Intelligent question-answering method and system for low-altitude economic education on interactive learning platform

Abstract

The present invention provides a low-altitude economic education intelligent question-answering method and system for an interactive learning platform, which relates to the technical field of natural language processing. First, a low-altitude economic education question data set containing multiple user interaction data is obtained, and a question feature set of the user interaction data is extracted, including semantic vector features, domain association features and context association features. Based on the question feature set, a domain knowledge feature set in a low-altitude economic education knowledge base is retrieved, and cross-dimensional feature fusion processing is performed on the question feature set and the domain knowledge feature set to generate an optimized answer statement for the user question statement. Finally, based on the difference characteristics between the optimized answer statement and the historical answer statement, the answer service strategy of the interactive learning platform is dynamically adjusted, thereby improving the accuracy of question and answer and the platform adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to an intelligent question-answering method and system for low-altitude economic education on an interactive learning platform. Background Art

[0002] In existing interactive learning platforms, the intelligent question-and-answer function for low-altitude economic education has many shortcomings. On the one hand, most traditional question-and-answer systems simply match the surface text of user questions, lacking a deep understanding of question semantics, domain relevance, and context, resulting in poor accuracy and pertinence of answers. For example, for some questions with similar expressions but different meanings, it is difficult for the question-and-answer system to accurately distinguish and give precise answers. On the other hand, these question-and-answer systems often search for answers in isolation from the knowledge base when answering questions, without effectively integrating question features with domain knowledge, making the answers relatively simple and one-sided. Moreover, existing question-and-answer systems lack a mechanism to dynamically adjust the answer service strategy, cannot optimize based on the user's historical answers, and cannot adapt to the diverse needs of different users, reducing the user's learning experience and effectiveness on the interactive learning platform. 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 low-altitude economic education on an interactive learning platform, the method comprising:

[0004] Obtaining a low-altitude economic education question data set submitted by a user on an interactive learning platform, wherein the low-altitude economic education question data set includes a plurality of user interaction data, each user interaction data consisting of at least one user question statement and a corresponding historical answer statement;

[0005] Extracting a question feature set from the user interaction data, the question feature set including a semantic vector feature of the user question statement, a field association feature between the user question statement and the field of low-altitude economic education, and a context association feature of the user question statement;

[0006] Retrieving a domain knowledge feature set in a low-altitude economic education knowledge base based on the problem feature set;

[0007] Performing cross-dimensional feature fusion processing on the question feature set and the domain knowledge feature set to generate an optimized response statement for the user question statement;

[0008] Based on the difference characteristics between the optimized response statement and the historical response statements in the user interaction data, the response service strategy of the interactive learning platform is dynamically adjusted.

[0009] On the other hand, an embodiment of the present invention also provides a low-altitude economic education intelligent question-and-answer system for an interactive learning platform, 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 obtains a low-altitude economic education problem data set containing multiple user interaction data, and extracts problem feature sets such as semantic vector features, domain-related features, and context-related features of the user interaction data, and then retrieves the domain knowledge feature set in the low-altitude economic education knowledge base. After cross-dimensional feature fusion, an optimized response statement is generated, which can fully combine the problem features with the domain knowledge to improve the accuracy and comprehensiveness of the response. At the same time, the response service strategy is dynamically adjusted based on the difference features between the optimized response statement and the historical response statement, so that the interactive learning platform can continuously optimize the response service according to the actual situation of the user, better meet the diverse needs of different users in low-altitude economic education learning, improve the user's learning experience and learning effect, and enhance the intelligence and adaptability of the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the intelligent question-answering method for low-altitude economic education on an interactive learning platform provided by an embodiment of the present invention.

[0012] Figure 2 Schematic diagram of exemplary hardware and software components of the low-altitude economic education intelligent question-answering system for an interactive learning platform 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 low-altitude economic education on an interactive learning platform provided by an embodiment of the present invention. The intelligent question-answering method for low-altitude economic education on an interactive learning platform is introduced in detail below.

[0014] Step S110: obtaining a low-altitude economic education question data set submitted by users on the interactive learning platform, wherein the low-altitude economic education question data set includes a plurality of user interaction data, and each user interaction data is composed of at least one user question statement and a corresponding historical answer statement.

[0015] In this embodiment, in the daily operation of the interactive learning platform, many users log in and raise questions related to low-altitude economic education. The interactive learning platform has a complete data receiving and storage mechanism. When the user enters a question in the question interface of the interactive learning platform, the user's question statement can be quickly captured. At the same time, if the user's question statement has a corresponding answer record in the past, the corresponding historical answer statement can be retrieved from the historical data storage area. Each group of user question statements and historical answer statements are combined into a user interaction data. With the passage of time and the participation of many users, a large amount of user interaction data is accumulated, which together constitutes a low-altitude economic education question data set.

[0016] To provide a more intuitive explanation, the interactive learning platform uses a distributed storage system to manage this data. When a user submits a question about the efficiency evaluation of new modes of transportation in the low-altitude economy, the question statement can be encoded according to a set format, and a search is performed to determine if there are historical responses to similar questions. If so, the corresponding historical response statement is associated with the current user's question statement and stored in a data node in the distributed storage system. As more users submit different types of questions, each data node is continuously updated and expanded, ultimately forming a complete dataset of low-altitude economy education questions.

[0017] Step S120: extracting a question feature set of the user interaction data, wherein the question feature set includes a semantic vector feature of the user question statement, a field association feature between the user question statement and the field of low-altitude economic education, and a context association feature of the user question statement.

[0018] Step S121: performing semantic coding processing on the user question sentence to generate a semantic vector feature of the user question sentence, wherein the semantic vector feature includes a keyword distribution weight of a core concept in the user question sentence.

[0019] Step S1211: calling a pre-trained low-altitude economic domain semantic encoder to perform word segmentation processing on the user question sentence to obtain multiple semantic word units.

[0020] The pre-trained semantic encoder for the low-altitude economy domain is trained based on a massive amount of low-altitude economy-related text data. It is designed to accurately identify specialized vocabulary, terminology, and common linguistic expressions within the field. When receiving a user question, the encoder performs word segmentation. For example, a user might ask, "Given the development of the low-altitude economy, what are the significant differences in cost control and timeliness between emerging low-altitude logistics and distribution models and traditional models?" The encoder will break the question into multiple semantic units, such as "low-altitude economy," "development trend," "emerging," "low-altitude logistics and distribution model," "traditional model," "cost control," "timeliness," and "significant differences." This process is based on the encoder's internal lexical analysis and language model, which accurately identifies each semantic unit based on the semantic and grammatical rules of the vocabulary.

[0021] Step S1212: performing context dependency analysis on the semantic word units to generate a local context weight for each semantic word unit, where the local context weight is used to represent the semantic contribution of the semantic word unit in the user question sentence.

[0022] After obtaining semantic word units, we need to analyze their dependencies within the context of the user's question. This is because within a sentence, individual semantic word units are interconnected and contribute differently to the semantics of the entire sentence. For example, in the question sentence above, "low-altitude logistics distribution model" and "traditional model" are the core comparison objects and play a key role in semantic expression, so their local context weights are relatively high. In contrast, expressions like "under..." primarily provide background information and have a relatively low local context weight.

[0023] To determine the local context weight, a deep learning-based attention mechanism can be used. This mechanism calculates the degree of association between each semantic word unit and other semantic word units, and by learning the semantic structure and information flow in the sentence, it assigns an appropriate weight value to each semantic word unit. For example, for the semantic word unit "low-altitude logistics distribution model," the attention mechanism will consider its association with other semantic word units such as "cost control" and "timeliness," and determine its local context weight based on the closeness of the association.

[0024] Step S1213: performing weighted aggregation processing on the semantic word units according to the local context weights to generate an initial semantic vector of the user question sentence.

[0025] After obtaining the local context weight of each semantic word unit, weighted aggregation processing can be performed. Assume that the set of semantic word units is {u1, u2, …, un}, and the corresponding set of local context weights is {w1, w2, …, wn}. The process of weighted aggregation is to multiply each semantic word unit with its corresponding local context weight, and then combine these products. Specifically, for each semantic word unit ui, multiply it with the corresponding weight wi to obtain wi*ui, and then combine all wi*ui to form a vector representation, which is the initial semantic vector of the user's question statement. This initial semantic vector comprehensively considers the semantic contribution of each semantic word unit and can more accurately express the semantic information of the user's question statement.

[0026] Step S1214: Input the initial semantic vector into the semantic optimization model of the low-altitude economy field, perform domain knowledge enhancement processing on the initial semantic vector through knowledge graph embedding and term context association analysis, and generate semantic vector features containing the distribution characteristics of low-altitude economy professional terms.

[0027] The low-altitude economy semantic optimization model is a model designed for the low-altitude economy sector that combines knowledge graphs and term context association analysis technology. The knowledge graph is a knowledge base that contains various concepts, entities, and the relationships between them in the low-altitude economy sector. When the initial semantic vector is input into the low-altitude economy semantic optimization model, the low-altitude economy semantic optimization model will embed the relevant knowledge in the knowledge graph into the initial semantic vector. For example, if the initial semantic vector contains the concept of "low-altitude logistics distribution model", the low-altitude economy semantic optimization model will search the knowledge graph for various related attributes and relationships, such as the operating process of the model, related technologies, etc., and integrate this knowledge information into the initial semantic vector.

[0028] At the same time, term context analysis further analyzes the meaning and associations of specialized terms in the initial semantic vector across different contexts. In the low-altitude economy sector, some specialized terms may have multiple meanings. Term context analysis accurately determines their specific semantics within the current user's question. Through knowledge graph embedding and term context analysis, the initial semantic vector is enhanced with domain knowledge, ultimately generating a semantic vector feature that incorporates the distribution characteristics of specialized low-altitude economy terms. This semantic vector feature more comprehensively and accurately reflects the semantic information of the user's question and includes the keyword distribution weights for core concepts.

[0029] Step S122: extracting the response semantic features of the historical response sentences in the user interaction data, performing time-series alignment processing on the response semantic features and the semantic vector features, and generating context-related features of the user question sentence.

[0030] In this embodiment, the semantic extraction module can be used to process the historical response sentences to extract the semantic features of the response. The semantic extraction module performs part-of-speech tagging, syntactic analysis and other operations on the historical response sentences based on natural language processing technology, so as to identify the key information and semantic structure in the sentences. For example, for the historical response sentence "The traditional low-altitude logistics distribution model mainly relies on economies of scale in terms of cost control, while the emerging model pays more attention to the efficiency improvement brought about by technological innovation", the semantic extraction module will identify key information such as "traditional model", "cost control", "scale effect", "emerging model", "technological innovation" and "efficiency improvement", and convert this information into a feature vector, which is the semantic feature of the response.

[0031] After obtaining the response semantic features, they need to be temporally aligned with the semantic vector features of the previously generated user question. The purpose of temporal alignment is to match the two features in the temporal dimension to reflect the contextual relationship between the user question and the historical responses. The dynamic time warping (DTW) algorithm can be used to achieve temporal alignment. This algorithm finds an optimal matching path between the two feature sequences, aligning them as closely as possible in time. In this way, the response semantic features and the semantic vector features are fused to generate contextual features of the user question. These contextual features reflect the historical context of the user question, providing richer contextual information for subsequent processing.

[0032] Step S123: Determine the domain-related features of the user question statement based on the term coverage of the user question statement in the field of low-altitude economic education. The domain-related features are used to quantify the strength of the association between the user question statement and the core knowledge system of low-altitude economic education.

[0033] To determine the domain-relevance characteristics of user question statements, we need to calculate the term coverage of the user question statement in the field of low-altitude economic education. First, we need to establish a term dictionary for the low-altitude economic education field. This term dictionary contains all professional terms and common words in this field. Then, we compare the user question statement with this term dictionary. For each term in the user question statement, we check whether it appears in the term dictionary. If so, the term is considered a domain term.

[0034] For example, for a user question like "Amid the development of the low-altitude economy, what are the significant differences between emerging low-altitude logistics and distribution models and traditional models in terms of cost control and timeliness?", a comparison with the term dictionary reveals that terms like "low-altitude economy," "low-altitude logistics and distribution model," "cost control," and "timeliness" all fall under domain terms. The proportion of these domain terms in the user's question is calculated, and this proportion represents the term coverage.

[0035] The higher the term coverage, the stronger the correlation between the user's question and the core knowledge system of low-altitude economic education. Therefore, term coverage can be directly used as a domain-related feature, which can quantify the strength of the correlation between the user's question and the core knowledge system of low-altitude economic education. This domain-related feature plays an important role in the subsequent retrieval and matching process, helping the system to more accurately screen domain knowledge related to the user's question.

[0036] Step S124: performing feature splicing on the semantic vector features, the context-related features, and the domain-related features to generate the problem feature set.

[0037] After obtaining the semantic vector features, contextual features, and domain-related features, they need to be concatenated. Feature concatenation involves connecting different feature vectors in a set order to form a longer feature vector. For example, if the dimension of the semantic vector features is m, the dimension of the contextual features is n, and the dimension of the domain-related features is p, then the dimension of the resulting problem feature set after concatenating them is m+n+p.

[0038] During the concatenation process, it's important to maintain the order and dimensionality of the feature vectors. Semantic features can be placed first, followed by contextual features, and finally domain-specific features. The resulting concatenated question feature set integrates the semantic information of the user's question, contextual information, and information related to the field of low-altitude economic education, providing a comprehensive and rich feature foundation for subsequent retrieval and answer generation.

[0039] Step S130: searching for a domain knowledge feature set in a low-altitude economic education knowledge base based on the problem feature set.

[0040] Step S131: performing multi-level knowledge matching in the low-altitude economic education knowledge base according to the problem feature set, wherein the multi-level knowledge matching includes a core concept matching layer, an application scenario matching layer, and a technical standard matching layer.

[0041] The Low-Altitude Economy Education Knowledge Base is a database containing extensive knowledge on the low-altitude economy, organized into different categories and hierarchies. Multi-level knowledge matching begins with the core concept matching layer. This layer focuses on the semantic vector features within the question feature set, specifically the keyword distribution weights. Based on these keywords, the Low-Altitude Economy Education Knowledge Base searches for directly related core knowledge points.

[0042] For example, for the keyword "low-altitude logistics distribution model" in the semantic vector feature in the problem feature set, all core knowledge points related to "low-altitude logistics distribution model" are searched in the low-altitude economic education knowledge base, such as the definition and basic principles of the model.

[0043] Next comes the application scenario matching layer. This layer primarily searches for extended knowledge points related to the application scenario implied by the user's question statement based on contextual features within the question feature set. For example, if the user's question statement mentions a comparison of low-altitude logistics distribution models under a specific development trend, the low-altitude economic education knowledge base will be searched for extended knowledge points such as application scenarios of this model under this development trend, as well as relevant cases and data.

[0044] Finally, there's the technical standards matching layer. This layer uses the domain-related features in the question feature set to filter out standard knowledge points that match the technical standards category of the user's question. For example, if the user's question involves cost control and timeliness in low-altitude logistics distribution models, the low-altitude economic education knowledge base will be searched for relevant industry technical standards and specifications.

[0045] Step S132: In the core concept matching layer, based on the keyword distribution weights in the semantic vector features, the core knowledge point features directly associated with the user question statement in the low-altitude economic education knowledge base are retrieved.

[0046] At the core concept matching layer, keyword distribution weights within the semantic vector features play a key role. These weights reflect the importance of different keywords in the user's question. These weights can be used to filter and sort knowledge within the knowledge base. For keywords with high weights, the search for directly related core knowledge points within the low-altitude economic education knowledge base is focused.

[0047] For example, for the keyword "low-altitude logistics distribution model," since it has a high weight in the semantic vector feature, we can search the low-altitude economic education knowledge base for all core knowledge points related to this model, including its core definition, basic components, and main types. These core knowledge points are extracted and characterized to form core knowledge point features. This core knowledge point feature contains core concept information directly related to the user's question statement.

[0048] Step S133: In the application scenario matching layer, based on the contextual association features of the user question sentence, extended knowledge point features associated with the implicit application scenario of the user question sentence in the low-altitude economic education knowledge base are retrieved.

[0049] The context-related features reflect the historical context and actual application information of the user's question statement. In the application scenario matching layer, the context-related features can be used to search for extended knowledge points related to the implicit application scenario of the user's question in the low-altitude economic education knowledge base.

[0050] For example, based on contextual features, we can determine that a user's question is about comparing emerging and traditional low-altitude logistics distribution models within the context of a specific trend in the low-altitude economy. We can then search the low-altitude economy education knowledge base for specific application cases of these models within this context, such as the cost control strategies and timeliness performance of emerging models adopted by different companies in actual operations. These extended knowledge points are extracted and characterized to form extended knowledge point features, enriching the relevant knowledge about the user's question and providing more practical application information.

[0051] Step S134: In the technical standard matching layer, based on the field association features, standard knowledge point features in the low-altitude economic education knowledge base that match the technical standard category to which the user question statement belongs are screened.

[0052] The domain-related feature quantifies the strength of the correlation between the user's question and the core knowledge system of low-altitude economic education. It also reflects the technical standard category to which the user's question belongs. At the technical standard matching layer, this domain-related feature can be used to filter out standard knowledge points in the low-altitude economic education knowledge base that match the technical standard category to which the user's question belongs.

[0053] For example, if domain-related features indicate that a user's question involves cost control and timeliness in low-altitude logistics distribution models, the relevant industry technical standards and specifications, such as cost accounting standards and timeliness evaluation indicators, can be searched for in the Low-Altitude Economic Education Knowledge Base. These standard knowledge points are extracted and characterized to form standard knowledge point features, ensuring that the response meets the industry's technical standards and specifications.

[0054] Step S135: hierarchically fuse the core knowledge point features, the extended knowledge point features and the standard knowledge point features according to preset priorities to generate the domain knowledge feature set, wherein the core knowledge point features retain the highest priority by default in subsequent dynamic weight allocation.

[0055] After obtaining the core knowledge point features, extended knowledge point features, and standard knowledge point features, they need to be hierarchically fused. Preset priorities dictate the importance of different features in the fusion process. Since core knowledge point features are the core concepts directly related to user questions, they are given the highest priority by default in the subsequent dynamic weighting.

[0056] The fusion process involves concatenating the three feature sets in order of priority. First, the core knowledge point features are placed first, followed by the extended knowledge point features and the standard knowledge point features. The resulting domain knowledge feature set integrates multiple aspects of knowledge, including core concepts, application scenarios, and technical standards, providing comprehensive and accurate knowledge support for subsequent response generation.

[0057] Step S140: performing cross-dimensional feature fusion processing on the question feature set and the domain knowledge feature set to generate an optimized response statement for the user question statement.

[0058] Step S141: performing knowledge adaptation analysis on the semantic vector features in the problem feature set to determine the semantic matching degree between the semantic vector features and each knowledge point in the domain knowledge feature set.

[0059] Knowledge fit analysis focuses on the semantic similarity between the semantic vector features in the question feature set and the knowledge points in the domain knowledge feature set. Vector similarity calculation methods can be used to determine semantic fit. For example, the cosine similarity algorithm can be used to compare the semantic vector features with the feature vector of each knowledge point in the domain knowledge feature set.

[0060] For the semantic vector feature and the feature vector of a knowledge point, calculate the cosine value between them. The closer the cosine value is to 1, the more similar the directions of the two vectors are, that is, the higher the semantic match; the closer the cosine value is to 0, the lower the semantic match. In this way, a semantic match value is calculated for each knowledge point in the domain knowledge feature set.

[0061] Step S142: constructing a dynamic weight allocation mechanism based on the semantic matching degree, prioritizing the knowledge points in the domain knowledge feature set, and obtaining a priority ranking result.

[0062] Based on the calculated semantic matching degree, a dynamic weighting mechanism is constructed. Knowledge points with higher semantic matching degrees receive higher weights in the weighting distribution. For example, semantic matching degree can be used as an important parameter for weighting, and converted into a weight value through a linear or nonlinear function.

[0063] Then, the knowledge points in the domain knowledge feature set are prioritized based on these weights. Knowledge points with higher weights are ranked higher in the ranking. This prioritized ranking reflects the relevance of each knowledge point to the user's question, providing a basis for subsequent feature fusion and answer generation.

[0064] Step S143: According to the priority sorting result, feature dimension projection processing is performed on the priority knowledge point features of the top N priority sorting points and the context-related features in the question feature set respectively to generate projected knowledge features and projected question features of the same dimension, and the projected knowledge features and projected question features are fused based on the bidirectional attention mechanism to generate knowledge-enhanced semantic features.

[0065] Based on the priority sorting results, select the top N priority knowledge point features. Perform feature dimension projection on these priority knowledge point features and the contextual features in the question feature set. The purpose of feature dimension projection is to convert feature vectors of different dimensions to the same dimension for subsequent fusion processing.

[0066] Feature dimension projection can be achieved using linear transformation or nonlinear mapping methods. For example, priority knowledge point features and context-related features can be projected into the same feature space through a projection matrix to generate projected knowledge features and projected question features.

[0067] The projected knowledge features and the projected question features are then fused using a bidirectional attention mechanism. This mechanism simultaneously focuses on the relationship between the projected knowledge features and the projected question features, fusing their information by learning the attention weights between them. Specifically, the attention mechanism calculates the degree of correlation between each element in the projected knowledge features and the projected question features, assigns attention weights based on the correlation, and then combines these weighted features to generate a knowledge-enhanced semantic feature. This knowledge-enhanced semantic feature integrates prioritized knowledge points and contextual information, enhancing semantic expression capabilities.

[0068] Step S144: input the knowledge-enhanced semantic features into a response generation model based on a multi-level attention mechanism to generate an initial response sentence containing low-altitude economic education field terms and knowledge point source identifiers.

[0069] The response generation model based on a multi-level attention mechanism consists of several key modules and layers to effectively process knowledge-enhanced semantic features and generate response statements. The model primarily comprises an input layer, a multi-level attention layer, and an output layer. The input layer receives the knowledge-enhanced semantic features and uses them as starting data for the response generation model. The multi-level attention layer is the core component of the response generation model and consists of multiple attention sub-layers, each with different focus and functions.

[0070] Within the multi-layer attention layer, the first is the word-level attention sub-layer. This word-level attention sub-layer focuses on each word in the knowledge-enhanced semantic feature and calculates the importance weight of each word when generating the response sentence. By allocating attention at the word level, the response generation model can highlight key low-altitude economic education terms and ensure that these terms are accurately expressed in the response. For example, for the knowledge-enhanced semantic feature involving low-altitude logistics and distribution models, the word-level attention sub-layer will identify key terms such as "low-altitude logistics" and "distribution model" and assign them higher weights.

[0071] Next comes the sentence-level attention sublayer, which considers the structure and semantic information of knowledge-enhanced semantic features at the sentence level and calculates the importance of each sentence in response generation. This helps the model organize the logical order of sentences, making the generated responses coherent and logical. For example, when describing the advantages and disadvantages of the low-altitude logistics distribution model, the sentence-level attention sublayer will arrange the sentence order of the advantages and disadvantages in a reasonable manner, making the response easier to understand.

[0072] Next comes the chapter-level attention sub-layer. This sub-layer analyzes the knowledge-enhanced semantic features from the perspective of the entire chapter, considering the relevance of the response sentence to the entire knowledge system of low-altitude economic education. This ensures that the generated response conforms to the logic and norms of the field at a macro level and covers the necessary knowledge points. For example, when answering a question about the development trend of the low-altitude economy, the chapter-level attention sub-layer comprehensively considers development trends in different aspects to ensure a comprehensive response.

[0073] Multi-layer attention layers integrate and filter information from knowledge-enhanced semantic features by continuously adjusting and assigning attention weights. During this process, the response generation model weights the knowledge-enhanced semantic features according to the attention weights, highlighting important information and deemphasizing less important information.

[0074] After being processed by the multi-layer attention layer, the information is passed to the output layer. The main task of the output layer is to generate a specific response sentence based on the processing results of the multi-layer attention layer. It combines the language expression and grammatical rules in the field of low-altitude economic education to convert the processed information into natural and fluent text. At the same time, the output layer will embed the knowledge point source identifier in the generated response sentence. The embedding of the knowledge point source identifier is achieved by extracting and annotating the knowledge point related information recorded during the processing of the multi-layer attention layer. For example, when the response generation model processes the knowledge enhancement semantic features, it will record the source of each knowledge point. When outputting the response sentence, this source information will be added to the sentence in the form of a set identifier so that the user can understand the knowledge source of the response content.

[0075] Through the above steps, the response generation model based on the multi-level attention mechanism can transform the knowledge-enhanced semantic features into initial response sentences containing low-altitude economic education field terms and knowledge point source identifiers.

[0076] Step S145: Perform interpretability optimization processing on the initial response statement, embed the knowledge point source identifier corresponding to the user question statement in the initial response statement, and generate the optimized response statement.

[0077] The purpose of optimizing the interpretability of initial responses is to increase the transparency and credibility of responses, allowing users to clearly understand the rationale behind the responses. This process begins with a structural analysis of the initial responses. Natural language processing techniques are used to identify the individual sentences and key information within the initial responses. For example, for an initial response regarding low-altitude logistics cost control strategies, the sentences describing different cost control strategies and the key terms involved are analyzed.

[0078] Then, based on the knowledge point source information previously recorded in the response generation model, the corresponding knowledge point source identifier is embedded into the initial response sentence. This embedding method can be to add a set identifier symbol after each key information or sentence, along with the corresponding knowledge point source description. For example, add an identifier symbol after a sentence describing a cost control strategy, such as "[Source: Document Y in Knowledge Base X]". This way, when users view the response sentence, they can intuitively understand the knowledge source of each part.

[0079] When embedding knowledge point source identifiers, pay attention to their accuracy and consistency. Ensure that each identifier corresponds to the correct knowledge point source and that all identifiers are formatted and expressed in a consistent manner. Also, avoid overly lengthy or complex identifiers that can affect the readability of your responses.

[0080] Through this interpretability optimization process, the knowledge point source identifier is accurately embedded in the initial response sentence, ultimately generating an optimized response sentence. This optimized response sentence not only provides an accurate answer but also allows users to clearly understand the basis for the answer, improving the quality and credibility of the response.

[0081] Step S150: Dynamically adjust the response service strategy of the interactive learning platform based on the difference characteristics between the optimized response statement and the historical response statements in the user interaction data.

[0082] Step S151: perform deep semantic parsing on the optimized response statement and the historical response statement respectively to generate a first semantic parsing tree and a second semantic parsing tree, and extract a first core knowledge point node set in the first semantic parsing tree and a second core knowledge point node set in the second semantic parsing tree.

[0083] Deep semantic parsing is the process of in-depth analysis of optimized and historical response statements, aiming to reveal the semantic structure and core knowledge points of the statements. For optimized response statements, a specialized semantic parser is used. The semantic parser performs part-of-speech tagging, syntactic analysis, and semantic understanding on the statements, constructing a first semantic parse tree. The first semantic parse tree is a tree structure in which nodes represent different semantic units in the statement and edges represent the semantic relationships between these units.

[0084] For example, for the optimized response sentence "In the development of the low-altitude economy, the emerging low-altitude logistics distribution model has improved timeliness through technological innovation, and cost control is also more effective [source: document B in knowledge base A]", the semantic parser will identify semantic units such as "low-altitude economic development", "emerging low-altitude logistics distribution model", "technological innovation", "timeliness", and "cost control", and construct the relationship between them to form the first semantic parsing tree.

[0085] Similarly, deep semantic parsing is performed on historical response sentences to generate a second semantic parse tree. For example, the historical response sentence "The traditional low-altitude logistics distribution model mainly relies on economies of scale to control costs" is semantically parsed to construct a second semantic parse tree.

[0086] After obtaining the first and second semantic parse trees, it is necessary to extract the core knowledge point node sets. Core knowledge point nodes refer to nodes that represent key knowledge information in the semantic parse tree. For the first semantic parse tree, by traversing and analyzing the tree, core knowledge point nodes such as "emerging low-altitude logistics distribution model", "technological innovation", "timeliness", and "cost control" are identified to form the first core knowledge point node set. For the second semantic parse tree, core knowledge point nodes such as "traditional low-altitude logistics distribution model", "scale effect", and "cost control" are extracted to form the second core knowledge point node set.

[0087] Step S152: Calculate the knowledge overlap and knowledge difference between the first core knowledge point node set and the second core knowledge point node set. The knowledge overlap is used to characterize the proportion of knowledge points commonly covered in the response statements before and after optimization. The knowledge difference is used to characterize the distribution difference between the newly added knowledge points in the response statement after optimization and the omitted knowledge points in the response statement before optimization.

[0088] The process of calculating knowledge overlap and knowledge difference is achieved by comparing and analyzing the first core knowledge point node set and the second core knowledge point node set. First, find the knowledge point nodes that are the same in the two sets. For example, the knowledge point node "cost control" exists in both the first core knowledge point node set and the second core knowledge point node set.

[0089] The knowledge overlap is calculated by comparing the number of identical knowledge point nodes with the total number of knowledge point nodes in the two sets. Specifically, let A be the number of elements in the first core knowledge point node set, B be the number of elements in the second core knowledge point node set, and C be the number of identical knowledge point nodes. The knowledge overlap can be calculated by dividing C by the larger of A and B. A higher knowledge overlap indicates a greater proportion of knowledge points shared by the response statements before and after optimization.

[0090] To calculate the knowledge difference, it is necessary to find the newly added knowledge point nodes that exist in the first core knowledge point node set but not in the second core knowledge point node set, and the missing knowledge point nodes that exist in the second core knowledge point node set but not in the first core knowledge point node set.

[0091] Next, analyze the distribution of newly added and omitted knowledge nodes. By classifying and counting knowledge nodes, we can calculate the difference in their number across different knowledge domains or topics. For example, there may be multiple newly added knowledge nodes in the area of technological innovation, while there may be a few omitted knowledge nodes in the area of traditional operating models. By comparing these differences in number, we can determine the degree of knowledge dissimilarity. This degree of knowledge dissimilarity reflects the change in knowledge coverage of the optimized response statement compared to before optimization.

[0092] Step S153: constructing a weighted difference scoring function based on the knowledge overlap and the knowledge difference to generate a semantic difference, where the semantic difference is used to quantify the difference in knowledge coverage before and after the answer optimization.

[0093] Constructing a weighted difference scoring function requires comprehensive consideration of the impact of knowledge overlap and knowledge difference. The weighting of these two factors in the scoring function needs to be adjusted based on the actual application scenario and requirements. For example, if the optimization emphasizes the newly added knowledge in the response, a higher weight can be assigned to knowledge difference; if the goal is to ensure knowledge coherence before and after optimization, a higher weight can be assigned to knowledge overlap.

[0094] Let K1 be the knowledge overlap and K2 be the knowledge difference. The corresponding weights are W1 and W2, respectively, with W1 + W2 = 1. The weighted difference scoring function can be implemented by weightedly combining K1 and K2. Specifically, the semantic difference D can be calculated using the formula D = W1 * (1 - K1) + W2 * K2. (1 - K1) represents the non-overlapping portion of knowledge, which is weighted together with the knowledge difference to produce a comprehensive score, the semantic difference.

[0095] Semantic Dissimilarity quantifies the difference in knowledge coverage before and after response optimization. A greater degree of semantic dissimilarity indicates a significant change in the knowledge content of the optimized response; a smaller degree of semantic dissimilarity indicates similar knowledge coverage before and after optimization.

[0096] Step S154: determining the response quality evaluation index of the interactive learning platform according to the semantic difference, wherein the response quality evaluation index includes knowledge accuracy increment, term coverage increment and user understanding increment.

[0097] Semantic difference reflects the difference in knowledge coverage before and after response optimization, and this can be used to determine the quality assessment indicators for interactive learning platforms. Regarding the incremental knowledge accuracy, a large semantic difference indicates that the optimized response statement has undergone significant changes in knowledge content, and it is necessary to evaluate whether these changes have improved knowledge accuracy. The accuracy of the knowledge points in the optimized response statement can be checked by comparing it with the authoritative low-altitude economic education knowledge base. If the knowledge points in the optimized response statement are more consistent with the content of the authoritative knowledge base, the incremental knowledge accuracy is positive; otherwise, it is negative.

[0098] The term coverage increment refers to the change in the coverage of low-altitude economic education terminology in the optimized response sentence compared to the pre-optimization level. This can be calculated by counting the number and types of domain terms in the response sentence before and after optimization. If the optimized response sentence covers more domain terms, the term coverage increment is positive; otherwise, it is negative.

[0099] The increase in user comprehension needs to be assessed through user feedback or simulated user comprehension tests. This can be done by inviting users to evaluate the responses before and after optimization, or by using natural language processing techniques to analyze the complexity and clarity of responses. If users find the optimized responses easier to understand, the increase in user comprehension is positive; otherwise, it is negative.

[0100] Step S155: Generate a service strategy adjustment instruction based on the response quality evaluation indicator, wherein the service strategy adjustment instruction includes a knowledge base update instruction, a semantic encoder optimization instruction, and a response generation adjustment instruction.

[0101] Based on the response quality evaluation indicators, corresponding service strategy adjustment instructions are generated. When the knowledge accuracy increment is low, it means that the knowledge in the knowledge base may be inaccurate or incomplete, and a knowledge base update instruction needs to be generated. The knowledge base update instruction will trigger the inspection and update operation of the low-altitude economic education knowledge base. Specifically, the knowledge points in the optimized response statement will be compared with the content in the knowledge base, and inaccurate or missing knowledge points will be corrected and supplemented. At the same time, the knowledge points in the knowledge base will be classified and organized to improve the quality and usability of the knowledge base.

[0102] If the term coverage increment is low, it indicates that the semantic encoder may be deficient in recognizing and processing low-altitude economic education terminology, and it is necessary to generate semantic encoder optimization instructions. These instructions guide the optimization of the pre-trained low-altitude economic domain semantic encoder. This can be achieved by adding more domain terminology data for training and adjusting the encoder's parameters and model structure to improve its ability to recognize and encode domain terminology.

[0103] If the incremental user comprehension is low, it indicates that the response generation model needs improvement in terms of readability and logic of the generated responses, and response generation adjustment instructions are required. These adjustment instructions adjust the response generation model based on the multi-level attention mechanism. These adjustments can optimize the model's output layer structure and improve the sentence generation algorithm, making the generated responses more natural, fluent, and easy to understand.

[0104] Step S156: Incrementally optimize the real-time response service process of the interactive learning platform according to the service strategy adjustment instruction. The incremental optimization includes updating the distribution weights of hot knowledge points in the low-altitude economic education knowledge base, adjusting the semantic optimization parameters of the semantic optimization model in the low-altitude economic field, and optimizing the response generation parameters of the response generation model.

[0105] Step S1561: Update the distribution weights of hot knowledge points in the low-altitude economic education knowledge base.

[0106] According to the knowledge point source identifier embedded in the optimized response statement, the core knowledge point identifier set associated with the optimized response statement is extracted. First, the knowledge point source identifier in the optimized response statement is parsed to identify the identification information of the core knowledge points contained therein. For example, for the content identified as "[Source: Document Y in Knowledge Base X, Knowledge Point Z]" in the optimized response statement, the identifier of Knowledge Point Z is extracted. All such core knowledge point identifiers are collected to form a core knowledge point identifier set.

[0107] Count the user query frequency and associated response citation frequency of the core knowledge point identifier set within a preset time window. By analyzing the log data of the interactive learning platform, count the number of times each core knowledge point identifier is queried by users within the preset time window to obtain the user query frequency. Simultaneously, count the number of times the core knowledge point identifier is cited in response statements to obtain the associated response citation frequency.

[0108] According to the normalized user query frequency and the associated answer citation frequency, a real-time popularity score for each core knowledge point is generated through a weighted fusion algorithm. Since the numerical ranges of user query frequency and associated answer citation frequency may be different, they need to be normalized so that they are on the same scale. Let the normalized user query frequency be F1, the associated answer citation frequency be F2, the corresponding weights be W3 and W4 respectively, and W3+W4=1. The real-time popularity score R can be calculated by the formula R=W3*F1+W4*F2. Among them, the weights W3 and W4 are dynamically adjusted according to the preset teaching scenario. For example, in a teaching scenario that focuses on user active queries, W3 can be assigned a higher weight; in a teaching scenario that focuses on answer citations, W4 can be assigned a higher weight.

[0109] The retrieval priority weight of the corresponding knowledge point in the low-altitude economic education knowledge base is adjusted based on the real-time popularity score. The higher the real-time popularity score of the knowledge point, the higher the feature matching weight is required in the subsequent response generation process. Therefore, the retrieval priority weight of the corresponding knowledge point in the knowledge base is adjusted according to the real-time popularity score. For example, the retrieval priority weight of the knowledge point with a high real-time popularity score is increased, so that the system can find these hot knowledge points faster when searching the knowledge base, thereby improving the accuracy and relevance of the response.

[0110] Step S1562: Adjust the semantic optimization parameters of the semantic optimization model for the low-altitude economic domain.

[0111] Extract the feature matching deviation between the semantic vector features of the user question statement and the knowledge graph embedding vector in the low-altitude economic education knowledge base. First, calculate the similarity between the semantic vector features of the user question statement and the knowledge graph embedding vector. A vector similarity calculation method, such as the cosine similarity algorithm, can be used. The calculated similarity is compared with a preset ideal similarity to obtain the feature matching deviation. The feature matching deviation reflects the degree of difference between the semantic vector features and the knowledge graph embedding vector.

[0112] A dynamic attenuation coefficient is constructed based on the domain-related features. The domain-related features quantify the strength of the association between the user's question statement and the core knowledge system of low-altitude economic education. A dynamic attenuation coefficient is constructed based on the magnitude of the domain-related features. Stronger domain-related features have smaller dynamic attenuation coefficients, while weaker domain-related features have larger dynamic attenuation coefficients.

[0113] The feature matching deviations are weighted and fused to generate a context-aware feature correction factor. The feature matching deviations are multiplied by the dynamic attenuation coefficient to obtain a context-aware feature correction factor. This feature correction factor takes into account the context of the user's question and its relevance to domain knowledge.

[0114] The feature corrections are back-propagated to the term context analysis layer of the semantic optimization model for the low-altitude economy. Through the back-propagation algorithm, the feature corrections are passed to the term context analysis layer. In this layer, the hierarchical attention weights and the span parameters of the term co-occurrence sliding window used in the knowledge graph embedding process are adjusted simultaneously. Adjusting the hierarchical attention weights allows the model to focus more on knowledge graph information relevant to the user's question, while adjusting the span parameters of the term co-occurrence sliding window better captures the contextual relationships between terms.

[0115] Based on the incremental knowledge accuracy in the response quality assessment metric, the term recognition error rate of the adjusted semantic optimization model is verified. The term recognition error rate is calculated by comparing the results of the adjusted semantic optimization model with authoritative term annotation data. When the term recognition error rate decreases and the strength of the domain-related features increases, the performance of the adjusted semantic optimization model has improved. At this point, the iterative update range of the attention weights at the current level is locked to avoid over-adjustment that may lead to unstable model performance.

[0116] Step S1563: Optimize the response generation parameters of the response generation model.

[0117] Extract the mapping deviation between the knowledge point source identifier of the optimized response statement and the hierarchical fusion priority of the domain knowledge feature set. First, encode the knowledge point source identifier of the optimized response statement to obtain an identifier embedding vector. At the same time, encode the hierarchical fusion priority of the domain knowledge feature set to obtain a hierarchical priority vector. Then, calculate the orthogonal projection distance between the identifier embedding vector and the hierarchical priority vector, and use the orthogonal projection distance as the mapping deviation. The mapping deviation reflects the degree of inconsistency between the knowledge point source identifier and the hierarchical fusion priority.

[0118] A cross-level gradient constraint function is constructed based on the semantic difference. The semantic difference quantifies the difference in knowledge coverage before and after answer optimization. Based on the semantic difference, a cross-level gradient constraint function is constructed. This function is used to limit the update amplitude of the bidirectional attention projection parameters in the multi-level attention mechanism to prevent excessive parameter updates from causing model instability.

[0119] Backpropagation optimization is performed on the bidirectional attention projection parameters in the multi-level attention mechanism. A cross-level gradient constraint function is applied to the backpropagation algorithm to adjust the bidirectional attention projection parameters. Through continuous iterative updates, the parameters gradually converge to the optimal value, adjusting the nonlinear alignment coefficient between the knowledge feature dimension and the problem feature dimension, and improving the model's ability to integrate knowledge and problem features.

[0120] The decoder parameters of the response generation model are reconstructed based on the user comprehension increment. The user comprehension increment reflects the user's understanding of the optimized response statement. The decoder parameters are adjusted according to the size of the user comprehension increment. The knowledge application scenario features corresponding to the knowledge point source identifier and the context-related features of the historical response statement can be temporally coupled to dynamically enhance the feature fusion strength of the scene adaptation layer during the decoding process. In this way, the generated response statement is more in line with the user's understanding habits, thereby improving user satisfaction.

[0121] By real-time updating of the standard knowledge point features, the dynamic interface parameters of the standard verification layer in the response generation model are synchronously adjusted, so that the technical standard compliance of the generated response statement is synchronized with the latest industry standard changes in the knowledge base.

[0122] Standard knowledge point features reflect the latest technical standards and specifications in the field of low-altitude economic education. Their real-time updates reflect the dynamic changes in industry standards. The standard validation layer in the response generation model is responsible for ensuring that the generated response statements comply with the current technical standards. To ensure that the standard validation layer can promptly adapt to changes in industry standards, its dynamic interface parameters need to be adjusted based on the real-time updates of the standard knowledge point features.

[0123] First, a monitoring mechanism is needed to monitor the real-time status of updates to standard knowledge points. This mechanism can regularly obtain updated information about standard knowledge points from the low-altitude economic education knowledge base. For example, when a knowledge point regarding low-altitude aircraft safety standards is updated in the knowledge base, the monitoring mechanism will capture this change and record the content and time of the update.

[0124] Then, based on the monitored update information, analyze the impact of these changes on the standard verification layer. Different standard knowledge point updates may affect different interface parameters. For example, if the update involves new cost accounting standards, then the interface parameters related to cost accounting in the standard verification layer need to be adjusted.

[0125] Next, specific adjustments are made to the dynamic interface parameters of the standard verification layer. This process determines the method and extent of the adjustments based on the characteristics of the updated standard knowledge points and the structure of the response generation model. For some simple updates, the preset values of the interface parameters can be directly modified. For example, if a new standard specifies a timeliness indicator range for a certain low-altitude logistics distribution model, the interface parameters corresponding to the timeliness verification in the standard verification layer can be directly updated to the new range.

[0126] For more complex updates, interface parameters may need to be retrained or optimized. For example, when industry standards undergo significant changes, involving adjustments to multiple technical specifications, the interface parameters of the standard verification layer need to be retrained using the new standard knowledge point feature data. This can be achieved by inputting the new data into the response generation model and using the backpropagation algorithm to adjust the interface parameters so that it can accurately verify whether the response statement conforms to the new standard.

[0127] When adjusting dynamic interface parameters, it's also important to consider their impact on the entire response generation model. To avoid performance degradation or instability due to parameter adjustments, thorough testing and verification are necessary after adjusting parameters. A set of test data can be used to assess whether the responses generated by the adjusted model meet the new technical standards and whether the overall model performance remains stable.

[0128] At the same time, a feedback mechanism is also needed to ensure that the technical standards for generated response statements remain consistent with the latest industry standard changes in the knowledge base. When the standard verification layer discovers that a response statement does not meet the latest standards during actual operation, this information can be promptly fed back to the monitoring mechanism and parameter adjustment module. The monitoring mechanism will recheck the update status of the standard knowledge point features, and the parameter adjustment module will further adjust the dynamic interface parameters based on this feedback, forming a closed-loop synchronous adjustment system.

[0129] Through the above steps, the dynamic interface parameters of the standard verification layer in the response generation model are continuously adjusted according to the real-time update status of the standard knowledge point features, so that the generated response statements can always meet the latest technical standards in the field of low-altitude economic education, and improve the accuracy and professionalism of the response.

[0130] Throughout the implementation of this intelligent question-answering method for low-altitude economic education on an interactive learning platform, each step is closely linked, forming a complete closed-loop system. From acquiring a user question dataset, through question feature extraction, domain knowledge retrieval, and feature fusion to generating optimized responses, and then adjusting the response service strategy based on the differences between the optimized response and historical responses, each step plays a key role in improving the performance and service quality of the question-answering system.

[0131] In terms of data collection, the platform continuously collects data sets on low-altitude economic education questions submitted by users on the interactive learning platform, ensuring that the system can obtain the latest and most comprehensive information on user needs. During the feature extraction phase, in-depth semantic analysis and feature mining are carried out on user question sentences, including the extraction of semantic vector features, domain-related features, and context-related features. This provides a rich and accurate feature foundation for subsequent knowledge retrieval and answer generation.

[0132] During the knowledge retrieval process, a multi-level knowledge matching process is used to filter out domain knowledge feature sets related to the user's question from the low-altitude economic education knowledge base, including core knowledge point features, extended knowledge point features, and standard knowledge point features. The integration and prioritization of these feature sets enables the system to more accurately locate and utilize relevant knowledge.

[0133] During the feature fusion and response generation phase, cross-dimensional feature fusion processing organically combines the question feature set with the domain knowledge feature set, generating an optimized response statement that includes low-altitude economic education terminology and knowledge point source identification. Furthermore, interpretability optimization processing is performed on the optimized response statement, improving the transparency and credibility of the response.

[0134] Finally, by comparing the differences between optimized responses and historical responses, the interactive learning platform's response service strategy is dynamically adjusted. This includes updating the distribution weights of hot knowledge points in the knowledge base, adjusting the semantic optimization parameters of the semantic optimization model, and optimizing the response generation parameters of the response generation model. These adjustments enable the system to continuously adapt to changes in industry standards and the diversity of user needs, improving response quality and service effectiveness.

[0135] Throughout the entire process, attention must also be paid to the protection of privacy-sensitive data. During the data collection phase, user-submitted question data may contain sensitive information. To prevent the leakage of this information, data encryption technology can be used to encrypt the data. For example, a symmetric encryption algorithm can be used to encrypt user question statements and historical response statements, allowing only authorized system modules to decrypt and process them. Furthermore, appropriate security measures must be implemented during data storage and transmission, such as using secure network protocols and storage devices, to ensure data security and integrity.

[0136] Furthermore, the training and optimization of AI models must adhere to certain principles and methods. Model training requires the use of a large amount of labeled data for supervised learning. This labeled data should be representative and accurate, reflecting the actual situation in the field of low-altitude economic education. During model training, appropriate training parameters, such as the learning rate and number of iterations, must be set to ensure the model converges to the optimal solution. Furthermore, the model must be evaluated and validated, using various evaluation metrics to measure its performance, such as accuracy and recall.

[0137] In terms of model optimization, it is necessary to continuously adjust the model's parameters and structure based on feedback from actual applications. For example, if a model performs poorly on certain types of problems, its performance can be improved by increasing the training data, adjusting the number of layers or neurons in the model, and so on.

[0138] In summary, by implementing the above-mentioned intelligent question-answering method for low-altitude economic education on an interactive learning platform, it is possible to provide users with more accurate, professional, and explainable answer services, while continuously optimizing the system's performance and service strategies to adapt to the development of the low-altitude economic education field and changes in user needs.

[0139] Figure 2 A schematic diagram illustrates exemplary hardware and software components of an intelligent question-answering system 100 for low-altitude economic education on an interactive learning platform, 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 intelligent question-answering system 100 for low-altitude economic education on an interactive learning platform to perform the functions described in the present application.

[0140] The intelligent question-answering system 100 for low-altitude economic education on an interactive learning platform can be a general-purpose server or a special-purpose server, both of which can be used to implement the intelligent question-answering method for low-altitude economic education on an interactive learning platform of the present 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.

[0141] For example, the low-altitude economic education intelligent question-answering system 100 for interactive learning platform can include a network port 110 connected to the 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 low-altitude economic education intelligent question-answering system 100 for interactive learning platform can also include a program instruction stored in a ROM, RAM or other types of non-transitory storage media or any combination thereof. The method of the present application can be realized according to these program instructions. The low-altitude economic education intelligent question-answering system 100 for interactive learning platform also includes an I / O interface 150 between a computer and other input and output devices.

[0142] For ease of explanation, only one processor is described in the low-altitude economic education intelligent question and answer system 100 for an interactive learning platform. However, it should be noted that the low-altitude economic education intelligent question and answer system 100 for an interactive learning platform in this application may also include multiple processors, so the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the low-altitude economic education intelligent question and answer system 100 for an interactive learning platform executes steps A and B, it should be understood that steps A and B may also be performed 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.

[0143] 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-answering method for low-altitude economic education on an interactive learning platform is implemented.

[0144] 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 low-altitude economic education on an interactive learning platform, characterized in that: The method comprises: Obtaining a low-altitude economic education question data set submitted by a user on an interactive learning platform, wherein the low-altitude economic education question data set includes a plurality of user interaction data, each user interaction data consisting of at least one user question statement and a corresponding historical answer statement; Extracting a question feature set from the user interaction data, the question feature set including a semantic vector feature of the user question statement, a field association feature between the user question statement and the field of low-altitude economic education, and a context association feature of the user question statement; Retrieving a domain knowledge feature set in a low-altitude economic education knowledge base based on the problem feature set; Performing cross-dimensional feature fusion processing on the question feature set and the domain knowledge feature set to generate an optimized response statement for the user question statement; Dynamically adjusting the response service strategy of the interactive learning platform based on the difference characteristics between the optimized response statement and the historical response statements in the user interaction data; The dynamically adjusting the response service strategy of the interactive learning platform based on the difference characteristics between the optimized response statement and the historical response statements in the user interaction data includes: Performing deep semantic parsing on the optimized response statement and the historical response statement, respectively, to generate a first semantic parse tree and a second semantic parse tree, and extracting a first core knowledge point node set in the first semantic parse tree and a second core knowledge point node set in the second semantic parse tree; Calculate the knowledge overlap and knowledge difference between the first core knowledge point node set and the second core knowledge point node set, where the knowledge overlap is used to characterize the proportion of knowledge points commonly covered in the response statements before and after optimization, and the knowledge difference is used to characterize the distribution difference between the newly added knowledge points in the response statement after optimization and the omitted knowledge points in the response statement before optimization; Constructing a weighted difference scoring function based on the knowledge overlap and the knowledge difference to generate a semantic difference, wherein the semantic difference is used to quantify the difference in knowledge coverage before and after answer optimization; Determining the response quality evaluation index of the interactive learning platform according to the semantic difference, wherein the response quality evaluation index includes knowledge accuracy increment, term coverage increment and user understanding increment; Generate a service strategy adjustment instruction based on the response quality evaluation indicator, wherein the service strategy adjustment instruction includes a knowledge base update instruction, a semantic encoder optimization instruction, and a response generation adjustment instruction; The real-time response service process of the interactive learning platform is incrementally optimized according to the service strategy adjustment instruction. The incremental optimization includes updating the distribution weights of hot knowledge points in the low-altitude economic education knowledge base, adjusting the semantic optimization parameters of the semantic optimization model in the low-altitude economic field, and optimizing the response generation parameters of the response generation model.

2. The intelligent question-answering method for low-altitude economic education on an interactive learning platform according to claim 1 is characterized in that: The extracting of the problem feature set of the user interaction data includes: Performing semantic encoding processing on the user question sentence to generate a semantic vector feature of the user question sentence, wherein the semantic vector feature includes a keyword distribution weight of a core concept in the user question sentence; Extracting response semantic features of historical response sentences in the user interaction data, performing time-series alignment processing on the response semantic features and the semantic vector features, and generating context-related features of the user question sentence; Determining, based on the term coverage of the user question statement in the field of low-altitude economic education, a domain-related feature of the user question statement, wherein the domain-related feature is used to quantify the strength of the association between the user question statement and the core knowledge system of low-altitude economic education; The semantic vector features, the context-related features, and the domain-related features are concatenated to generate the problem feature set.

3. The intelligent question-answering method for low-altitude economic education on an interactive learning platform according to claim 2 is characterized in that: The performing semantic encoding processing on the user question statement to generate a semantic vector feature of the user question statement includes: Calling a pre-trained low-altitude economic domain semantic encoder to perform word segmentation processing on the user question sentence to obtain multiple semantic word units; Performing context dependency analysis on the semantic word units to generate a local context weight for each semantic word unit, wherein the local context weight is used to represent the semantic contribution of the semantic word unit in the user question sentence; Performing weighted aggregation processing on the semantic word units according to the local context weights to generate an initial semantic vector of the user question sentence; The initial semantic vector is input into the semantic optimization model of the low-altitude economy field. Through knowledge graph embedding and term context association analysis, the initial semantic vector is enhanced with domain knowledge to generate semantic vector features containing the distribution characteristics of low-altitude economy professional terms.

4. The intelligent question-answering method for low-altitude economic education on an interactive learning platform according to claim 1 is characterized in that: The step of searching the domain knowledge feature set in the low-altitude economic education knowledge base based on the problem feature set includes: Perform multi-level knowledge matching in the low-altitude economic education knowledge base according to the problem feature set, wherein the multi-level knowledge matching includes a core concept matching layer, an application scenario matching layer, and a technical standard matching layer; In the core concept matching layer, based on the keyword distribution weights in the semantic vector features, the core knowledge point features directly associated with the user question statement in the low-altitude economic education knowledge base are retrieved; In the application scenario matching layer, based on the contextual association features of the user question sentence, extended knowledge point features associated with the implicit application scenario of the user question sentence are retrieved from the low-altitude economic education knowledge base; In the technical standard matching layer, based on the domain-related features, standard knowledge point features in the low-altitude economic education knowledge base that match the technical standard category to which the user's question statement belongs are screened; The core knowledge point features, the extended knowledge point features and the standard knowledge point features are hierarchically fused according to preset priorities to generate the domain knowledge feature set, wherein the core knowledge point features retain the highest priority by default in subsequent dynamic weight allocation.

5. The intelligent question-answering method for low-altitude economic education on an interactive learning platform according to claim 1 is characterized in that: The performing cross-dimensional feature fusion processing on the question feature set and the domain knowledge feature set to generate an optimized response statement for the user question statement includes: Performing knowledge adaptability analysis on the semantic vector features in the problem feature set to determine the semantic matching degree between the semantic vector features and each knowledge point in the domain knowledge feature set; Building a dynamic weight allocation mechanism based on the semantic matching degree, prioritizing the knowledge points in the domain knowledge feature set, and obtaining a priority ranking result; According to the priority sorting result, feature dimension projection processing is performed on the priority knowledge point features of the top N priority points and the context-related features in the question feature set to generate projected knowledge features and projected question features of the same dimension, and the projected knowledge features and projected question features are fused based on a bidirectional attention mechanism to generate knowledge-enhanced semantic features; Inputting the knowledge-enhanced semantic features into a response generation model based on a multi-level attention mechanism to generate an initial response sentence containing low-altitude economic education field terms and knowledge point source identifiers; The initial response statement is subjected to interpretability optimization processing, and a knowledge point source identifier corresponding to the user question statement is embedded in the initial response statement to generate the optimized response statement.

6. The intelligent question-answering method for low-altitude economic education on an interactive learning platform according to claim 1 is characterized in that: The updating of the distribution weights of hot knowledge points in the low-altitude economic education knowledge base includes: Extracting a set of core knowledge point identifiers associated with the optimized response statement based on the knowledge point source identifier embedded in the optimized response statement; Counting the user query frequency and the associated answer citation frequency of the core knowledge point identifier set within a preset time window; Based on the normalized user query frequency and the associated answer citation frequency, a weighted fusion algorithm is used to generate a real-time popularity score for each core knowledge point. The weights of user query frequency and associated answer citation frequency are dynamically adjusted according to the preset teaching scenario. The retrieval priority weight of the corresponding knowledge point in the low-altitude economic education knowledge base is adjusted based on the real-time popularity score, so that the high-frequency hot knowledge points obtain a higher feature matching weight in the subsequent answer generation process.

7. The intelligent question-answering method for low-altitude economic education on an interactive learning platform according to claim 1 is characterized in that: The adjustment of the semantic optimization parameters of the semantic optimization model in the low-altitude economic field includes: Extracting the feature matching deviation between the semantic vector feature of the user question statement and the knowledge graph embedding vector in the low-altitude economic education knowledge base, wherein the feature matching deviation is calculated by normalized vector cosine similarity; Constructing a dynamic attenuation coefficient based on the domain-related features, performing weighted fusion on the feature matching deviations, and generating a context-aware feature correction amount; Back-propagating the feature correction amount to the term context association analysis layer of the semantic optimization model for the low-altitude economic field, and synchronously adjusting the hierarchical attention weights and the span parameters of the term co-occurrence sliding window in the knowledge graph embedding process; Based on the knowledge accuracy increment in the answer quality evaluation index, the term recognition error rate of the adjusted semantic optimization model is verified. When the term recognition error rate decreases and the strength of the domain-related features increases, the iterative update amplitude of the current level attention weight is locked.

8. The intelligent question-answering method for low-altitude economic education on an interactive learning platform according to claim 1 is characterized in that: The optimized response generation parameters of the response generation model include: Extracting a mapping deviation between the knowledge point source identifier of the optimized response sentence and the hierarchical fusion priority of the domain knowledge feature set, wherein the mapping deviation is calculated by the orthogonal projection distance between the identifier embedding vector and the hierarchical priority vector; A cross-level gradient constraint function is constructed based on the semantic difference, and back-propagation optimization is performed on the bidirectional attention projection parameters in the multi-level attention mechanism to adjust the nonlinear alignment coefficient between the knowledge feature dimension and the problem feature dimension. Reconstructing the decoder parameters of the response generation model based on the user comprehension increment, temporally coupling the knowledge application scenario features corresponding to the knowledge point source identifier with the contextual association features of the historical response sentences, and dynamically enhancing the feature fusion strength of the scene adaptation layer during the decoding process; By real-time updating of the standard knowledge point features, the dynamic interface parameters of the standard verification layer in the response generation model are synchronously adjusted, so that the technical standard compliance of the generated response statement is synchronized with the latest industry standard changes in the knowledge base.

9. An intelligent question-answering system for low-altitude economic education on an interactive learning platform, characterized in that: 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 low-altitude economic education intelligent question-answering method for the interactive learning platform as described in any one of claims 1 to 8.

Images