s_c=wf*normalization(f)+ww*logarithmic transformation(w)+wc*c.

[0160] Step S1415: Filter the connection edges with a strength above a threshold value according to the comprehensive strength value to form a core connection edge set, and extract the node subsequence corresponding to the core connection edge set as the key knowledge point distribution.

[0161] Set a strength threshold t2 and select edges with a comprehensive strength greater than t2 to form the core edge set E_c. The core edge set reflects the most important semantic relationships between nodes. Extract the node subsequences corresponding to the core edge set. These node subsequences contain the most critical knowledge points and serve as the key knowledge point distribution K_d.

[0162] Step S1416: Based on the node subsequence of the core connection edge set, extract the temporal change pattern of the node type label and the combination rule of the semantic association type to generate the tutoring topic evolution logic. The temporal change pattern is determined by analyzing the appearance order and frequency distribution of the node type label in the node subsequence, and the combination rule is obtained by statistically analyzing the co-occurrence frequency and conditional probability of the semantic association type in the node subsequence.

[0163] Analyze the node subsequences corresponding to the core edge set to extract the temporal variation patterns of node type labels. For example, a "Concept Explanation" node might appear first, followed by a "Strategy Recommendation" node, and finally a "Case Analysis" node. Simultaneously, calculate the co-occurrence frequency and conditional probability of semantic association types within the node subsequences to derive the combination rules of semantic association types. Combine the temporal variation patterns of node type labels and the combination rules of semantic association types to generate the tutoring topic evolution logic L_e.

[0164] Step S142: Match a basic question-and-answer strategy template from a preset question-and-answer strategy library according to the tutoring topic evolution logic. The basic question-and-answer strategy template includes content generation rules, interaction process design, and resource reference method.

[0165] Step S1421: extracting path attributes in the tutoring topic evolution logic, where the path attributes include the number of path stages, node type distribution, and semantic association combination pattern.

[0166] Path attributes are extracted from the tutoring topic evolution logic L_e. The number of path stages reflects the phase divisions in the tutoring topic evolution process, the node type distribution describes the distribution of different node types in the path, and the semantic association combination pattern reflects the combination of semantic association types between nodes.

[0167] Step S1422: Convert the number of path stages into a stage division vector, convert the node type distribution into a type density matrix, and convert the semantic association combination pattern into an association coding sequence.

[0168] The number of path stages is converted into a stage partition vector, where each element of the vector represents the relevant information for each stage. The node type distribution is converted into a type density matrix, where the rows represent different node types, the columns represent different stages, and the matrix elements represent the density of the node type in that stage. The semantic association combination pattern is converted into an association encoding sequence, where each element in the sequence represents the encoding of a semantic association type.

[0169] Step S1423: Call the strategy matching model to calculate the first matching degree between the stage division vector and the stage label of each strategy template in the preset question and answer strategy library, the second matching degree between the type density matrix and the node compatibility matrix of each strategy template, and the third matching degree between the associated coding sequence and the association rule library of each strategy template.

[0170] The phase division vector, type density matrix, and associated coding sequence are input into the strategy matching model. The strategy matching model calculates the first matching degree between the phase division vector and the phase label of each strategy template in the preset question-answer strategy library, the second matching degree between the type density matrix and the node compatibility matrix of each strategy template, and the third matching degree between the associated coding sequence and the association rule library of each strategy template. Let the phase division vector be v_p, the phase label of the strategy template be t_p, the type density matrix be M_t, the node compatibility matrix of the strategy template be M_c, the associated coding sequence be s_a, and the association rule library of the strategy template be R_a. Then the first matching degree m1 = strategy matching model (v_p, t_p), the second matching degree m2 = strategy matching model (M_t, M_c), and the third matching degree m3 = strategy matching model (s_a, R_a).

[0171] Step S1424: normalize the first matching degree, the second matching degree, and the third matching degree, and perform weighted fusion based on a preset weight coefficient to generate a comprehensive adaptation score for each policy template.

[0172] The first, second, and third matching degrees m1, m2, and m3 are normalized to the same dimensional range. Each normalized matching degree is then assigned a preset weight coefficient, denoted as wm1, wm2, and wm3, respectively. Through weighted fusion, a comprehensive adaptation score s_m for each policy template is generated: s_m = wm1 * normalized(m1) + wm2 * normalized(m2) + wm3 * normalized(m3).

[0173] Step S1425: Filter the policy templates whose comprehensive adaptation scores are higher than the adaptation threshold to form a candidate policy set, and sort and optimize the candidate policy set according to the temporal change pattern in the tutoring topic evolution logic. The sorting optimization is achieved by matching the temporal change pattern with the stage evolution consistency in the historical application scenarios of the candidate policy templates.

[0174] Set an adaptation threshold t3 and select policy templates with a comprehensive adaptation score higher than t3 to form the candidate policy set S_c. Then, optimize the ranking of the candidate policy set S_c based on the temporal change pattern in the tutoring topic evolution logic L_e. Match the temporal change pattern with the stage evolution consistency of the candidate policy templates in their historical application scenarios, prioritizing policy templates with high stage evolution consistency.

[0175] Step S1426: Select the highest-ranked policy template in the sorted and optimized candidate policy set as the basic question-answering policy template.

[0176] The policy templates in the sorted and optimized candidate policy set S_c are ranked, and the highest-ranked policy template is selected as the basic question-answering policy template T_b. This basic question-answering policy template contains information such as content generation rules, interaction process design, and resource reference methods.

[0177] Step S143: Dynamically adjust the basic question-answering strategy template based on the distribution of key knowledge points to obtain an adjusted basic question-answering strategy template.

[0178] Step S1431: Identify the priority parameters of each knowledge node in the key knowledge point distribution. The priority parameters are calculated based on the semantic weight, historical user interaction frequency and level attributes of the knowledge node.

[0179] For each knowledge node in the key knowledge point distribution K_d, its priority parameter is calculated. The priority parameter is calculated based on the semantic weight, historical user interaction frequency, and hierarchical attributes of the knowledge node. The semantic weight reflects the importance of the node in the knowledge base, the historical user interaction frequency reflects the popularity of the node, and the hierarchical attributes represent the hierarchical position of the node in the knowledge base. The priority parameter can be calculated by weighted summation. Let the semantic weight be w, the historical user interaction frequency be f, the hierarchical attribute be l, and the weighting coefficients be ww, wf, and wl respectively. Then the priority parameter p=ww*w+wf*f+wl*l.

[0180] Step S1432: Divide the key knowledge points into a core knowledge point set and an auxiliary knowledge point set according to the priority parameter, wherein the priority parameter of the core knowledge point set is higher than a preset priority threshold.

[0181] Set a preset priority threshold t4 and divide the key knowledge point distribution K_d into a core knowledge point set K_c and an auxiliary knowledge point set K_a based on the priority parameter p. Knowledge nodes in the core knowledge point set have a higher priority parameter than t4 and contain the most critical information. Knowledge nodes in the auxiliary knowledge point set have a lower priority parameter than t4 and are used to further supplement and expand the core knowledge points.

[0182] Step S1433: extract the content generation rules in the basic question-answering strategy template, set the content generation rules corresponding to the core knowledge point set as the default activation module, and generate extension trigger conditions based on the semantic relevance of the auxiliary knowledge point set.

[0183] Extract content generation rules from the basic question-and-answer strategy template T_b. Set the content generation rules corresponding to the core knowledge point set K_c as the default activation modules. These modules will be called first when generating optimized tutoring content. Generate expansion trigger conditions based on the semantic relevance of the auxiliary knowledge point set K_a. For example, if an auxiliary knowledge point has a causal relationship with a core knowledge point, then when the core knowledge point is mentioned and the set conditions are met, content generation for the auxiliary knowledge point is triggered.

[0184] Step S1434: Analyze the unvisited knowledge point sequence in the user's historical interaction data, calculate the matching degree between the unvisited knowledge point sequence and the auxiliary knowledge point set, and select the unvisited knowledge points with a matching degree higher than a preset matching threshold as insertion nodes for supplementary cases.

[0185] Analyze historical user interaction data to extract a sequence of unvisited knowledge points, U. Calculate the degree of match between the unvisited knowledge point sequence U and the set of auxiliary knowledge points, K_a, using a semantic similarity model. Set a preset matching threshold, t5, and select unvisited knowledge points with a matching degree higher than t5. These are inserted as nodes I_n in the supplementary case.

[0186] Step S1435: Based on the auxiliary logic features in the semantic understanding features, determine the logical embedding position of the insertion node of the supplementary case in the basic question-and-answer strategy template. The logical embedding position is obtained by matching the dependency relationship in the auxiliary logic features with the interaction process node sequence in the template.

[0187] Based on the auxiliary logic feature f_l in the semantic understanding feature F, the logical embedding position of the supplementary case insertion node I_n in the basic question-answering strategy template T_b is determined. The appropriate embedding position is found by matching the dependency relationships in the auxiliary logic feature with the sequence of interaction process nodes in the template. For example, if the auxiliary logic feature indicates that a certain knowledge point has a sequential relationship with another knowledge point, the supplementary case insertion node should be inserted at the corresponding position.

[0188] Step S1436: Integrate the parameters of the default activation module, extended trigger conditions, and logic embedding position to generate an adjusted basic question-and-answer strategy template.

[0189] The default activation module, extended trigger conditions, and logical embedding locations are parameterized to generate the adjusted basic question-answering strategy template T_a. This template dynamically adjusts based on the distribution of key knowledge points based on the basic question-answering strategy template to better meet user needs.

[0190] Step S144: Generate optimized coaching content based on the adjusted basic question-and-answer strategy template; the optimized coaching content includes step-by-step guidance text, reference cases adapted to user needs, and extended resource links for related knowledge points.

[0191] Based on the adjusted basic Q&A strategy template T_a, optimized tutoring content C is generated. This optimized tutoring content includes step-by-step guidance, providing detailed instruction based on the evolution of tutoring topics and the distribution of key knowledge points. It also includes case studies tailored to user needs, which can be drawn from historical success stories or real-world application scenarios. Furthermore, links to additional resources related to related knowledge points are provided to facilitate in-depth learning.

[0192] Step S150: updating the node connection weights and association relationships in the innovation and entrepreneurship coaching knowledge base based on the user's feedback data on the optimized coaching content.

[0193] Step S151: Collecting interactive behavior data between users and optimized coaching content, where the interactive behavior data includes content click locations, case review times, resource link opening rates, and subsequent question statements.

[0194] By monitoring users' interactions with optimized tutoring content C, we collect interaction behavior data. The click locations reflect the user's focus on different content items, the number of case studies reflects the user's interest in the case, the resource link open rate indicates the user's utilization of extended resources, and subsequent questions reflect the problems encountered and further needs of the user during the learning process.

[0195] Step S152: Clean and structure the interactive behavior data to generate a feedback data set.

[0196] The collected interactive behavior data is cleaned to remove noise and invalid data. Then, the cleaned data is structured and converted into a format that is easy to analyze to generate the feedback dataset D.

[0197] Step S153: Calculate the contribution score of each knowledge node in the semantic association path based on the feedback data set.

[0198] Step S1531: extracting the interaction indicator set of each knowledge node from the feedback data set, the interaction indicator set including click rate increment, case review time, resource link opening rate change value and subsequent question matching degree.

[0199] A set of interaction metrics for each knowledge node is extracted from the feedback dataset D. The click-through rate increment represents the increase in the click-through rate of the knowledge node in this interaction relative to the historical click-through rate. The case review duration reflects the time the user spent reviewing the case corresponding to the knowledge node. The resource link open rate change value represents the change in the open rate of the resource link corresponding to the knowledge node. The subsequent question matching degree is calculated using a semantic similarity model and reflects the semantic relevance of the user's subsequent question to the knowledge node.

[0200] Step S1532: Perform time decay weighting processing on the click rate increment to generate a time-corrected click rate. The time decay weighting processing is calculated based on the interval between the time when the click behavior occurs and the current time and a preset decay factor.

[0201] The click-through rate increment is weighted by time decay to account for the timeliness of clicks. The click-through rate increment is weighted based on the interval between the click and the current time and a preset decay factor. Let the click-through rate increment be i_c, the interval between the click and the current time be t, and the preset decay factor be α. Then, the time-corrected click-through rate i_c_t = i_c * decay function (t, α), where the decay function can be an exponential decay function, for example.

[0202] Step S1533: Convert the case review duration into a duration distribution percentile, convert the resource link opening rate change value into a change intensity coefficient, and calculate the subsequent question matching degree through a semantic similarity model to generate a matching score.

[0203] The case review duration is converted to a percentile of the duration distribution. This is done by comparing the case review duration of the knowledge node with that of all knowledge nodes to determine its percentile in the duration distribution. The change in resource link open rate is converted into a change intensity coefficient by normalizing the change value. The subsequent question matching is calculated using a semantic similarity model to generate a matching score. The semantic similarity model can learn the semantic relationship between the question statement and the knowledge node content based on a deep learning algorithm.

[0204] Step S1534: normalize the time-corrected click-through rate, duration distribution percentile, change intensity coefficient, and matching score to obtain a standardized indicator set.

[0205] Normalize the time-corrected click-through rate (i_c_t), duration distribution percentile (p_t), change intensity coefficient (c_r), and match score (s_m) to the same dimensional range. This can be done using methods such as Z-score normalization, resulting in the standardized indicator set I_s = {i_c_t_s, p_t_s, cr_r_s, s_m_s}.

[0206] Step S1535: Allocate the weight of each indicator in the standardized indicator set according to the preset contribution weight, and generate an initial contribution score through linear weighting.

[0207] Assign a preset contribution weight to each indicator in the standardized indicator set I_s, denoted as wi, wp, wc, and ws respectively. Generate the initial contribution score s_i through linear weighting, that is, s_i=wi*i_c_t_s+wp*p_t_s+wc*c_r_s+ws*s_m_s.

[0208] Step S1536: Perform path correction on the initial contribution score based on the position attribute of the knowledge node in the semantic association path. The path correction is achieved by calculating the weight coefficient of the path level where the node is located and the jump distance between the node and the core intention node.

[0209] Based on the position attribute of the knowledge node in the semantic association path p_t, the initial contribution score s_i is path-corrected. The weight coefficient of the path level where the node is located is calculated. The higher the level, the greater the weight coefficient of the node. At the same time, the jump distance between the node and the core intention node is calculated. The closer the jump distance, the greater the weight coefficient of the node. Suppose the weight coefficient of the path level where the node is located is wl, the jump distance between the node and the core intention node is d, and the weight coefficient of the jump distance is wd. The score s_f after path correction can be obtained by weighting the initial contribution score s_i, that is, s_f=s_i*(wl+wd / (d+1)). The addition of 1 here is to avoid the situation where the denominator is 0 when the jump distance is 0. The score s_f after path correction is used as the final contribution score of the knowledge node.

[0210] Step S154: For nodes whose contribution scores are higher than the second preset threshold, increase the connection weight between the node and the node corresponding to the core intent feature.

[0211] A second preset threshold t6 is set, and for nodes in the semantic association path whose contribution scores are higher than t6, the connection weight between the node and the node corresponding to the core intent feature is increased. In the innovation and entrepreneurship counseling knowledge base, each connection between nodes has a corresponding weight, which reflects the strength of the association between them. By increasing the connection weight, the association between these nodes and the core intent can be strengthened, making them easier to be selected in the subsequent matching process. For example, if a knowledge node has a high contribution score for "risk control of seed round financing" and the core intent feature corresponding node is "financing and market expansion strategy", then the connection weight between the two nodes is increased.

[0212] Step S155: For nodes whose contribution scores are lower than a third preset threshold, reduce the connection weight of the node or remove invalid connections.

[0213] A third preset threshold t7 is set. For nodes with a contribution score lower than t7, the connection weight of the node is reduced or invalid connections are removed according to the specific situation. If a node has a very low contribution in user interaction, it means that it is not closely related to the user's needs. By reducing the connection weight, its priority in the matching process can be lowered. If the connection between the node and other nodes is completely useless, that is, it is an invalid connection, it will be removed to optimize the structure of the knowledge base. For example, if a knowledge node about "niche marketing methods for early-stage startups" has a very low contribution score, its connection weight with other nodes can be reduced. If it is found that its connection with the core business-related node has never been used, the connection will be removed.

[0214] Step S156: Based on the newly added semantic understanding features extracted from the user's subsequent question sentences, a new node is created in the innovation and entrepreneurship guidance knowledge base and an association relationship is established with the existing node.

[0215] The user's subsequent question statements are processed and new semantic understanding features are extracted according to the method of steps S110-S115. If the content represented by the new semantic understanding feature does not have a corresponding node in the existing innovation and entrepreneurship guidance knowledge base, then a new node is created in the knowledge base. Based on the semantic association between the new semantic understanding feature and the existing node, an association relationship is established between the new node and the existing node. For example, if the user's subsequent question involves "the integrated application of blockchain technology in artificial intelligence entrepreneurship", the extracted new semantic understanding feature finds that there is no relevant node in the knowledge base, so a new node "integrated entrepreneurial application of artificial intelligence and blockchain" is created, and a connection relationship is established with existing nodes such as "artificial intelligence entrepreneurship" and "blockchain technology" based on the semantic association.

[0216] Furthermore, the training method of the semantic coding model includes the following steps:

[0217] Step S210: Collect sample question sentences and their corresponding annotation intention labels in historical innovation and entrepreneurship coaching scenarios to construct a training dataset.

[0218] Collect a large number of sample questions from historical innovation and entrepreneurship coaching scenarios. These sample questions can come from historical user interactions with the coaching system, entrepreneurship-related questions in online forums, and other sources. Each sample question is annotated with a corresponding intent label, such as "financing strategy consultation" or "market expansion method inquiry." Combine the sample questions and their annotated intent labels to construct a training dataset. During this construction process, ensure the data is diverse and representative, covering a wide range of innovation and entrepreneurship issues.

[0219] Step S220: Optimize the parameters of the semantic encoding model with the goal of minimizing the cross entropy loss between the predicted intent label and the annotated intent label of the sample question sentence.

[0220] The training dataset is input into the semantic encoding model, which processes the sample question sentences and outputs a predicted intent label. The cross-entropy loss is calculated between the predicted intent label and the annotated intent label. The cross-entropy loss reflects the difference between the predicted and true results. With the goal of minimizing the cross-entropy loss, the parameters of the semantic encoding model are updated using an optimization algorithm (such as stochastic gradient descent). In each round of training, the model adjusts the parameters based on the gradient of the loss function, gradually reducing the loss. After multiple rounds of training, the model performance stabilizes, resulting in a trained semantic encoding model.

[0221] Furthermore, the training method of the node matching model includes the following steps:

[0222] Step S310: extracting positive sample knowledge nodes from historical successfully matched question and answer data, and recording the user question semantic understanding features corresponding to the positive sample knowledge nodes.

[0223] Positive knowledge nodes are selected from historically successfully matched question-and-answer data. These nodes are highly compatible with the user's question. The semantic understanding features of the user's question corresponding to each positive knowledge node are recorded. These semantic understanding features are extracted using the method in steps S110-S115. The positive knowledge nodes and their corresponding semantic understanding features constitute the positive dataset.

[0224] Step S320: Process the positive sample knowledge nodes to generate negative sample knowledge nodes.

[0225] The positive sample knowledge nodes are processed to generate negative sample knowledge nodes. There are two specific methods. One is to retain the hierarchical attributes and associated labels of the positive sample knowledge nodes, and replace the core keywords of the positive sample knowledge nodes to generate semantically contradictory node description text. For example, if the positive sample knowledge node is "equity financing strategies for innovative entrepreneurship", "equity financing" can be replaced with "debt financing" to generate a semantically contradictory node description. Another method is to select nodes from the same tutoring field that are irrelevant to the semantics of the user's question as negative samples. The generated negative sample knowledge nodes and the corresponding user question semantic understanding features constitute a negative sample dataset.

[0226] Step S330: calling the initialized node matching model to respectively calculate the positive sample similarity between the node features of the positive sample knowledge node and the corresponding semantic understanding features, and the negative sample similarity between the node features of the negative sample knowledge node and the semantic understanding features.

[0227] The node features and corresponding semantic understanding features of the positive sample knowledge nodes are input into the initialization node matching model to calculate the positive sample similarity. Similarly, the node features and semantic understanding features of the negative sample knowledge nodes are input into the initialization node matching model to calculate the negative sample similarity. The initialization node matching model can be a simple neural network model that learns the mapping relationship between node features and semantic understanding features and outputs a similarity score.

[0228] Step S340: Construct a contrast loss function to maximize the difference between the positive sample similarity and the negative sample similarity, and update the parameters of the initialized node matching model through the back propagation algorithm until the initialized node matching model converges to obtain a trained node matching model.

[0229] A contrastive loss function is constructed. Its goal is to maximize the difference between the similarity of positive samples and the similarity of negative samples. Using the backpropagation algorithm, the parameters of the initialized node matching model are updated based on the gradient of the loss function. In each round of training, the model adjusts the parameters to maximize the similarity of positive samples and minimize the similarity of negative samples. This process is repeated until the model performance reaches stability, i.e., the loss function value no longer decreases significantly. At this point, the initialized node matching model converges, resulting in a trained node matching model.

[0230] During the data collection process of this embodiment, the principles of legality, legitimacy, and necessity are strictly adhered to, fully safeguarding the user's right to know and right to choose, and ensuring that all data collection is explicitly authorized by the user. For example, in the initial stage of user interaction with the system, a clear and easy-to-understand privacy policy and user agreement can be used to explain in detail the purpose, method, scope of data collection, and the use and protection of the data. The user is informed that the collected data includes question statements, interactive behavior data (such as content click locations, case review times, resource link open rates, and subsequent question statements), and that this data will be used to optimize the system's question-and-answer matching service and provide users with more accurate and personalized innovation and entrepreneurship guidance. At the same time, the user is clearly informed that they have the right to withdraw authorization at any time, and that withdrawing authorization will not affect their normal use of the basic functions of the system.

[0231] When a user registers or uses the system for the first time, a pop-up window or a checkbox will be displayed to request authorization for data collection. The system will only begin collecting relevant data after the user explicitly consents. For the collection of privacy-sensitive data (such as a small amount of user background information), specific authorization will be obtained from the user to ensure that the user is fully aware of the data's purpose and potential risks. Furthermore, a comprehensive authorization management mechanism will be established to record user authorization information and authorization time. If the system needs to adjust the scope, method, or purpose of data collection, the user will be notified promptly and authorization will be re-obtained. At the same time, the user's authorization status will be regularly checked to ensure its validity and legitimacy.

[0232] Furthermore, during the data collection process, various technical measures are employed to protect privacy and prevent data leakage when handling privacy-sensitive data. First, collected user data is encrypted using advanced encryption algorithms (such as AES) to convert the data into ciphertext for storage and transmission. For data storage, a secure database system is employed with strict access rights, ensuring that only authorized personnel can access the data. Furthermore, data is regularly backed up to prevent data loss. During data use, data desensitization technology is employed to remove sensitive information (such as user names and contact information), retaining only non-sensitive information relevant to the business. For example, user names are replaced with numbers and contact information is partially masked. During model training, technologies such as federated learning are employed to eliminate the need for centralized data storage and processing. Instead, model training is performed locally on the device, with only model parameters transmitted, thus preventing the leakage of private data. These technical measures ensure the security and confidentiality of user privacy-sensitive data.

[0233] Figure 2 A schematic diagram illustrates exemplary hardware and software components of a semantic understanding-based innovation and entrepreneurship coaching question-and-answer matching system 100, provided in some embodiments of the present application, that can implement the concepts of the present application. For example, a processor 120 can be used in the semantic understanding-based innovation and entrepreneurship coaching question-and-answer matching system 100 to perform the functions of the present application.

[0234] The semantic understanding-based innovation and entrepreneurship coaching question-and-answer matching system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the semantic understanding-based innovation and entrepreneurship coaching question-and-answer matching method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0235] For example, the innovation and entrepreneurship coaching question and answer matching system 100 based on semantic understanding may include a network port 110 connected to the network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the innovation and entrepreneurship coaching question and answer matching system 100 based on semantic understanding may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The innovation and entrepreneurship coaching question and answer matching system 100 based on semantic understanding also includes an I / O interface 150 between the computer and other input and output devices.

[0236] For ease of explanation, only one processor is described in the innovation and entrepreneurship coaching question and answer matching system 100 based on semantic understanding. However, it should be noted that the innovation and entrepreneurship coaching question and answer matching system 100 based on semantic understanding in this application may also include multiple processors, so the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the innovation and entrepreneurship coaching question and answer matching system 100 based on semantic understanding executes step A and step B, it should be understood that step A and step B may also be executed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0237] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned innovation and entrepreneurship coaching question-and-answer matching method based on semantic understanding is implemented.

[0238] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. The innovation and entrepreneurship coaching question-answer matching method based on semantic understanding is characterized by: The method comprises: Extract semantic understanding features of the question sentence input by the user; Matching a set of knowledge nodes corresponding to the semantic understanding features based on a preset innovation and entrepreneurship coaching knowledge base; Generating a semantic association path according to the hierarchical relationship and semantic association of the knowledge node set; Determine the target question-answering strategy based on the semantic association path and generate optimized tutoring content; Updating the node connection weights and association relationships in the innovation and entrepreneurship coaching knowledge base based on user feedback data on the optimized coaching content; Determining the target question-answering strategy based on the semantic association path and generating optimized tutoring content includes: Parsing the node sequence in the semantic association path, and identifying the tutoring topic evolution logic and key knowledge point distribution corresponding to the node sequence; Matching a basic question-and-answer strategy template from a preset question-and-answer strategy library according to the tutoring topic evolution logic, wherein the basic question-and-answer strategy template includes content generation rules, interaction process design, and resource reference method; Dynamically adjusting the basic question-answering strategy template based on the distribution of key knowledge points to obtain an adjusted basic question-answering strategy template; The optimized coaching content is generated based on the adjusted basic question-and-answer strategy template; the optimized coaching content includes step-by-step guidance text, reference cases adapted to user needs, and extended resource links for related knowledge points.

2. The innovation and entrepreneurship coaching question-answer matching method based on semantic understanding according to claim 1 is characterized in that: The step of extracting semantic understanding features of the question sentence input by the user includes: Performing word segmentation on the question sentence, dividing it into multiple semantic word units and removing stop words from the multiple semantic word units; Calling a pre-trained semantic encoding model to perform context encoding on the semantic word unit to generate a context vector for each semantic word unit; Input the context vector into the multi-head attention layer to calculate the semantic association weight between each semantic word unit; Performing weighted fusion on the context vectors based on the semantic association weights to generate a global semantic feature of the question sentence; Input the global semantic features into the feature dimensionality reduction layer to extract core intent features and auxiliary logic features as the semantic understanding features; wherein the core intent features are used to represent the core objectives of the user's question, and the auxiliary logic features are used to describe the dependency relationship and semantic hierarchy between multiple semantic word units in the question sentence; The semantic coding model is trained by the following steps: Collect sample question sentences and their corresponding annotated intent labels from historical innovation and entrepreneurship coaching scenarios to construct a training dataset; The parameters of the semantic encoding model are optimized with the goal of minimizing the cross entropy loss between the predicted intent label and the annotated intent label of the sample question sentence.

3. The innovation and entrepreneurship coaching question-answer matching method based on semantic understanding according to claim 2 is characterized in that: The calling of the pre-trained semantic encoding model to perform context encoding on the semantic word unit to generate a context vector for each semantic word unit includes: Inputting the semantic word unit into a bidirectional long short-term memory network to obtain a forward hidden state sequence and a backward hidden state sequence of each semantic word unit; Concatenate the forward hidden state sequence and the backward hidden state sequence of the same semantic word unit to generate the initial context vector; Performing layer normalization on the initial context vector, and fusing the normalized context vector with the initial context vector through a residual connection to obtain an enhanced context vector; The enhanced context vector is input into a nonlinear transformation layer to generate a context vector set with uniform dimension.

4. The innovation and entrepreneurship coaching question-answer matching method based on semantic understanding according to claim 2 is characterized in that: The preset innovation and entrepreneurship coaching knowledge base matches a set of knowledge nodes corresponding to the semantic understanding features, including: Traversing the candidate knowledge nodes in the innovation and entrepreneurship coaching knowledge base, calling a node matching model to calculate a first similarity score between a node feature of each candidate knowledge node and the core intent feature; Filtering candidate knowledge nodes whose first similarity scores are higher than a first preset threshold to form an initial matching set; Calling a logic verification model to analyze whether the node description texts of the candidate knowledge nodes in the initial matching set satisfy the dependency relationship and semantic hierarchy corresponding to the auxiliary logical features; if not, determining them as conflicting nodes and eliminating the conflicting nodes to obtain candidate knowledge nodes after logic verification; The candidate knowledge nodes after logical verification are arranged in descending order according to the first similarity score, and the top N nodes are selected as the optimized knowledge node set.

5. The innovation and entrepreneurship coaching question-answer matching method based on semantic understanding according to claim 4 is characterized in that: The training method of the node matching model includes: Extract positive sample knowledge nodes from historical successfully matched question and answer data, and record the user question semantic understanding features corresponding to the positive sample knowledge nodes; Processing the positive sample knowledge nodes to generate negative sample knowledge nodes, specifically including: retaining the hierarchical attributes and associated labels of the positive sample knowledge nodes, replacing the core keywords of the positive sample knowledge nodes to generate semantically contradictory node description texts, or selecting nodes from the same tutoring field that are semantically irrelevant to the user's question; Calling the initialized node matching model to respectively calculate the positive sample similarity between the node feature of the positive sample knowledge node and the corresponding semantic understanding feature, and the negative sample similarity between the node feature of the negative sample knowledge node and the semantic understanding feature; A contrast loss function is constructed to maximize the difference between the similarity of positive samples and the similarity of negative samples, and the parameters of the initialized node matching model are updated through a back propagation algorithm until the initialized node matching model converges to obtain a trained node matching model.

6. The innovation and entrepreneurship coaching question-answer matching method based on semantic understanding according to claim 1 is characterized in that: Generating a semantic association path according to the hierarchical relationship and semantic association of the knowledge node set includes: Obtaining hierarchical attributes, associated tags, and historical user interaction data of each node in the knowledge node set; Constructing a vertical hierarchical connection relationship between nodes according to the hierarchical attributes; Expanding the horizontal semantic association relationship between nodes according to the association label; Generate an initial semantic network based on the vertical hierarchical connection relationship and the horizontal semantic association relationship; The path generation algorithm is called to traverse the initial semantic network to generate multiple candidate semantic paths. Each candidate semantic path is scored based on the path length, the jump probability between nodes, the node semantic weight and the user feedback index, combined with the historical user interaction data, and the candidate semantic path with the highest score is selected as the target semantic association path.

7. The innovation and entrepreneurship coaching question-answer matching method based on semantic understanding according to claim 6 is characterized in that: The call path generation algorithm traverses the initial semantic network to generate multiple candidate semantic paths, and scores each candidate semantic path based on path length, inter-node jump probability, node semantic weight, and user feedback indicators, combined with the historical user interaction data, and selects the candidate semantic path with the highest score as the target semantic association path, including: Starting from the core intent node in the knowledge node set, traverse the child nodes downward according to the vertical hierarchical connection relationship, and expand the sibling nodes according to the horizontal semantic association relationship; During the traversal process, the visited node sequence and jump relationship are recorded to generate the initial path set; Performing redundant path filtering on the initial path set to obtain multiple candidate semantic paths, specifically including: merging paths containing the same node sequence but different jump orders, and eliminating paths containing loop nodes or repeated jumps; Obtaining a path length of each candidate semantic path, where the path length is obtained by counting the total number of all nodes in the candidate semantic path; Calculate the inter-node jump probability of the candidate semantic path, which is determined based on the ratio of the number of jumps between adjacent nodes in the historical user interaction data to the total number of jumps to the starting node. Extracting the node semantic weights of the candidate semantic paths, wherein the node semantic weights read the preset weight values of each node from the innovation and entrepreneurship counseling knowledge base, and the weight values are calculated and generated based on the node citation frequency and the semantic relevance annotated by the user; Obtaining a user feedback index, wherein the user feedback index is obtained by counting the average satisfaction score generated by all nodes in the candidate semantic path after being clicked by users in historical sessions and the matching degree of associated subsequent questions; The path length, inter-node jump probability, node semantic weight and user feedback index of each candidate semantic path are normalized and then weighted summed to generate a comprehensive score for each candidate semantic path; The candidate semantic path with the highest comprehensive score is selected as the target semantic association path.

8. The innovation and entrepreneurship coaching question-answer matching method based on semantic understanding according to claim 1 is characterized in that: The dynamically adjusting the basic question-answering strategy template based on the key knowledge point distribution to obtain the adjusted basic question-answering strategy template includes: Identifying a priority parameter for each knowledge node in the key knowledge point distribution, wherein the priority parameter is calculated based on the semantic weight, historical user interaction frequency, and hierarchical attributes of the knowledge node; Dividing the key knowledge points into a core knowledge point set and an auxiliary knowledge point set according to the priority parameter, wherein the priority parameter of the core knowledge point set is higher than a preset priority threshold; Extracting content generation rules from the basic question-answering strategy template, setting the content generation rules corresponding to the core knowledge point set as a default activation module, and generating an expansion trigger condition based on the semantic relevance of the auxiliary knowledge point set; Analyze the unvisited knowledge point sequence in the user's historical interaction data, calculate the matching degree between the unvisited knowledge point sequence and the auxiliary knowledge point set, and select the unvisited knowledge points with a matching degree higher than a preset matching threshold as insertion nodes for supplementary cases; Determine, based on the auxiliary logic feature in the semantic understanding feature, a logical embedding position of the insertion node of the supplementary case in the basic question-answering strategy template, wherein the logical embedding position is obtained by matching the dependency relationship in the auxiliary logic feature with the interaction process node sequence in the basic question-answering strategy template; The default activation module, extended trigger conditions and logic embedding position are parameterized to generate an adjusted basic question-and-answer strategy template.

9. An innovation and entrepreneurship coaching question-answer matching system based on semantic understanding, characterized by: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the innovation and entrepreneurship coaching question and answer matching method based on semantic understanding as described in any one of claims 1 to 8.

Images