Intelligent question-answering method and system for entrepreneurial mentors based on dialogue system

Abstract

The present invention provides an intelligent question-answering method and system for entrepreneurial mentors based on a dialogue system, which relates to the technical field of natural language processing. First, a target conversation request containing entrepreneurial guidance-related question content input by a user is obtained, and semantic analysis is performed to generate intent categories and context-related features. Then, based on the intent categories, associated nodes in a preset entrepreneurial knowledge graph are matched, and the adjacent paths of the associated nodes are traversed according to the context-related features to extract knowledge combination relationships. Finally, target reply content is generated based on the knowledge combination relationship and returned to the user end to complete the question-answering interaction. In this way, the user's question intention can be deeply understood, the knowledge structure in the field of entrepreneurial guidance can be fully utilized, and accurate, relevant and in-depth reply content can be generated, which significantly improves the performance of the intelligent question-answering system in the field of entrepreneurial guidance.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to an intelligent question-answering method and system for entrepreneurial mentors based on a dialogue system. Background Art

[0002] In the field of entrepreneurial guidance, entrepreneurs often face various challenges and problems, requiring professional guidance and advice. Traditional question-and-answer methods for entrepreneurial mentors often rely on manual consultation, which is limited by the mentor's time and energy and cannot meet the immediate needs of a large number of entrepreneurs. With the development of artificial intelligence technology, intelligent question-and-answer methods based on dialogue systems have gradually become a research hotspot in the field of entrepreneurial guidance. However, when dealing with complex and multi-layered entrepreneurial guidance questions, existing intelligent question-and-answer methods often suffer from inaccurate understanding, irrelevant responses, or a lack of depth. This is primarily because these methods fail to fully utilize the knowledge structure of the entrepreneurial guidance field and the contextual information in the user's questions, resulting in the generated responses failing to meet the actual needs of entrepreneurs. 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 an intelligent question-answering method for entrepreneurial mentors based on a dialogue system, the method comprising:

[0004] Obtaining a target conversation request input by a user, wherein the target conversation request includes user question content related to entrepreneurship guidance;

[0005] Performing semantic parsing on the target conversation request to generate a semantic parsing result of the user's question content, wherein the semantic parsing result includes an intent category and context-related features;

[0006] Matching associated nodes in a preset entrepreneurship knowledge graph based on the intention category, wherein the associated nodes include entrepreneurship guidance knowledge entities corresponding to the intention category;

[0007] Traversing the adjacent path of the associated node according to the context association feature, and extracting the knowledge combination relationship between the associated node and the adjacent node;

[0008] Target reply content is generated based on the knowledge combination relationship, and the target reply content is returned to the user terminal to complete the question-answer interaction.

[0009] On the other hand, an embodiment of the present invention also provides an intelligent question-and-answer system for entrepreneurial mentors based on a dialogue system, including a processor and a machine-readable storage medium, wherein 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.

[0010] Based on the above aspects, the embodiment of the present invention can accurately identify the intent category and context-related features of the user's question by obtaining the target conversation request input by the user and performing semantic parsing. Furthermore, based on the intent category, the associated nodes in the preset entrepreneurial knowledge graph are matched, and the adjacent paths of the associated nodes are traversed according to the context-related features to extract the knowledge combination relationship between the associated nodes and the adjacent nodes. This fully utilizes the knowledge structure in the field of entrepreneurial guidance and the contextual information in the user's question, so that the generated target reply content can closely revolve around the core needs of the user's question, providing accurate, relevant and in-depth entrepreneurial guidance suggestions, significantly improving the performance and practicality of the intelligent question-answering system in the field of entrepreneurial guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the intelligent question-answering method for entrepreneurial mentors based on the dialogue system provided by an embodiment of the present invention.

[0012] Figure 2 Schematic diagram of exemplary hardware and software components of an intelligent question-and-answer system for entrepreneurial mentors based on a dialogue system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an intelligent question-answering method for entrepreneurial mentors based on a dialogue system provided by an embodiment of the present invention. The following is a detailed introduction to the intelligent question-answering method for entrepreneurial mentors based on a dialogue system.

[0014] Step S110: obtaining a target conversation request input by a user, wherein the target conversation request includes user question content related to entrepreneurship guidance.

[0015] For example, in the field of internet product startups, user-entered target conversation requests often contain in-depth explorations of practical entrepreneurial issues. For example, a user might enter a complex and detailed message like this: "Amidst the fierce competition in the current internet social product market, with emerging social models like short videos and live streaming rapidly emerging and seizing market share, and increasingly stringent data privacy regulations, a startup team of recent college graduates, comprised of computer science students, plans to develop an internet product that integrates virtual reality (VR) and augmented reality (AR) technologies, aiming to create an immersive social experience for young people. However, the team faces challenges: insufficient technical R&D experience, extreme funding constraints, relying on limited angel investment, and a lack of effective marketing channels and industry connections. Therefore, how can they comprehensively consider the product's technical architecture selection, user interface and interaction design, content operation strategy, profit model innovation, and strategies to counter potential competitors, in order to develop a business plan that not only ensures technological leadership and superior user experience, but also quickly establishes a foothold in the market and achieves profitable growth?"

[0016] This target conversation request closely fits the actual scenario of Internet product entrepreneurship, covering multiple key dimensions such as market environment, team status, product characteristics, technical challenges, operational strategies, and profit goals. It is a very representative user question related to Internet product entrepreneurship guidance.

[0017] When obtaining user input for a target session request, we strictly adhere to relevant laws and regulations and user authorization, clearly inform users of the purpose and scope of their data use, and obtain their consent. We also employ anonymization and encryption technologies to process user data to ensure its privacy and security.

[0018] Step S120: performing semantic parsing on the target conversation request to generate a semantic parsing result of the user's question content, wherein the semantic parsing result includes an intent category and context-related features.

[0019] To accurately understand the user's intent regarding internet product startups and construct relevant contextual features, comprehensive and detailed semantic parsing of the target conversation request is necessary. The following steps will explain this in detail.

[0020] Step S121: performing word segmentation processing on the user's question content to obtain multiple semantic unit sequences.

[0021] When segmenting complex user questions like these, we must fully consider the internet industry's specialized terminology, unique linguistic expressions, and the inherent semantic connections between words. First, we should make a preliminary segmentation based on common lexical boundaries. At the same time, we should retain specialized terms and fixed expressions with specific meanings as complete semantic units.

[0022] For example, "virtual reality (VR) and augmented reality (AR) technology", "immersive social experience", "data privacy protection regulations", "technology R&D experience", "angel investment", "marketing channels", "industry network resources", "technology architecture selection", "user interface and interaction design", "content operation strategy", "profit model innovation", "sniping strategy of potential competitors", etc. These are all professional expressions with specific meanings in the field of Internet product entrepreneurship and should be treated as independent semantic units.

[0023] After word segmentation, the resulting semantic unit sequence may be as follows: "at", "at present", "Internet social products", "competition", "extremely fierce", ","", "short video", ",","live broadcast", "etc.", "emerging social models", "continuously emerging", "and", "rapidly", "seize", "market share", ",","at the same time", "data privacy protection regulations", "increasingly strict", "the", "environment", "under", ","for", "a", "composed of", "several", "just graduated from", "colleges", "computer science students", "of", "startup team", ","they", "intend", "develop", "a", "integration", "virtual reality (VR) and augmented reality (AR) technology", ","aims", "to", "create", "immersive social experience", "Internet product", ".", "However", ","the team", "faces", "technology R&D experience" ", "insufficient", ","funds", "extremely scarce", "only", "relying on", "a small amount of", "angel investment", "maintaining", "and", "lack of", "effective", "market promotion channels", "and", "industry network resources", "dilemma". "So", ","how", "comprehensively consider", "product", "technical architecture selection", "user interface and interaction design", "content operation strategy", "profit model innovation", "and", "responding to", "potential competitors", "sniper strategy", "thereby", "developing", "a set of", "that can", "ensure", "product", "leading", "in", "technology", "user experience", "excellence", "and", "quickly", "gain" a foothold" in the", "market" and "achieve", "profit growth", "entrepreneurship plan".

[0024] Step S122: calling a pre-trained language parsing model to perform intent recognition processing on the semantic unit sequence to determine the initial intent category of the user's question content.

[0025] When processing semantic information related to internet products, the pre-trained language parsing model accurately identifies the intent of user questions through complex multi-layered operations. This model primarily consists of an encoding layer, an attention layer, and a classification layer. The following details the specific processing steps of each layer.

[0026] Step S1221: input the semantic unit sequence into the encoding layer of the language parsing model for vector conversion processing to generate a context vector representation of the semantic unit sequence.

[0027] The encoding layer's primary task is to convert each semantic unit into a corresponding vector representation, while fully considering the contextual relationships between semantic units to generate a comprehensive contextual vector representation. Specifically, the encoding layer assigns each semantic unit an initial word vector. These word vectors can be obtained from pre-trained word embedding models such as Word2Vec and GloVe.

[0028] Assume that the sequence of semantic units is S = [s1, s2, ..., sn], where si represents the i-th semantic unit. The encoding layer uses a sequence model, such as a long short-term memory (LSTM), to process these semantic units. LSTM has a unique gating mechanism that effectively handles long-range dependencies in sequence data.

[0029] In LSTM, the forget gate (ft) determines how much information in the cell state Ct-1 from the previous time step should be forgotten. It processes the current input xt and the hidden state ht-1 from the previous time step using a sigmoid function, generating a vector between 0 and 1, where 0 indicates complete forgetfulness and 1 indicates complete retention. The calculation process can be described as follows: first, the current input xt and the hidden state ht-1 from the previous time step are concatenated. Then, a linear transformation is performed using a weight matrix Wf. The bias vector bf is added, and finally, the forget gate output ft is obtained using a sigmoid function σ: ft = σ(Wf[ht-1, xt] + bf).

[0030] The input gate (it) determines how much information in the current input xt needs to be added to the cell state Ct. Similarly, it first concatenates the current input xt and the hidden state ht-1 of the previous time step, performs a linear transformation through the weight matrix Wi, adds the bias vector bi, and then passes through the sigmoid function σ to obtain the output it of the input gate, that is, it=σ(Wi[ht-1,xt]+bi). At the same time, a candidate cell state Ct~ is also calculated. It is obtained by concatenating the current input xt and the hidden state ht-1 of the previous time step, linearly transforming it through the weight matrix Wc, adding the bias vector bc, and then passing through the hyperbolic tangent function tanh, that is, Ct~=tanh(Wc[ht-1,xt]+bc).

[0031] The cell state Ct is updated according to the output of the forget gate and the input gate, which is the cell state Ct-1 of the previous time step multiplied by the output ft of the forget gate plus the output it of the input gate multiplied by the candidate cell state Ct~, that is, Ct=ft*Ct-1+it*Ct~.

[0032] The output gate (ot) determines how much information in the current cell state Ct is output to the hidden state ht. Its calculation is similar to that of the forget gate and input gate. First, the current input xt and the hidden state ht-1 of the previous time step are concatenated. A linear transformation is performed using the weight matrix Wo. The bias vector bo is added, and the output is passed through the sigmoid function σ to obtain the output of the output gate, ot: ot = σ(Wo[ht-1, xt] + bo). Finally, the hidden state ht is the output of the output gate ot multiplied by the cell state Ct processed by the hyperbolic tangent function tanh: ht = ot * tanh(Ct).

[0033] After processing at the encoding layer, each semantic unit corresponds to a hidden state vector. Combining these hidden state vectors yields a context vector representation of the semantic unit sequence. For example, for the first semantic unit s1, after LSTM processing, the hidden state vector h1 is obtained. The second semantic unit s2 is obtained by h2, and so on, ultimately resulting in a context vector representation H = [h1, h2, ..., hn].

[0034] Step S1222: Input the context vector representation into the attention layer of the language parsing model for feature weight allocation processing to generate a weighted context vector representation.

[0035] The core function of the attention layer is to make the model pay more attention to the semantic information related to intent recognition, thereby improving the accuracy of intent recognition. Assume that the context vector is represented as H = [h1, h2, ..., hn], where hi represents the hidden state vector corresponding to the i-th semantic unit.

[0036] The attention layer calculates the attention score for each hidden state vector. Specifically, each hidden state vector hi is first linearly transformed using a weight matrix W, then a bias vector b is added. Then, it is nonlinearly transformed using the hyperbolic tangent function tanh. Finally, a dot product is performed with a learnable vector v to obtain the attention score ei, i.e., ei = vT*tanh(W*hi+b).

[0037] After obtaining the attention scores, they are normalized using the softmax function to convert them into a probability distribution. The softmax function converts each attention score ei into a probability value αi between 0 and 1, where the sum of all probabilities is 1. The calculation method is αi = exp(ei) / ∑j = 1nexp(ej).

[0038] Next, the attention weights αi are weighted concatenated with the context vector representation H. For each hidden state vector hi, it is multiplied by the corresponding attention weight αi to obtain the weighted hidden state vector αi*hi. These weighted hidden state vectors are then concatenated in a predefined manner to generate the weighted context vector representation. For example, they can be concatenated along the vector dimensions to form a new high-dimensional vector to preserve more semantic information.

[0039] Step S1223: Input the weighted context vector representation into the classification layer of the language parsing model for intention probability prediction processing to generate a probability distribution of the user question content under each preset intention category.

[0040] The classification layer is typically a fully connected layer. Its primary task is to map the weighted context vector representation to various predefined intent categories and calculate the probability of each intent category. Assuming there are m predefined intent categories, the classification layer performs a linear transformation on the weighted context vector representation using a weight matrix Wc and a bias vector bc.

[0041] Specifically, the weighted context vector representation, denoted as h_att, is multiplied by the weight matrix Wc and added to the bias vector bc to obtain an intermediate result. This intermediate result is then converted into a probability distribution using the softmax function. The softmax function converts each element in the intermediate result into a probability value between 0 and 1, and the sum of all probability values ​​is 1. This yields the probability distribution y = [p1, p2, ..., pm] of the user's question content under each preset intent category, where pi represents the probability that the user's question content belongs to the i-th intent category.

[0042] Step S1224: Determine, based on the probability distribution, the intent category corresponding to the maximum probability as the initial intent category.

[0043] By comparing the probability values ​​of each intent category in the probability distribution, the intent category with the highest probability is selected as the initial intent category. For example, if the probability distribution is y = [p1, p2, ..., pm], by traversing the probability distribution and finding the element with the highest probability value, assuming it is pj, then the corresponding intent category Cj is the initial intent category.

[0044] Step S1225: obtaining an intermediate feature vector generated by the language parsing model in the process of generating the probability distribution, and using the intermediate feature vector as a latent semantic feature of the user's question content.

[0045] During the computational process of the language parsing model, intermediate feature vectors are generated. These vectors contain the latent semantic information of the user's question. For example, the output of the hidden layer before the classification layer can serve as an intermediate feature vector. Specifically, before the classification layer performs a linear transformation on the weighted context vector representation, the hidden layer performs a series of calculations and transformations on the input to produce an intermediate feature representation. This intermediate feature representation is extracted as the latent semantic features of the user's question.

[0046] Step S1226: Constructing a semantic feature index of the user's question content based on the association relationship between the potential semantic features and the initial intention category.

[0047] Semantic feature indexing is used to quickly locate and retrieve information related to user questions. This can be accomplished by mapping latent semantic features to initial intent categories. For example, a latent semantic feature vector can be hashed to generate a unique index value, which is then associated with the initial intent category. This allows for rapid retrieval of information related to the user's question using the index value in subsequent processing.

[0048] Step S123: traverse the historical session records of the target session request, and extract a set of context keywords associated with the user's question content in the historical session records.

[0049] To more accurately understand the user's question intent, we need to consider the historical session records of the target session request. These historical session records may contain contextual information related to the current user's question. The following details the steps for extracting the contextual keyword set.

[0050] Step S1231: Obtain historical conversation records within a preset time period before the current conversation round from the user terminal, wherein the historical conversation records include a plurality of historical question contents and their corresponding historical reply contents.

[0051] The preset time period can be set based on actual circumstances, such as the last week or the last month. Historical conversation records within this time period are obtained from the user. These records include the user's previous questions and the system's responses. For example, in the scenario of an internet product startup, historical questions may involve product feature planning, marketing strategies, and other aspects, while historical responses contain suggestions and solutions to these questions.

[0052] Step S1232: performing keyword extraction processing on the historical question content to obtain a historical question keyword set.

[0053] Keyword extraction can be done using statistical methods (such as TF-IDF) or machine learning methods (such as TextRank). Keyword extraction algorithms are used to extract important keywords from historical questions. For example, for the historical question "How to improve user retention rates for internet products," extracted keywords might include "internet products" and "user retention rates." When processing multiple historical questions, keyword extraction is performed on each question separately. The extracted keywords are then aggregated, and duplicate keywords are removed to create a historical question keyword set.

[0054] Step S1233: Perform keyword extraction processing on the historical reply content to obtain a historical reply keyword set.

[0055] Similarly, keyword extraction is performed on historical replies to obtain a set of historical reply keywords. For example, if a historical reply mentions "improving user retention can be achieved by optimizing the product's user experience and providing personalized services," the extracted keywords may include "user experience" and "personalized services." When processing multiple historical replies, keyword extraction is performed on each reply separately. The extracted keywords are then aggregated, and duplicate keywords are removed to obtain a set of historical reply keywords.

[0056] Step S1234: performing similarity matching processing on the semantic unit sequence of the user question content and the historical question keyword set to determine a first similarity score between the user question content and each historical question content.

[0057] Similarity matching can be performed using methods such as cosine similarity and edit distance. The sequence of semantic units in the user's question content is converted into a vector representation, for example, using the context vector representation generated by the encoding layer. Each keyword in the historical question keyword set is then converted into a vector representation. These vectors can be obtained from a pre-trained word embedding model.

[0058] For the vector representation S_vec of the semantic unit sequence of the user's question content and each keyword vector ki_vec in the historical question keyword set, calculate the cosine similarity between them. Cosine similarity is obtained by calculating the dot product of two vectors divided by the product of their moduli. For each keyword vector ki_vec, a cosine similarity value sim1i is calculated, and these similarity values ​​are combined into a vector sim1=[sim11, sim12, ..., sim1m], where m is the number of keywords in the historical question keyword set. These similarity values ​​reflect the degree of similarity between the user's question content and each historical question keyword, that is, the first similarity score between the user's question content and each historical question content.

[0059] Step S1235: performing similarity matching processing on the semantic unit sequence of the user question content and the historical reply keyword set to determine a second similarity score between the user question content and each historical reply content.

[0060] Similarly, the vector representation S_vec of the semantic unit sequence of the user's question content is matched with each keyword vector ri_vec in the historical reply keyword set for similarity. Using the same method as step S1234, the cosine similarity between them is calculated to obtain a second similarity score vector sim2=[sim21, sim22, ..., sim2n], where n is the number of keywords in the historical reply keyword set. These similarity values ​​reflect the degree of similarity between the user's question content and each historical reply keyword, that is, the second similarity score between the user's question content and each historical reply content.

[0061] Step S1236: Filter out a target historical keyword set associated with the user question content according to a weighted sum of the first similarity score and the second similarity score.

[0062] To comprehensively consider the relevance of historical questions and responses, a weighted sum is performed on the first and second similarity scores. Assume that the weight of the first similarity score is w1, the weight of the second similarity score is w2, and w1 + w2 = 1. Multiply each element in the first similarity score vector sim1 by weight w1, and each element in the second similarity score vector sim2 by weight w2. Then, add the elements in corresponding positions together to obtain the weighted sum result vector sim = w1 * sim1 + w2 * sim2.

[0063] Based on the similarity values ​​in the weighted summation vector sim, keywords with high similarity are selected. A similarity threshold can be set to select keywords with similarity values ​​greater than the threshold to form the target historical keyword set. These keywords are highly relevant to the user's question and provide important context for subsequent intent calibration and response generation.

[0064] Step S1237: De-duplicate and sort the target historical keyword set to generate the context keyword set.

[0065] When deduplicating the target historical keyword set, each keyword in the set needs to be compared one by one. For example, in the case of an internet product startup, if there are duplicates of the keyword "internet marketing" in the target historical keyword set, the redundant entries will be removed to ensure the uniqueness of each keyword in the set.

[0066] The sorting process can be performed based on the weighted sum of similarity values ​​of the keywords. The higher the similarity value of the keyword, the higher it is in the sorting. For example, the keyword "VR social function design" has a higher similarity value after weighted summation, so it will be ranked higher in the sorting. Through such sorting, when the context keyword set is used later, priority will be given to keywords that are more relevant to the content of the user's question. After completing the deduplication and sorting process, a context keyword set is generated, which can accurately reflect the context information closely related to the content of the current user's question.

[0067] Step S124: inputting the context keyword set into the language parsing model, calibrating the initial intent category, and generating a calibrated intent category.

[0068] To further improve the accuracy of intent recognition, it is necessary to calibrate the previously determined initial intent category using the contextual keyword set. The specific steps of the calibration process are described in detail below.

[0069] Step S1241: input the context keyword set into the encoding layer of the language parsing model for vector conversion processing to generate a context vector set of the context keyword set, and perform average pooling processing on the context vector set to generate an aggregated context vector.

[0070] Each keyword in the context keyword set is sequentially input into the encoding layer of the language parsing model. Just as with the semantic unit sequence, the encoding layer assigns each keyword an initial word vector, derived from a pretrained word embedding model. Next, a sequence model, such as an LSTM, is used to process the keywords, generating a corresponding context vector for each keyword. Assuming the context keyword set is K = [k1, k2, ..., kn], after processing by the encoding layer, the corresponding context vector set V = [v1, v2, ..., vn] is obtained.

[0071] Average pooling is the process of combining all vectors in the context vector set V. Specifically, the values ​​of the corresponding dimensions of all vectors in the context vector set V are added together and then divided by the number of keywords, n. For example, the values ​​of the first dimension of vectors v1, v2, and so on are added together and divided by n to obtain the value of the aggregate context vector along the first dimension. This process is repeated several times to generate the aggregate context vector. This aggregate context vector represents the overall semantic information of the context keyword set.

[0072] Step S1242: Acquire the potential semantic features of the user question content according to the semantic feature index of the user question content.

[0073] The semantic feature index constructed earlier allows for the rapid and accurate acquisition of latent semantic features of user questions. The semantic feature index maps latent semantic features to initial intent categories. For example, using the previously constructed hash index, the corresponding index value directly identifies the associated latent semantic feature vector. This latent semantic feature vector contains key semantic information about the user's question as processed by the language parsing model.

[0074] Step S1243: Fusing the aggregated context vector with the latent semantic feature to generate a fused semantic feature vector.

[0075] The fusion process requires consideration of the rationality of weighted concatenation. Both the aggregated context vector and the latent semantic feature vector are multi-dimensional vectors with different semantic emphases. In the context of internet product startups, the aggregated context vector reflects the contextual semantics related to the current question in historical conversations, while the latent semantic feature vector reflects the latent semantics of the current question itself.

[0076] To perform weighted concatenation, we first need to determine the weight of each vector. Weights can be assigned based on the relevance of the contextual keyword set to the user's question and the importance of the latent semantic features. Assume that the weight of the aggregated context vector is w1, the weight of the latent semantic feature vector is w2, and w1 + w2 = 1. Multiply each dimension of the aggregated context vector by w1, and each dimension of the latent semantic feature vector by w2. Then, concatenate the two processed vectors along the dimension to form a new high-dimensional vector, the fused semantic feature vector. This fusion method fully preserves the semantic information of both vectors, providing a more comprehensive basis for subsequent intent calibration.

[0077] Step S1244: inputting the fused semantic feature vector into the classification layer of the language parsing model to perform intent probability prediction processing to generate a calibrated intent probability distribution.

[0078] The classification layer of the language parsing model processes the input fused semantic feature vector. The classification layer contains a weight matrix and a bias vector. The fused semantic feature vector is multiplied by the weight matrix and then added with the bias vector to obtain an intermediate result. This intermediate result is then converted into a probability distribution using the softmax function. The softmax function converts each element in the intermediate result into a probability value between 0 and 1, and the sum of all probability values ​​is 1. This results in a calibrated intent probability distribution, which reflects the probability that the user's question content belongs to each preset intent category after considering the contextual information.

[0079] Step S1245: Determine the intent category corresponding to the maximum probability as the candidate intent category based on the calibrated intent probability distribution, and record the probability difference between the probability value of the candidate intent category and the probability value of the initial intent category.

[0080] After obtaining the calibrated intent probability distribution, we need to find the intent category with the highest probability. By traversing each probability value in the calibrated intent probability distribution, we determine the intent category with the highest probability and use it as the candidate intent category. We also record the difference between the probability value of the candidate intent category and the probability value of the initial intent category previously determined. This probability difference reflects the degree to which the calibration process affected the intent recognition results.

[0081] Step S1246: Calculate the semantic distance score between the candidate intent category and the initial intent category.

[0082] The semantic distance score measures the degree of semantic difference between the candidate intent category and the initial intent category. Methods such as cosine similarity and Euclidean distance can be used to calculate the semantic distance score. Taking cosine similarity as an example, the candidate intent category and the initial intent category are first represented as vectors. These vectors can be the feature vectors corresponding to each intent category during the language parsing model training process. Then, the cosine similarity of these two vectors is calculated. The closer the cosine similarity value is to 1, the greater the semantic similarity between the two intent categories; the closer the value is to 0, the greater the semantic difference. Subtracting the cosine similarity value from 1 will yield the semantic distance score.

[0083] Step S1247: If the semantic distance score exceeds the preset distance threshold and the probability difference exceeds the preset probability threshold, the candidate intent category is used as the calibrated intent category; otherwise, the initial intent category is retained as the calibrated intent category.

[0084] The preset distance threshold and the preset probability threshold are pre-set standards used to determine whether the candidate intent category needs to be used as the final calibration result. If the semantic distance score exceeds the preset distance threshold, it means that the candidate intent category is semantically different from the initial intent category; at the same time, the probability difference exceeds the preset probability threshold, which means that the probability distribution after calibration has changed significantly. In this case, the candidate intent category is considered to more accurately reflect the user's question intention and is used as the calibrated intent category. Conversely, if the semantic distance score does not exceed the preset distance threshold or the probability difference does not exceed the preset probability threshold, it means that the calibration process has little impact on the intent recognition result, and the initial intent category is retained as the calibrated intent category.

[0085] Step S125: constructing context-related features of the user's question content based on the mapping relationship between the calibrated intention category and the context keyword set.

[0086] The contextual association feature can reflect the semantic association strength between the calibrated intent category and the context keyword set. The construction process is described in detail below.

[0087] Step S1251: Obtain a preset intention feature vector corresponding to the calibrated intention category, and extract a pre-trained semantic vector for each keyword in the context keyword set.

[0088] The preset intent feature vectors are determined during the model training phase. Each intent category has a corresponding feature vector. These feature vectors represent the location and characteristics of the intent category in the semantic space. Using the calibrated intent category, the corresponding preset intent feature vector can be directly obtained.

[0089] For each keyword in the context keyword set, its corresponding semantic vector is extracted from a pre-trained word embedding model. These pre-trained word embedding models are trained on large amounts of text data and are able to learn the semantic information of vocabulary. For example, for the keyword "Internet product promotion strategy," its corresponding high-dimensional semantic vector is extracted from the word embedding model.

[0090] Step S1252: Determine the association strength between each keyword and the calibrated intent category based on the positional relationship between the intent feature vector and the semantic vector of each keyword in the same vector space.

[0091] In the same vector space, the intent feature vector and the semantic vector of each keyword have their own position. The strength of association can be determined by calculating the similarity between them. A commonly used similarity calculation method is cosine similarity. Cosine similarity is calculated between the intent feature vector and the semantic vector of each keyword. The closer the cosine similarity value is to 1, the stronger the association strength between the keyword and the calibrated intent category; the closer the value is to 0, the weaker the association strength. In this way, an association strength value is determined for each keyword in the context keyword set.

[0092] Step S1253: dynamically assigning weights to the keywords in the context keyword set according to the association strength, and generating a weighted keyword semantic vector set.

[0093] Based on the association strength determined in the previous step, a weight is assigned to each keyword in the context keyword set. Keywords with greater association strength are assigned a higher weight. Multiply the semantic vector of each keyword by the corresponding weight to obtain a weighted keyword semantic vector. Combining all weighted keyword semantic vectors generates a set of weighted keyword semantic vectors. For example, for the keyword "VR technology application" with a high association strength, its weight is larger, and the weighted semantic vector obtained after multiplying it by the semantic vector will further highlight the importance of this keyword in subsequent processing.

[0094] Step S1254: performing multi-dimensional vector concatenation on the weighted keyword semantic vector set and the intention feature vector to generate a fused semantic feature vector.

[0095] Multidimensional vector concatenation involves dimensional expansion and combination of the weighted keyword semantic vector set and the intent feature vector. Since the weighted keyword semantic vector set contains semantic information for multiple keywords, while the intent feature vector represents a calibrated intent category, concatenating them can fuse contextual and intent information. Specifically, the intent feature vector and the weighted keyword semantic vector set are arranged in a predefined order across the vector dimensions to form a new high-dimensional vector, the fused semantic feature vector.

[0096] Step S1255: performing a linear dimensional transformation on the fused semantic feature vector using a preset mapping relationship matrix to generate a standardized multidimensional feature vector corresponding to the context-related feature.

[0097] The preset mapping relationship matrix is ​​learned during model training and performs a linear dimensionality transformation on the fused semantic feature vector. Multiplying the fused semantic feature vector by the mapping relationship matrix yields a new vector, which may have different dimensions than the fused semantic feature vector. This dimensionality transformation converts the fused semantic feature vector into a standardized multidimensional feature vector more suitable for representing contextual features. This standardization ensures that the contextual features generated from different user questions are consistent in dimensionality and numerical range, facilitating subsequent processing and comparison.

[0098] Step S1256: constructing the contextual association feature reflecting the semantic association strength between the calibrated intent category and the contextual keyword set according to the numerical distribution of each dimension in the standardized multidimensional feature vector.

[0099] The numerical value of each dimension in the standardized multidimensional feature vector has a specific meaning, and the distribution of these values ​​reflects the strength of the semantic association between the calibrated intent category and the contextual keyword set. The standardized multidimensional feature vector can be further processed and analyzed based on the magnitude and distribution of the numerical values. For example, dimensions with larger numerical values ​​can be considered to have stronger semantic associations with the calibrated intent category and the contextual keyword set. By comprehensively considering the numerical values ​​of each dimension, a contextual association feature is constructed that accurately reflects the strength of the semantic association. This contextual association feature provides an important basis for subsequently searching for relevant knowledge in the entrepreneurial knowledge graph.

[0100] Step S126: combining the calibrated intent category with the context-related features to generate the semantic parsing result.

[0101] The calibrated intent categories and contextual features are combined to form the final semantic parsing result. The calibrated intent categories clarify the core intent of the user's question, while the contextual features provide contextual information related to that intent. This combination comprehensively and accurately reflects the semantic information of the user's question content. In the context of internet product startups, the semantic parsing results can help the system better understand user needs and provide accurate input for subsequently matching relevant nodes in the pre-set startup knowledge graph.

[0102] In the above embodiments, a pre-trained language parsing model serves as an AI model. The pre-trained language parsing model is primarily composed of an encoding layer, an attention layer, and a classification layer. The encoding layer is at the front end of the model, and its function is to convert the input sequence of semantic units into a context vector representation. It receives a sequence of semantic units as input, maps each semantic unit into a word vector using a pre-trained word embedding model, and then processes these word vectors using a sequence model such as a long short-term memory network (LSTM), taking into account the contextual relationships between semantic units and outputting a context vector representation. The encoding layer is directly connected to the attention layer, passing the context vector representation to the attention layer.

[0103] The attention layer follows the encoding layer and receives the context vector representation output by the encoding layer. The attention layer's core task is to assign different weights to different parts of the context vector representation, highlighting semantic information relevant to intent recognition. It calculates attention scores and attention weights, performs weighted concatenation on the context vector representations, and generates a weighted context vector representation, which it then passes to the classification layer.

[0104] The classification layer is the final layer of the model, receiving the weighted context vector representation output by the attention layer. The classification layer is a fully connected layer that performs a linear transformation on the input using a weight matrix and a bias vector. It then uses the softmax function to convert the output into a probability distribution representing the probability of the user's question falling under each predefined intent category.

[0105] Training a pre-trained language parsing model requires a large amount of text data and a clear training procedure. The first step is data preparation. This involves collecting a large amount of text data related to internet product startups, including user questions and historical responses. This data is then annotated, with each sample assigned a corresponding intent category.

[0106] During training, stochastic gradient descent (SGD) or its variants (such as Adagrad, Adadelta, and Adam) are used as the optimization algorithm. The learning rate is a key hyperparameter that controls the step size of the model's parameter updates at each iteration. An appropriate learning rate ensures rapid model convergence and avoids overfitting during training. For example, the initial learning rate can be set to a small value and then gradually decreased as training progresses.

[0107] Batch size is also a key parameter, determining the number of samples used in each training run. Larger batch sizes can improve training stability but may cause the model to become trapped in a local optimum. Smaller batch sizes can increase the model's randomness, helping it escape local optima, but may also slow training. Choosing an appropriate batch size depends on your computing resources and data size.

[0108] The training process typically proceeds through multiple iterations. In each iteration, the data is divided into batches and fed into the model for training. During each batch, the model's loss function is calculated. A commonly used loss function is the cross-entropy loss function, which measures the difference between the model's predicted probability distribution and the true labels. Based on the loss function's value, an optimization algorithm is then used to update the model's parameters, gradually reducing the loss function. After multiple rounds of iterative training, the model's performance gradually improves.

[0109] In the internet product startup scenario, the model's input data primarily consists of target user conversation requests. These requests undergo word segmentation to produce a sequence of semantic units, which serves as input to the encoding layer. Furthermore, during intent alignment, a set of contextual keywords is also input into the model. These keywords are extracted from historical session records of the target conversation requests and are relevant to the current user's question. This input data is closely centered around the internet product startup sector, covering various issues users encounter during the startup process, such as product development, marketing, and profit models.

[0110] The model's output data is the probability distribution of user questions across various pre-defined intent categories. These pre-defined intent categories are based on the characteristics and common problems of internet product startups, such as "Internet product technology selection," "Internet product marketing strategy," and "Internet product profit model design." By outputting this probability distribution, the intent category with the highest probability is determined, which serves as the initial intent category or calibrated intent category for the user's question. This output data is directly relevant to the internet product startup scenario and provides an accurate basis for subsequently matching related nodes within the pre-defined startup knowledge graph.

[0111] Step S130: Matching associated nodes in a preset entrepreneurship knowledge graph based on the intention category, wherein the associated nodes include entrepreneurship guidance knowledge entities corresponding to the intention category.

[0112] The pre-defined entrepreneurial knowledge graph is a large and complex knowledge network. Its nodes represent various entrepreneurial guidance knowledge entities, and its edges represent the relationships between these knowledge entities. To find entrepreneurial guidance knowledge relevant to the user's question, we need to match it within the knowledge graph based on the intent category. The following details the matching process.

[0113] Step S131: query a preset intention-node mapping table according to the intention category to determine an initial associated node set corresponding to the intention category.

[0114] The preset intent-node mapping table is a pre-established mapping relationship table that records the correspondence between each intent category and the nodes in the knowledge graph. By querying this mapping relationship table, you can quickly find the initial set of associated nodes corresponding to the calibrated intent category. For example, when the calibrated intent category is "Internet product marketing strategy planning", by querying the intent-node mapping table, you can determine that the corresponding initial associated nodes may include "social media promotion", "search engine optimization", "word-of-mouth marketing" and other nodes.

[0115] Step S132: extracting node attribute information of each node in the initial associated node set from the entrepreneurial knowledge graph, wherein the node attribute information includes node type, node weight, and connection strength between nodes.

[0116] After determining the initial set of associated nodes, it's necessary to extract attribute information for these nodes from the entrepreneurial knowledge graph. Node types can be categorized into technology, marketing, operations, and other categories, representing different areas of entrepreneurial knowledge. Node weight reflects the importance of a node in the knowledge graph; the higher the weight, the more critical the knowledge it represents. The connection strength between nodes indicates the closeness of the relationship between the two nodes; higher connection strength indicates a stronger correlation between the knowledge represented by the two nodes.

[0117] For example, for the "Social Media Promotion" node, whose node type is marketing, its node weight might be set to a relatively high value based on its importance in marketing. The connection strength between it and the "User Growth" node might also be high, as social media promotion often promotes user growth. Extracting this node attribute information can provide a basis for subsequent sorting and filtering of the initial set of connected nodes.

[0118] Step S133: Sorting the initial associated node set according to the node attribute information to generate a sorted candidate node sequence.

[0119] Based on the extracted node attribute information, the initial set of associated nodes is sorted. The sorting rules can comprehensively consider the node type, node weight, and the connection strength between nodes. For example, the nodes can be sorted from high to low according to the node weight, with nodes with high weight placed in front. For nodes with the same node weight, they can be further sorted according to the connection strength between the nodes, with nodes with high connection strength being prioritized. Through such a sorting process, a sorted candidate node sequence is generated, and the nodes in the sorted candidate node sequence are arranged according to their relevance and importance to the user's question intention.

[0120] Step S134: Calculate the matching score between each node in the candidate node sequence and the context-related features according to the context-related features in the semantic parsing result of the user question content.

[0121] To further select nodes that are more relevant to the user's question, it is necessary to calculate the match score between each node in the candidate node sequence and the contextual features. Each node can be represented as a feature vector, which can be generated during the knowledge graph construction process based on the node's attribute information and related knowledge. The node's feature vector is then similar to the contextual features. A common similarity calculation method is cosine similarity. The calculated similarity value is the match score between the node and the contextual features. The higher the match score, the more relevant the node is to the contextual information of the user's question.

[0122] Step S135: Filtering a target associated node set from the candidate node sequence according to the matching score.

[0123] Based on the calculated match scores, nodes with high matching scores are selected from the candidate node sequence to form the target associated node set. A matching threshold can be set to select nodes with a matching score greater than the threshold. These nodes are highly relevant to the user's question intent and context, providing more precise entrepreneurial guidance.

[0124] Step S136: performing redundant node elimination processing on the target associated node set to generate an optimized associated node set as the associated node.

[0125] In the target associated node set, there may be some redundant nodes, and the knowledge represented by these nodes has a large overlap with other nodes. In order to improve the efficiency of knowledge utilization and the accuracy of responses, the target associated node set needs to be processed to eliminate redundant nodes. The similarity between nodes can be calculated to determine whether the nodes are redundant. If the similarity between two nodes exceeds a predefined threshold, it means that the knowledge they represent has a large overlap, and one of the nodes can be retained. After the redundant node elimination process, an optimized associated node set is generated. The nodes in the optimized associated node set are both relevant to the content of the user's question and avoid knowledge redundancy. They are used as the final associated nodes for subsequent knowledge extraction and response generation.

[0126] Step S140: traverse the adjacent path of the associated node according to the context association feature, and extract the knowledge combination relationship between the associated node and the adjacent nodes.

[0127] After identifying the associated nodes, we need to further explore the relationships between these associated nodes and other nodes to obtain richer entrepreneurial guidance knowledge. The following details the steps for extracting knowledge combination relationships.

[0128] Step S141: extracting the direct adjacent node set and the indirect adjacent node set of the associated node from the entrepreneurial knowledge graph.

[0129] In the entrepreneurial knowledge graph, a directly adjacent node of an associated node refers to a node directly connected to the associated node, with a direct relationship between them. An indirect adjacent node refers to a node indirectly connected to the associated node through other nodes. By traversing the edge information of the entrepreneurial knowledge graph, the set of directly adjacent nodes and the set of indirectly adjacent nodes of the associated node can be extracted. For example, for the associated node "social media promotion," its direct adjacent nodes may include nodes such as "hot topic marketing" and "influencer collaboration," while its indirect adjacent nodes may include nodes such as "user data analysis" (indirectly connected to "social media promotion" through "hot topic marketing").

[0130] Step S142: Filtering out a first target adjacent node subset according to a first correlation score between each node in the directly adjacent node set and the context association feature.

[0131] To select directly adjacent nodes that are more relevant to the user's question, it's necessary to calculate the first correlation score between each node in the directly adjacent node set and the contextually associated features. This calculation is similar to the previously described matching score between the node and the contextually associated features, and methods such as cosine similarity can be used. Based on the calculated first correlation score, a correlation threshold is set, and nodes with a first correlation score greater than this threshold are selected to form the first target adjacent node subset. These nodes have a high correlation with the contextual information of the user's question and can provide more valuable information for subsequent knowledge combination.

[0132] Step S143: Filtering out a second target adjacent node subset according to a second correlation score between each node in the indirect adjacent node set and the context association feature.

[0133] Similarly, for the set of indirect adjacent nodes, the second correlation score between each node and the context-related feature is calculated. By setting a suitable correlation threshold, the nodes with a second correlation score greater than the threshold are screened out to form the second target adjacent node subset. Although these indirect adjacent nodes are not directly connected to the associated nodes, by calculating the correlation score, the nodes with a high correlation with the context information of the user's question content are screened out, which can provide a wider source of knowledge for knowledge combination. For example, in the Internet product startup scenario, the associated node is "social media advertising", and the indirect adjacent node "data privacy regulations" may be connected to it through the intermediate node "advertising compliance". If its second correlation score with the context-related feature is high, it will be selected into the second target adjacent node subset, because in the current startup environment, data privacy regulations have an important impact on social media advertising.

[0134] Step S144: Merge the first target adjacent node subset and the second target adjacent node subset to generate a comprehensive adjacent node set.

[0135] Merge the first target adjacent node subset and the second target adjacent node subset to form a comprehensive adjacent node set that includes directly and indirectly related adjacent nodes. During the merging process, nodes need to be deduplicated to ensure the uniqueness of each node in the set. Doing so can integrate nodes that are closely related to the associated nodes and have certain indirect associations, providing a basis for subsequent comprehensive mining of knowledge combination relationships. For example, the first target adjacent node subset includes "hot topic planning" and "Internet celebrity cooperation plan", and the second target adjacent node subset includes "data privacy regulations" and "user psychological analysis". The comprehensive adjacent node set obtained after merging and deduplication includes these nodes that supplement and expand the relevant knowledge of the associated nodes.

[0136] Step S145: traverse the connection paths between each node in the comprehensive adjacent node set and the associated node, and extract the relationship type, weight coefficient and text description information predefined in the entrepreneurial knowledge graph of the connection path as path attribute information.

[0137] In the entrepreneurial knowledge graph, each connection path between nodes has predefined attribute information. The connection paths between each node in the comprehensive adjacent node set and the associated node are traversed, and the relationship type, weight coefficient, and text description information of these paths are extracted from the knowledge graph. The relationship type can be "causal relationship", "supplementary relationship", "precondition relationship", etc., which describes the logical connection of knowledge between the two nodes. The weight coefficient reflects the importance of this connection path. The higher the weight, the closer the relationship between the two nodes. The text description information provides a detailed textual description of this relationship. For example, the connection path between the associated node "social media promotion" and the adjacent node "user growth" may have a "causal relationship" relationship type and a high weight coefficient. Because social media promotion usually has a direct impact on user growth, the text description information may be "Effective social media promotion activities can attract more users, thereby promoting the growth of the number of users."

[0138] Step S146: constructing a knowledge combination relationship between the associated node and each adjacent node according to the weight coefficient and relationship type in the path attribute information, wherein the knowledge combination relationship includes the relationship type, the normalized relationship strength and the keyword label extracted from the text description information.

[0139] First, the weight coefficients in the path attribute information are normalized to ensure that the weight coefficients of different connection paths are comparable. Normalization can use the common linear normalization method to map the weight coefficients to a specific range, such as between 0 and 1. The normalized weight coefficients are used as the relationship strength. Then, combining the relationship type and the keyword tags extracted from the text description information, a knowledge combination relationship between the associated node and each adjacent node is constructed. The keyword tags can summarize the core content of the text description information. For example, for the connection path of "social media promotion" and "user growth" mentioned above, the keyword tags extracted from the text description information may be "social media promotion", "user growth", "attracting users", etc. The knowledge combination relationship constructed in this way can clearly reflect the knowledge connection and importance between the associated node and the adjacent node.

[0140] Step S150: Generate target reply content based on the knowledge combination relationship, and return the target reply content to the user terminal to complete the question-answer interaction.

[0141] Once we have the knowledge combination relationship between the associated nodes and the adjacent nodes, we can use these relationships to generate target response content that can meet user needs. The following are the specific steps.

[0142] Step S151: performing knowledge integration processing on the associated nodes and adjacent nodes according to the relationship type and relationship description text in the knowledge combination relationship to generate an initial knowledge fragment set.

[0143] Based on the relationship type within the knowledge combination relationship, such as causal relationships and supplementary relationships, the knowledge represented by the associated and adjacent nodes is integrated. For causal relationships, the knowledge of the cause and effect nodes is combined according to causal logic; for supplementary relationships, the knowledge of the supplementary node is added to the corresponding part of the knowledge of the associated node. At the same time, the relationship description text is referenced to ensure the accuracy and completeness of the knowledge integration. For example, the associated node "social media promotion" and the adjacent node "user growth" are causally related, and the relationship description text is "Effective social media promotion activities can attract more users, thereby promoting user growth." During integration, the specific methods of social media promotion and relevant indicators of user growth can be combined to form a knowledge fragment about the impact of social media promotion on user growth. By performing this knowledge integration process on all associated and adjacent nodes, an initial set of knowledge fragments is generated.

[0144] Step S152: Sorting the initial knowledge fragment set according to the relationship strength to generate a sorted knowledge fragment sequence.

[0145] Sort the initial set of knowledge fragments according to the strength of the relationships in the knowledge combination relationship. The stronger the relationship strength of the knowledge fragment, the more important it is in the knowledge system, and the more it should be presented to the user first. Sorting can be done in descending order, with knowledge fragments with high relationship strength placed at the front. For example, for the knowledge fragment related to the associated node "social media promotion", the knowledge fragment "Hot topic marketing significantly improves the effectiveness of social media promotion" has a higher relationship strength and will be ranked higher in the sequence, while "Sharing tips in social media promotion" with a relatively lower relationship strength will be ranked at the back. This sorted sequence of knowledge fragments allows users to obtain key entrepreneurial guidance knowledge more quickly.

[0146] Step S153: Filtering out a set of core knowledge fragments that meet a preset strength threshold from the sorted knowledge fragment sequence.

[0147] Set a preset strength threshold, and filter out knowledge fragments with relationship strength greater than the threshold from the sorted knowledge fragment sequence to form a core knowledge fragment set. The setting of this threshold needs to comprehensively consider the importance of knowledge and the conciseness of the reply content. If the threshold is set too high, the core knowledge fragment set may be too streamlined, missing some valuable knowledge; if the threshold is set too low, the core knowledge fragment set may contain too much information, increasing the difficulty for users to obtain key knowledge. For example, in the Internet product startup scenario, for the knowledge fragment sequence associated with the node "product user retention strategy", the preset strength threshold can be set according to actual conditions to filter out knowledge fragments with higher relationship strength, such as "the key role of personalized recommendations in user retention" and "the relationship between the frequency of high-quality content updates and user retention" to form a core knowledge fragment set.

[0148] Step S154: performing semantic coherence verification on the core knowledge fragment set to generate a coherent knowledge text.

[0149] In order to ensure that the generated response content is readable and logical, it is necessary to perform semantic coherence verification on the core knowledge fragment set. The specific processing steps are as follows:

[0150] Step S1541: extract the subject keyword set of each knowledge segment in the core knowledge segment set, and convert the keyword set into a vector representation based on a pre-trained word vector model.

[0151] Extract topic keywords from each core knowledge segment to form a topic keyword set. For example, for the knowledge segment "The Key Role of Personalized Recommendations in User Retention," the extracted topic keywords might be "personalized recommendations" and "user retention." These keywords are then converted into vector representations using pre-trained word embedding models such as Word2Vec and GloVe. These vectors reflect the keyword's position in the semantic space and its semantic information.

[0152] Step S1542: Calculate the average cosine similarity between the topic keyword vectors of each two knowledge fragments as the semantic overlap score.

[0153] For every two knowledge fragments in the core knowledge fragment set, calculate the cosine similarity between their subject keyword vectors. Cosine similarity can measure the degree of similarity between two vectors in direction. The closer the value is to 1, the more similar the semantics of the two knowledge fragments are. For each pair of knowledge fragments, average the cosine similarities between their subject keyword vectors to obtain a semantic overlap score. For example, the subject keyword vectors of knowledge fragment A are vA1, vA2, etc., and the subject keyword vectors of knowledge fragment B are vB1, vB2, etc. Calculate the cosine similarity between vA1 and vB1, vA1 and vB2, etc., and then take the average to obtain the semantic overlap score of knowledge fragments A and B.

[0154] Step S1543: Sort the knowledge fragments from high to low according to the semantic overlap scores, and construct a priority logical connection relationship between the knowledge fragments.

[0155] Knowledge segments are sorted from high to low according to their semantic overlap scores, with segments with high semantic overlap placed adjacent to each other. Based on this sorting result, a prioritized logical connection relationship is constructed between the knowledge segments. For example, if knowledge segments A and B have high semantic overlap, they are likely to be closely connected logically. Knowledge segment A can be presented first, followed by knowledge segment B, with the relationship between them indicated by appropriate connectives, such as "in addition" or "and."

[0156] Step S1544: performing sequence adjustment processing on the core knowledge fragment set based on the priority logical connection relationship to generate an adjusted knowledge fragment sequence.

[0157] Based on the established priority logical connections, the order of the knowledge fragments within the core knowledge fragment set is adjusted. Knowledge fragments with close logical connections are arranged together to form a more semantically coherent sequence. For example, if the original order of the knowledge fragments is C, D, E, and analysis reveals that D and E have a high degree of semantic overlap and a logical connection, the adjusted order may become C, E, D.

[0158] Step S1545: inserting logical connectives into the adjusted knowledge fragment sequence to generate a preliminary coherent text.

[0159] In the adjusted sequence of knowledge fragments, insert appropriate logical connectives based on the logical relationships between the knowledge fragments. For example, for knowledge fragments that demonstrate causality, insert "because" or "so"; for knowledge fragments that demonstrate parallel relationships, insert "at the same time" or "and." By inserting logical connectives, the transitions between knowledge fragments become more natural, generating a preliminarily coherent text.

[0160] Step S1546: performing grammar correction and redundant information deletion processing on the preliminary coherent text to generate the coherent knowledge text.

[0161] A grammar check is performed on the preliminary coherent text to correct any grammatical errors, such as subject-verb inconsistency and inappropriate collocations. Redundant information, such as repeated expressions and irrelevant modifiers, is also removed to make the text more concise and clear. After grammatical correction and redundant information removal, a coherent knowledge text is generated. This coherent knowledge text is semantically coherent and logically clear, providing users with high-quality entrepreneurial guidance information.

[0162] Step S155: According to the intention category in the semantic analysis result of the user's question content, the coherent knowledge text is structurally reorganized to generate a structured reply text.

[0163] Based on the intent categories in the semantic analysis results of the user's question content, the coherent knowledge text is restructured and reorganized. Different intent categories may require different structures to present knowledge. For example, if the intent category is "Internet product market promotion strategy planning," the coherent knowledge text can be restructured according to market research, promotion channel selection, promotion event planning, etc. Categorizing relevant knowledge content into different structural sections makes the response content more organized and easier for users to understand. For example, knowledge fragments about market research can be placed in the "Market Research" section, and knowledge about promotion channel selection can be placed in the "Promotion Channel Selection" section to form a clearly structured response text.

[0164] Step S156: Fusing the structured reply text with a preset dialogue template to generate the target reply content in natural language form.

[0165] Pre-set conversation templates are pre-designed text frameworks with a natural language expression style. Structured response text is integrated with the conversation template to embed structured knowledge content into the conversation template. Conversation templates can be designed based on different scenarios and intent categories. For example, responses to entrepreneurial consultations can adopt a friendly, professional language style. During the integration process, ensure that the knowledge content matches the language style and logical structure of the conversation template. For example, a conversation template might read, "In response to your question about starting an internet product business, here are the detailed answers: First, in [Specific Aspect 1], we have [Related Knowledge Content 1]; Second, in [Specific Aspect 2], [Related Knowledge Content 2]..." The corresponding content from the structured response text is entered into the template to generate the target response content in natural language. This target response content not only contains accurate entrepreneurial guidance knowledge but also has good natural language expression, making it easier for users to understand and accept.

[0166] Step S157: Return the target reply content to the user terminal to complete the question-answer interaction.

[0167] The generated target response is sent to the user via the communication interface between the system and the user. Once the user receives the target response, a question-and-answer interaction is complete. In the context of internet product startup consulting, users can obtain targeted entrepreneurial guidance based on the target response, helping them solve problems encountered during the startup process.

[0168] Step S210: monitoring the user's feedback behavior on the target reply content, and generating a feedback behavior record.

[0169] After the question-and-answer interaction is complete, it's necessary to monitor the user's feedback on the target response. This feedback can include actions such as liking, commenting, forwarding, and asking questions again. The system records these feedback actions in real time, generating a feedback log. For example, if a user likes a target response, the system will record information such as the time of the like and user ID. If a user comments on a response, the system will record the specific content and time of the comment. These feedback logs can reflect the user's satisfaction with the response and their concerns.

[0170] Step S220: updating the weight parameter of the semantic parsing result of the user's question content according to the feedback behavior record.

[0171] Analyze feedback behavior records and update the weight parameters of the semantic parsing results of user question content based on user feedback. If the user is satisfied with the reply content, such as giving a like or positive comment, it means that the semantic parsing result is relatively accurate, and the corresponding weight parameter can be appropriately increased; if the user questions the reply content or asks again, it means that the semantic parsing result may be insufficient, and the corresponding weight parameter needs to be appropriately reduced. For example, if the user likes the reply content about "Internet product user growth strategy", then the weight parameters of the intent category and context-related features related to the strategy in the semantic parsing result can be increased; if the user comments that the reply content does not solve their core problem, then the relevant weight parameters need to be reduced. By updating the weight parameters, the semantic parsing results can more accurately reflect the user's needs.

[0172] Step S230: adjusting the node weights of associated nodes in the preset entrepreneurial knowledge graph according to the updated weight parameters.

[0173] Apply the updated weight parameters of the semantic parsing results to the preset entrepreneurial knowledge graph to adjust the node weights of the associated nodes. The node weight of the associated node reflects the importance of the node in the knowledge graph and its relevance to user needs. Adjusting the node weights according to the weight parameters of the semantic parsing results can make the knowledge graph more focused on the entrepreneurial knowledge that users are concerned about. For example, if the weight parameters related to "social media promotion" in the semantic parsing results are increased, then the node weights of the associated nodes related to "social media promotion" in the entrepreneurial knowledge graph will also increase accordingly. In this way, in the subsequent matching process, these nodes are more likely to be selected, providing users with knowledge that better meets their needs.

[0174] Step S240: Dynamically optimize the entrepreneurial knowledge graph based on the adjusted node weights to generate an optimized entrepreneurial knowledge graph.

[0175] The entrepreneurial knowledge graph is dynamically optimized based on the adjusted node weights. This optimization process includes reordering nodes, adjusting edge weights, and updating knowledge relationships. For example, for associated nodes with increased node weights, their positions in the knowledge graph are adjusted to make them more accessible. Edge weights connected to these nodes are adjusted based on changes in node weights to reflect changes in the closeness of the relationships between nodes. Through dynamic optimization, the entrepreneurial knowledge graph can better adapt to changing user needs and improve the accuracy of knowledge retrieval and matching.

[0176] Step S250: Apply the optimized entrepreneurial knowledge graph to the associated node matching process in the subsequent question-answer interaction process.

[0177] In subsequent Q&A interactions, the optimized entrepreneurial knowledge graph is used to match related nodes. When a new user asks a question, the question is semantically parsed according to the previous process, and then the related nodes in the optimized entrepreneurial knowledge graph are matched based on the semantic parsing results. Because the optimized entrepreneurial knowledge graph better meets user needs, it can more accurately find related nodes related to the user's question, thereby generating more precise targeted responses and improving the quality and effectiveness of the Q&A interaction. By continuously monitoring user feedback, updating weight parameters, and optimizing the knowledge graph, the system can continuously learn and improve, providing users with better entrepreneurial guidance services.

[0178] Figure 2 A schematic diagram illustrates exemplary hardware and software components of a dialogue-based intelligent question-answering system 100 for entrepreneurial mentors, 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 dialogue-based intelligent question-answering system 100 for entrepreneurial mentors to perform the functions of the present application.

[0179] The dialogue-based entrepreneurial mentor intelligent question-answering system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the dialogue-based entrepreneurial mentor intelligent question-answering 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.

[0180] For example, the intelligent question-answering system 100 for entrepreneurial mentors based on a dialogue system may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the intelligent question-answering system 100 for entrepreneurial mentors based on a dialogue system 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 intelligent question-answering system 100 for entrepreneurial mentors based on a dialogue system also includes an I / O interface 150 between the computer and other input and output devices.

[0181] For ease of explanation, only one processor is described in the dialogue system-based entrepreneurial mentor intelligent question and answer system 100. However, it should be noted that the dialogue system-based entrepreneurial mentor intelligent question and answer system 100 in this application can also include multiple processors, so the steps performed by one processor described in this application can also be performed jointly or individually by multiple processors. For example, if the processor of the dialogue system-based entrepreneurial mentor intelligent question and answer system 100 executes steps A and B, it should be understood that steps A and B can 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.

[0182] 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 intelligent question-and-answer method for entrepreneurial mentors based on the dialogue system is implemented.

[0183] 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. An intelligent question-answering method for entrepreneurial mentors based on a dialogue system, characterized in that: The method comprises: Obtaining a target conversation request input by a user, wherein the target conversation request includes user question content related to entrepreneurship guidance; Performing semantic parsing on the target conversation request to generate a semantic parsing result of the user's question content, wherein the semantic parsing result includes an intent category and context-related features; Matching associated nodes in a preset entrepreneurship knowledge graph based on the intention category, wherein the associated nodes include entrepreneurship guidance knowledge entities corresponding to the intention category; Extracting a set of direct adjacent nodes and a set of indirect adjacent nodes of the associated node from the entrepreneurial knowledge graph; Filtering a first target adjacent node subset according to a first correlation score between each node in the set of directly adjacent nodes and the contextual correlation feature; Screening out a second target adjacent node subset based on a second correlation score between each node in the indirect adjacent node set and the contextual correlation feature; Merging the first target adjacent node subset and the second target adjacent node subset to generate a comprehensive adjacent node set; Traversing the connection paths between each node in the comprehensive adjacent node set and the associated node, extracting the relationship type, weight coefficient, and text description information predefined in the entrepreneurial knowledge graph for the connection path as path attribute information, wherein the weight coefficient reflects the importance of the connection path. The higher the weight coefficient, the closer the relationship between the two nodes. Constructing a knowledge combination relationship between the associated node and each adjacent node based on the weight coefficient and relationship type in the path attribute information, wherein the knowledge combination relationship includes the relationship type, the normalized weight coefficient, and the keyword tags extracted from the text description information, and the normalized weight coefficient serves as the relationship strength; Target reply content is generated based on the knowledge combination relationship, and the target reply content is returned to the user end to complete the question-answer interaction.

2. The intelligent question-answering method for entrepreneurial mentors based on a dialogue system according to claim 1 is characterized in that: The performing semantic parsing on the target session request to generate a semantic parsing result of the user's question content includes: Performing word segmentation processing on the user's question content to obtain multiple semantic unit sequences; Calling a pre-trained language parsing model to perform intent recognition processing on the semantic unit sequence to determine the initial intent category of the user's question content; Traversing the historical session records of the target session request, and extracting a set of context keywords associated with the user's question content in the historical session records; Inputting the context keyword set into the language parsing model, calibrating the initial intent category, and generating a calibrated intent category; Constructing context-related features of the user's question content based on the mapping relationship between the calibrated intent category and the context keyword set; The calibrated intent category is combined with the context-related features to generate the semantic parsing result.

3. The intelligent question-answering method for entrepreneurial mentors based on a dialogue system according to claim 2 is characterized in that: The calling of a pre-trained language parsing model to perform intent recognition processing on the semantic unit sequence to determine the initial intent category of the user's question content includes: Inputting the semantic unit sequence into the encoding layer of the language parsing model for vector conversion processing to generate a context vector representation of the semantic unit sequence; Inputting the context vector representation into the attention layer of the language parsing model for feature weight assignment processing to generate a weighted context vector representation; Inputting the weighted context vector representation into the classification layer of the language parsing model for intent probability prediction processing to generate a probability distribution of the user question content under each preset intent category; Determining, according to the probability distribution, the intent category corresponding to the maximum probability as the initial intent category; Obtaining an intermediate feature vector generated by the language parsing model in the process of generating the probability distribution, and using the intermediate feature vector as a latent semantic feature of the user's question content; According to the association relationship between the potential semantic features and the initial intention category, a semantic feature index of the user question content is constructed.

4. The intelligent question-answering method for entrepreneurial mentors based on a dialogue system according to claim 2 is characterized in that: The traversing the historical session records of the target session request and extracting a set of context keywords associated with the user question content in the historical session records includes: Obtaining from the user terminal a historical conversation record within a preset time period before the current conversation round, the historical conversation record including a plurality of historical question contents and their corresponding historical reply contents; Perform keyword extraction processing on the historical question content to obtain a historical question keyword set; Perform keyword extraction processing on the historical reply content to obtain a historical reply keyword set; Performing similarity matching processing on the semantic unit sequence of the user question content and the historical question keyword set to determine a first similarity score between the user question content and each historical question content; Performing similarity matching processing on the semantic unit sequence of the user question content and the historical reply keyword set to determine a second similarity score between the user question content and each historical reply content; Filtering a target historical keyword set associated with the user's question content according to a weighted sum of the first similarity score and the second similarity score; The target historical keyword set is deduplicated and sorted to generate the context keyword set.

5. The intelligent question-answering method for entrepreneurial mentors based on a dialogue system according to claim 3 is characterized in that: The step of inputting the context keyword set into the language parsing model, calibrating the initial intent category, and generating a calibrated intent category includes: Inputting the context keyword set into the encoding layer of the language parsing model for vector conversion processing to generate a context vector set of the context keyword set, and performing average pooling processing on the context vector set to generate an aggregated context vector; Acquire the potential semantic features of the user's question content according to the semantic feature index of the user's question content; fusing the aggregated context vector with the latent semantic feature to generate a fused semantic feature vector; Inputting the fused semantic feature vector into the classification layer of the language parsing model for intention probability prediction processing to generate a calibrated intention probability distribution; Determine, based on the calibrated intent probability distribution, the intent category corresponding to the maximum probability as the candidate intent category, and record the probability difference between the probability value of the candidate intent category and the probability value of the initial intent category; Calculating a semantic distance score between the candidate intent category and the initial intent category; If the semantic distance score exceeds a preset distance threshold and the probability difference exceeds a preset probability threshold, the candidate intent category is used as the calibrated intent category; otherwise, the initial intent category is retained as the calibrated intent category.

6. The intelligent question-answering method for entrepreneurial mentors based on a dialogue system according to claim 1 is characterized in that: The matching of associated nodes in a preset entrepreneurial knowledge graph based on the intention category includes: Querying a preset intent-node mapping table according to the intent category to determine an initial associated node set corresponding to the intent category; Extracting node attribute information of each node in the initial associated node set from the entrepreneurial knowledge graph, the node attribute information including node type, node weight, and connection strength between nodes. The node weight reflects the importance of the node in the entrepreneurial knowledge graph. The higher the node weight, the more critical the knowledge represented by the node. The connection strength between nodes represents the closeness of the relationship between two nodes. The higher the connection strength, the stronger the association between the knowledge represented by the two nodes. Sorting the initial associated node set according to the node attribute information to generate a sorted candidate node sequence; Calculating a matching score between each node in the candidate node sequence and the context-related features in the semantic parsing result of the user's question content; Filtering a target associated node set from the candidate node sequence according to the matching score; Redundant nodes are eliminated from the target associated node set to generate an optimized associated node set as the associated node.

7. The intelligent question-answering method for entrepreneurial mentors based on a dialogue system according to claim 1 is characterized in that: Generating target reply content based on the knowledge combination relationship includes: Performing knowledge integration processing on the associated nodes and adjacent nodes according to the relationship type and relationship description text in the knowledge combination relationship to generate an initial knowledge fragment set; Sorting the initial knowledge fragment set according to the relationship strength to generate a sorted knowledge fragment sequence; Filtering a set of core knowledge fragments that meet a preset strength threshold from the sorted knowledge fragment sequence; Performing semantic coherence verification on the core knowledge fragment set to generate a coherent knowledge text; According to the intention category in the semantic analysis result of the user's question content, the coherent knowledge text is structurally reorganized to generate a structured reply text; Fusing the structured reply text with a preset dialogue template to generate the target reply content in natural language form; The step of performing semantic coherence verification on the core knowledge fragment set to generate a coherent knowledge text includes: Extracting a subject keyword set of each knowledge segment in the core knowledge segment set, and converting the keyword set into a vector representation based on a pre-trained word vector model; Calculate the average cosine similarity between the topic keyword vectors of each two knowledge fragments as the semantic overlap score; Sorting the semantic overlap scores from high to low to construct a priority logical connection relationship between the knowledge fragments; Performing sequence adjustment processing on the core knowledge fragment set based on the priority logical connection relationship to generate an adjusted knowledge fragment sequence; inserting logical connectives into the adjusted sequence of knowledge fragments to generate a preliminary coherent text; The preliminary coherent text is subjected to grammatical correction and redundant information deletion processing to generate the coherent knowledge text.

8. The intelligent question-answering method for entrepreneurial mentors based on a dialogue system according to claim 1 is characterized in that: After returning the target reply content to the user terminal to complete the question-answer interaction, the method further includes: Monitoring the user's feedback behavior on the target reply content and generating a feedback behavior record; Updating the weight parameter of the semantic parsing result of the user's question content according to the feedback behavior record; Adjusting the node weights of associated nodes in the preset entrepreneurial knowledge graph according to the updated weight parameters; Dynamically optimizing the entrepreneurial knowledge graph based on the adjusted node weights to generate an optimized entrepreneurial knowledge graph; The optimized entrepreneurial knowledge graph is applied to the associated node matching process in the subsequent question-answering interaction process.

9. An intelligent question-answering system for entrepreneurial mentors based on a dialogue system, 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 entrepreneurial mentor intelligent question-answering method based on the dialogue system as described in any one of claims 1 to 8.

Images