Intelligent service method and system based on deep reasoning and knowledge enhancement

Abstract

The application is suitable for the technical field of intelligent question and answer, and particularly relates to an intelligent service method and system based on deep reasoning and knowledge enhancement, which comprises the following steps: obtaining the identity information and target question information of a student who asks questions; constructing question and answer characteristic parameters by using the target question information, improving the depth and accuracy of question understanding; screening candidate question and answer units from a preset question and answer knowledge base based on the question and answer characteristic parameters, quickly positioning a historical similar question and a corresponding answer, obtaining an associated document segment, and providing a rich and authoritative knowledge source for answer generation; when generating a candidate answer scheme based on the candidate question and answer units, expanding and optimizing the content with the candidate question and answer units as the core carrier; screening the candidate answer scheme by using a pre-trained deep reasoning screening model according to the identity information of the student, dynamically adjusting the professionalism and expression style of the answer, and displaying a target answer scheme, so as to realize the pertinence and authority from question understanding, knowledge retrieval to answer generation.

Description

Technical Field

[0001] This application belongs to the field of intelligent question answering technology, and in particular relates to an intelligent service method and system based on deep reasoning and knowledge enhancement. Background Technology

[0002] With the deep application of artificial intelligence technology in education, office work, and other fields, intelligent question-answering services based on large models have become a key technology for improving service efficiency and accuracy. Currently, most mainstream intelligent question-answering services rely on preset rules or simple keyword matching to retrieve answers from a knowledge base when processing user questions. For example, when a user asks a question, the system simply analyzes the keywords in the question and searches for similar questions and corresponding answers in the question-answering knowledge base.

[0003] However, existing intelligent question-answering services lack in-depth reasoning based on users' underlying intentions, knowledge background, and behavioral patterns. When faced with semantically ambiguous questions involving the integration of knowledge from multiple domains, they often fail to understand the essential meaning of the question, returning answers with low relevance or even errors. For example, in a campus Q&A scenario, when a student asks "What are the specific regulations regarding failing a course at a certain college?", existing question-answering models only match answers to superficial questions such as "What is failing a course?" These answers lack in-depth analysis and sufficient coverage of professional knowledge in specific domains, resulting in responses that lack relevance and authority. Summary of the Invention

[0004] This application provides an intelligent service method and system based on deep reasoning and knowledge enhancement, which can solve the problem that the answers provided by existing question-answering systems lack specificity and authority.

[0005] In a first aspect, embodiments of this application provide an intelligent service method based on deep reasoning and knowledge enhancement, comprising:

[0006] Obtain the identity information of the student asking the question and the target question information;

[0007] Construct question-and-answer feature parameters corresponding to the student asking the question based on the target question information;

[0008] Based on the question-and-answer feature parameters, at least one candidate question-and-answer unit is selected from a preset question-and-answer knowledge base; wherein, the candidate question-and-answer unit includes historical similar questions and corresponding answers, and related document fragments;

[0009] At least one candidate answer scheme is generated based on the candidate question-and-answer unit; wherein, any candidate answer scheme uses any candidate question-and-answer unit as its core content carrier;

[0010] Based on the identity information, a preset deep reasoning filtering model is used to filter the candidate answer schemes, determine the target answer scheme, and display the target answer scheme to the student who asked the question; wherein, the deep reasoning filtering model is a pre-trained machine learning model.

[0011] The technical solutions described in this application embodiment have at least the following technical effects:

[0012] The intelligent service method based on deep reasoning and knowledge enhancement provided in this application accurately targets user needs by acquiring the identity information of the student asking the question and the target question information. It constructs question-and-answer feature parameters using the target question information to improve the depth and accuracy of question understanding. Based on the question-and-answer feature parameters, it filters candidate question-and-answer units from a pre-set question-and-answer knowledge base, quickly locates similar historical questions and corresponding answers, and obtains related document fragments, providing rich and authoritative knowledge sources for answer generation. When generating candidate answer schemes based on candidate question-and-answer units, it expands and optimizes the content using these units as the core carrier, ensuring the answer content has a solid foundation. Based on the student's identity information, it uses a pre-trained deep reasoning filtering model to filter candidate answer schemes, dynamically adjusting the professionalism, difficulty, and expression style of the answers and displaying the target answer scheme, achieving relevance and authority from question understanding and knowledge retrieval to answer generation.

[0013] Secondly, embodiments of this application provide an intelligent service system based on deep reasoning and knowledge enhancement, comprising:

[0014] The acquisition unit is used to acquire the identity information of the student asking the question and the target question information;

[0015] The parameter unit is used to construct the question-and-answer feature parameters corresponding to the questioning student based on the target question information;

[0016] The first filtering unit is used to filter at least one candidate question-answer unit from a preset question-answer knowledge base based on the question-answer feature parameters; wherein, the candidate question-answer unit includes historical similar questions and corresponding answers, and related document fragments;

[0017] A generation unit is used to generate at least one candidate answer scheme based on the candidate question-and-answer unit; wherein any candidate answer scheme takes any candidate question-and-answer unit as its core content carrier;

[0018] The second filtering unit is used to filter the candidate answer schemes using a preset deep reasoning filtering model based on the identity information, determine the target answer scheme, and display the target answer scheme to the student asking the question; wherein, the deep reasoning filtering model is a pre-trained machine learning model.

[0019] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any of the foregoing aspects.

[0020] Fourthly, embodiments of this application provide a computer program product that, when run on an electronic device, causes the electronic device to perform the method described in any of the above aspects.

[0021] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the above aspects, and will not be repeated here. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating an embodiment of the intelligent service method based on deep reasoning and knowledge enhancement provided in this application.

[0024] Figure 2 This is a schematic diagram of the structure of an intelligent service system based on deep reasoning and knowledge enhancement provided in an embodiment of this application;

[0025] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0027] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0028] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determination" or "if the described condition or event is detected" may be interpreted, depending on the context, as "once determination," "in response to determination," "once the described condition or event is detected," or "in response to the detection of the described condition or event."

[0030] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0031] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0032] Existing intelligent question-answering services often lack in-depth reasoning based on users' underlying intentions, knowledge background, and behavioral patterns. When faced with semantically ambiguous questions involving the integration of knowledge from multiple domains, they frequently fail to grasp the essential meaning of the question, returning answers with low relevance or even errors. For example, in a campus Q&A scenario, when a student asks "What are the specific regulations regarding failing a course at a certain college?", existing question-answering models only match answers to superficial questions such as "What does failing a course mean?" These answers lack in-depth analysis and sufficient coverage of specific domain expertise, resulting in responses that are neither targeted nor authoritative.

[0033] To address the aforementioned issues, this application provides an intelligent service method and system based on deep reasoning and knowledge enhancement. In this method, user needs are accurately identified by acquiring the student's identity information and the target question information. Question-answering feature parameters are constructed using the target question information to improve the depth and accuracy of question understanding. Based on these feature parameters, candidate question-answering units are selected from a pre-set question-answering knowledge base to quickly locate similar historical questions and corresponding answers, and obtain related document fragments, providing rich and authoritative knowledge sources for answer generation. When generating candidate answer schemes based on candidate question-answering units, content expansion and optimization are performed using these units as the core carrier to ensure the answer content has a solid foundation. Based on the student's identity information, a pre-trained deep reasoning screening model is used to screen candidate answer schemes, dynamically adjusting the professionalism, difficulty, and expression style of the answers, and displaying the target answer scheme, achieving relevance and authority from question understanding and knowledge retrieval to answer generation.

[0034] The intelligent service method based on deep reasoning and knowledge enhancement provided in this application can be applied to electronic devices. In this case, the electronic device is the executing entity of the intelligent service method based on deep reasoning and knowledge enhancement provided in this application. This application does not impose any restrictions on the specific type of electronic device.

[0035] For example, electronic devices can be mobile phones, tablets, wearable devices, in-vehicle devices, augmented reality (AR) / virtual reality (VR) devices, laptops, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), desktop computers, smart screens, smart TVs, and other terminal devices.

[0036] To better understand the intelligent service method based on deep reasoning and knowledge enhancement provided in the embodiments of this application, the specific implementation process of the intelligent service method based on deep reasoning and knowledge enhancement provided in the embodiments of this application will be described by way of example below.

[0037] Figure 1 A schematic flowchart of an intelligent service method based on deep reasoning and knowledge enhancement provided in an embodiment of this application is shown. The intelligent service method based on deep reasoning and knowledge enhancement includes:

[0038] S100: Obtain the identity information of the student asking the question and the target question information.

[0039] It is understandable that the student's identity information is a basic identifier, which may include implicit characteristics such as the student's historical learning records and knowledge mastery level. This identity information can be obtained through user input, system account authentication, and integration with the campus information platform. For example, when a student logs into the learning system, their account-linked identity data is automatically synchronized, or they submit personal information through a form. The target question information refers to the specific question content raised by the student, including the question text, text converted from speech input, and accompanying images or documents. When obtaining the target question information, the question text can be directly collected through a text input box. Speech questions can be automatically converted to text using ASR, and questions in images can be extracted using Optical Character Recognition (OCR). Simultaneously, the obtained raw information undergoes format standardization processing, such as removing special characters and standardizing text encoding, to ensure the integrity and processability of the information, providing accurate basic data for subsequent question-and-answer processing.

[0040] In some embodiments, before obtaining the identity information of the student asking the question and the target question information, the method further includes:

[0041] S600: Obtain documents related to the question-and-answer content, and construct a question-and-answer knowledge base based on these documents; wherein, the question-and-answer knowledge base includes a first knowledge repository constructed based on domain tags, and a second knowledge repository constructed based on semantic feature vectors.

[0042] It's understandable that building a question-and-answer knowledge base is a process of structuring unstructured documents. Relevant documents such as textbooks, papers, and online Q&A platforms can be obtained through web crawling, document import, and manual collection. Seed question-and-answer units with labeled domain tags can then be extracted. These seed units refer to the most basic and core question-and-answer combinations that constitute the question-and-answer system. For example, regular expressions can be used to match the format "question: ... answer: ...", or information extraction models (such as BERT sequence labeling) can be used to identify question-and-answer boundaries. Semantic analysis is performed on the seed units, and feature vectors are generated using models such as SBERT. A semantic index database (such as Milvus, which stores the mapping between vectors and question-and-answer units) is then constructed. For unlabeled question-and-answer units, after generating feature vectors, their similarity is calculated with vectors in the index database. The most similar seed unit is found using the KNN algorithm (K=5), and the domain tag with the majority vote is used for automatic labeling. Finally, the seed question-and-answer units are clustered by domain tag to form the first knowledge warehouse, and stored by semantic feature vector to form the second knowledge warehouse, constructing a composite knowledge base that supports domain classification and semantic retrieval.

[0043] The dual-warehouse design, which uses domain-based tags to construct the first knowledge warehouse and semantic feature vectors to construct the second knowledge warehouse, effectively prevents the lack of retrieval dimensions and the mismatch of complex questions in a single knowledge management model (eliminating the mismatch of cross-domain questions with irrelevant knowledge in a single domain that affects authority). By limiting the domain scope through the structured classification of the tag warehouse and supplementing semantic associations through vector similarity of the semantic warehouse, a two-way verification architecture of "classification to prevent out-of-bounds and semantic to prevent omission" is formed. This ensures the comprehensiveness of knowledge retrieval, the stability of updates, and the accuracy of matching, providing a robust knowledge management solution for intelligent question answering that combines rule determinism and data flexibility.

[0044] In one possible implementation, S600 retrieves documents related to the question-and-answer content and constructs a question-and-answer knowledge base based on these documents, including:

[0045] S610: Obtain documents related to the question and answer content, and extract seed question and answer units with labeled domain tags from the documents related to the question and answer content.

[0046] It is understandable that when obtaining documents related to Q&A content, one can collect internal regulations documents such as the "Student Handbook," "Teaching Management Regulations," and "Laboratory Safety Standards," as well as structured Q&A data such as counselor Q&A records and FAQs from the academic affairs system, through channels such as the school's OA system, the academic affairs office platform, and the college's official website. Seed Q&A units are basic Q&A data units that possess authority, typicality, and the ability to generate related Q&As. For example, from the "XX University Student Disciplinary Action Measures" document, one can use regular expressions to match typical Q&A formats such as "Q: How is cheating on an exam handled? A: According to Article X of the school regulations, a demerit will be given and the degree application qualification will be revoked," or use syntactic analysis to identify the Q&A boundaries of "Q: What are the penalties for returning to the dormitory late? A: Returning to the dormitory after 23:00 is considered late, and more than 3 times will result in a public reprimand." Domain tagging can be combined with the hierarchical structure of the school's regulations system. During manual annotation, domains can be tagged according to document chapter affiliation, such as tagging the Q&A section of the "Student Status Management Regulations" as "Internal Regulations - Teaching Management - Student Status Management". During semi-supervised learning annotation, tags can be automatically inherited by matching document title keywords to a pre-defined tag dictionary (e.g., if the title contains "dormitory", it can be associated with the tag "Internal Regulations - Life Management - Dormitory Management"), or by utilizing an internal regulations classification tree (e.g., "Internal Regulations → Teaching Management → Course Assessment → Cheating Handling"). Extracted seed Q&A units must be associated with specific clauses of internal regulations, such as "Q: How to apply for course retake? A: Applications must be submitted through the academic affairs system before the 4th week of each semester. See Article 5 of the 'XX University Course Retake Management Regulations' tag: Internal Regulations - Teaching Management - Course Retake," ensuring that the Q&A content is traceable to specific school regulations, providing an authoritative basis for subsequent student inquiries.

[0047] S620, perform semantic analysis on each seed Q&A unit to generate a feature vector corresponding to the seed Q&A unit.

[0048] It can be understood that semantic analysis is a process of converting text into numerical feature vectors. Perform preprocessing such as word segmentation, stop word removal (such as "de", "le"), and word form reduction (such as "running" → "run") on the question and answer texts of the seed Q&A unit to ensure the consistency of lexical representation. The feature vectors generated by semantic analysis can represent the semantic essence of the Q&A unit and provide a numerical basis for subsequent semantic retrieval and similarity calculation.

[0049] S630, construct a semantic index database based on the feature vectors, and use the semantic index database to automatically assign domain labels to Q&A units without labeled domain labels.

[0050] It can be understood that when constructing a semantic index database based on feature vectors, the feature vectors of seed Q&A units such as school regulations and school activities (such as 768-dimensional vectors generated by BERT) can be stored in vector databases such as Milvus or Faiss, establishing a mapping relationship of "feature vector - Q&A unit - labeled domain label" to form an index structure that supports efficient semantic retrieval. When processing Q&A units without labeled domain labels, first generate their target feature vectors through the same semantic analysis process, and then retrieve K candidate seed Q&A units (K = 5) in the index database whose cosine similarity with this vector exceeds a threshold (such as 0.7). Calculate the text semantic consistency between the domain labels of the candidate Q&A units and the current Q&A unit (such as calculating the question text similarity using Sentence-BERT), and select the domain label of the seed Q&A unit with the highest matching degree (such as "School Regulations - Teaching Management - Scholarship Management") and assign it to the Q&A unit without labeling, realizing the automatic domain classification of school regulation Q&A and ensuring that new Q&A can be accurately mapped to the corresponding category of the school regulation system.

[0051] Optionally, in step S630, using the semantic index database to automatically assign domain labels to Q&A units without labeled domain labels includes:

[0052] S631, perform semantic analysis on the Q&A unit without labeled domain label to generate a target feature vector.

[0053] It is understandable that for unlabeled Q&A units on school regulations (such as "Q: What procedures are required to resume classes after a leave of absence?"), text preprocessing can be performed first to remove non-critical information such as "handle" and "already", standardize terms such as "leave of absence and resumption of classes" and "procedures", and then convert the Q&A text into a 768-dimensional target feature vector by extracting semantic features. The target feature vector contains semantic representations of keywords such as "leave of absence", "resumption of classes" and "procedures" as well as implicit relationships such as "process" and "school regulations clauses", providing a numerical basis for subsequent vector similarity calculation with seed Q&A units in the index.

[0054] S632, calculate the similarity between the target feature vector and each feature vector in the semantic index database, and select candidate seed question-answering units whose similarity exceeds the threshold.

[0055] Understandably, the ANN retrieval function of the Milvus vector database can be used to calculate the cosine similarity between the target feature vector of the unlabeled question-answer unit and the feature vectors of all seed question-answer units in the index. A threshold (such as 0.7) can be set to filter out candidate seed units with similarity higher than the threshold. For example, five seed question-answer units with similarity > 0.7 to the question vector of "procedures for taking a leave of absence and returning to school" can be retrieved from the school regulations index, including "Q: How do I apply for a leave of absence?" and "Q: What materials do I need to submit for returning to school?", to ensure that the candidate question-answer units are highly relevant to the current question at the semantic level.

[0056] S633, perform semantic consistency verification on the candidate seed question-answering units, and determine the target seed question-answering unit with the highest matching degree.

[0057] It is understandable that, in addition to vector similarity, the semantic consistency of the question text is further calculated using text matching algorithms (such as edit distance and Jaccard coefficient) and semantic models (such as Sentence-BERT) for the selected candidate seed question-answer units. For example, the unlabeled question "What procedures are required to resume classes after a leave of absence?" has a text similarity of 0.85 with the candidate question-answer seed unit "Question: What is the process for students who have taken a leave of absence to resume classes?", which is higher than other candidate question-answer units. Both of them involve core semantics such as "leave of absence to resume classes" and "procedures". Therefore, this candidate question-answer unit is determined to be the target seed unit with the highest matching degree, avoiding labeling errors caused by high vector similarity but actual semantic ambiguity (such as the vectors of "leave of absence" and "withdrawal" being similar but in different domains).

[0058] S634, assign the domain label of the target seed question-answering unit to the question-answering unit.

[0059] It is understandable that the domain label of the target seed unit with the highest matching degree (such as "campus regulations - life management - student status change") is directly assigned to the unlabeled question and answer unit to complete the domain classification. For example, the question "procedures for suspension of study and resumption of classes" is assigned the same label as the target seed unit. At the same time, the source of the school regulations corresponding to the label is recorded (such as Chapter 7 of "Regulations on Student Status Management of XX University") to ensure that the labeling results can be traced back to the specific category of the school regulations system, so as to provide accurate classification for subsequent question and answer matching based on domain labels.

[0060] S640 clusters all seed question-and-answer units based on domain labels to form the first knowledge repository.

[0061] Understandably, based on the hierarchical system of school regulations (such as "school regulations → teaching management → course assessment → cheating handling"), all seed question-and-answer units are hierarchically clustered according to domain labels. Using the K-means algorithm or a label tree-based clustering method, question-and-answer units with the same or similar domain labels are grouped into the same category. For example, questions and answers such as "scholarship application" and "GPA calculation" are grouped into the category of "school regulations - teaching management - academic awards". This forms a hierarchical first knowledge repository, with each category corresponding to a sub-domain of school regulations. This supports quick retrieval of relevant question-and-answer units through domain labels, improving the retrieval efficiency and organization of the knowledge base.

[0062] S200: Construct question-and-answer feature parameters corresponding to the student asking the question based on the target question information.

[0063] It is understandable that constructing question-answering feature parameters is the process of transforming question text into structured semantic features through natural language processing techniques. This involves segmenting the target question information, tagging parts of speech, and performing dependency parsing to extract keywords (such as "deep learning," "model," and "training" in "deep learning model training") and entity relationships, forming a semantic knowledge subgraph (nodes representing entities and edges representing relationships). The question is then encoded to generate a high-dimensional feature vector (e.g., 768-dimensional) containing contextual semantics. This high-dimensional feature vector captures the implicit semantics of the question (e.g., the association between "optimizing training efficiency" and "learning rate adjustment strategy"). Domain labels (e.g., "computer science - machine learning") can be determined by keyword matching with a domain dictionary or text classification model (e.g., TextCNN), and semantic index fragments (e.g., combinations of core terms in the question) can be extracted. The generated question-answering feature parameters include domain labels, semantic feature vectors, and knowledge subgraph structures, comprehensively representing the semantic essence of the question and providing accurate matching criteria for knowledge base retrieval.

[0064] In one possible implementation, S200, the question-and-answer feature parameters corresponding to the student asking the question are constructed based on the target question information, including:

[0065] S210, Perform semantic parsing based on the target query information to determine the semantic knowledge subgraph corresponding to the target query information and the semantic index fragment of the semantic knowledge subgraph.

[0066] It is understandable that when performing semantic parsing on the target query information, key entities (such as "cheating on exams" and "guaranteed admission to graduate school") and predicate relations (such as "cancel" and "impact") can be identified through word segmentation and part-of-speech tagging. A semantic knowledge subgraph can then be constructed—with entities as nodes and relations as edges. For example, the node "cheating on exams" is connected to the node "canceled admission to graduate school" via the edge "cause". Simultaneously, semantic index fragments are extracted, namely the core term combinations in the question (such as "cheating on exams" and "canceled admission to graduate school"). These core term combination fragments serve both as keywords for retrieving the knowledge base and as subsequent semantic structure features, ensuring that the knowledge subgraph can intuitively reflect the semantic network and core retrieval elements of the question.

[0067] S220, construct the semantic adjacency tensor corresponding to the semantic knowledge subgraph based on the semantic index fragment.

[0068] It is understandable that when converting a semantic knowledge subgraph into a semantic adjacency tensor, a multidimensional tensor representation of the association strength between nodes is constructed using entities in the semantic index fragment as dimensions. For example, regarding the question "Does cheating on an exam affect postgraduate admission?", the corresponding position value of the nodes "cheating on an exam" and "postgraduate admission qualification" in the tensor is 0.9 (based on co-occurrence frequency and syntactic relation weights). The semantic adjacency tensor reflects the topological structure of the knowledge subgraph, and the numerical value represents the tightness of semantic associations. This representation can efficiently store semantic relationships and provides a mathematical basis for subsequent topological sorting.

[0069] S230, perform topological sorting on the semantic adjacency tensor to generate a semantic subgraph dependency tree; wherein, the semantic subgraph dependency tree explicitly expresses the hierarchical dependency relationship between semantic knowledge subgraphs in a tree structure.

[0070] It's understandable that when performing topological sorting on a semantic adjacency tensor, the logical order is determined based on the association weights between nodes, generating a tree-like semantic subgraph dependency tree. For example, "cheating on an exam" is the root node, "disciplinary action" and "graduate school admission policy" are first-level child nodes, and "disqualification" is a second-level child node. The edge weights represent the dependency strength (e.g., "cheating on an exam → disciplinary action" has a weight of 0.8). The semantic subgraph dependency tree clearly displays the hierarchical relationships (such as causality and subordination) of the semantic elements in the question, transforming the abstract semantic network into an intuitive logical hierarchy, facilitating subsequent feature extraction and structured answer generation.

[0071] S240, perform hierarchical decomposition of the semantic subgraph dependency tree to determine the question-and-answer feature parameters corresponding to the student asking the question.

[0072] It is understandable that when performing hierarchical decomposition on the semantic subgraph dependency tree, core features are extracted according to the tree's hierarchical structure (root node → child node → leaf node): the root node corresponds to the domain label of the question (such as "school regulations - teaching management - examination discipline"), the child nodes correspond to the subdivided knowledge points (such as "cheating punishment" and "graduate school admission qualification conditions"), and the edge weights correspond to the feature importance. Simultaneously, combined with semantic feature vectors generated by pre-trained language models (such as BERT), the final question-answering feature parameters include the domain label, semantic feature vectors, and hierarchical structure weights. For example, after decomposition, the domain label "school regulations," semantic vectors (encoding "exam cheating," "graduate school admission," etc.), and hierarchical features (weight vectors for "cheating punishment → impact on graduate school admission") are obtained, comprehensively depicting the semantic essence and logical hierarchy of the question. This transforms the semantics from a text string into a computable semantic link through the semantic subgraph dependency tree, which is particularly suitable for domains such as school regulations and laws that require precise, logical, and authoritative matching.

[0073] S300, at least one candidate question-answer unit is selected from the preset question-answer knowledge base based on question-answer feature parameters; wherein, the candidate question-answer unit includes historical similar questions and corresponding answers, and related document fragments.

[0074] It is understandable that question-and-answer feature parameters include domain labels and semantic feature vectors. Based on the domain labels, a set of question-and-answer units in the same domain can be quickly located in the first knowledge warehouse through a tree index (such as all FAQs under the category of "school regulations - teaching management - examination discipline"). Then, the semantic feature vectors are used to retrieve semantically similar question-and-answer units in the vector database of the second knowledge warehouse (the semantic association between "whether switching screens during online exams counts as cheating" and the historical question "Zoom exam window switching judgment rules" is matched by calculating cosine similarity). Then, the matching coefficient of each question-and-answer unit in the two sets is calculated separately, namely the first matching coefficient and the second matching coefficient. Finally, through cross-fusion, historical similar questions and answers and related document fragments that are related to the domain and semantically similar are selected by sorting by score to form a set of candidate question-and-answer units, providing highly relevant knowledge materials for subsequent generation of answer solutions.

[0075] In one possible implementation, the question-answering feature parameters include domain labels and semantic feature vectors; S300, based on the question-answering feature parameters, at least one candidate question-answering unit is selected from a pre-defined question-answering knowledge base, including:

[0076] S310: Match the corresponding first question-answering unit set from the first knowledge warehouse based on the domain label, and retrieve the corresponding second question-answering unit set from the second knowledge warehouse based on the semantic feature vector.

[0077] It is understandable that candidate question-and-answer units can be simultaneously retrieved from the first and second knowledge repositories based on the domain labels and semantic feature vectors in the question-and-answer feature parameters. The first knowledge repository retrieval can use the domain label as the index key and directly locate the corresponding set of question-and-answer units through a hierarchical tree structure (e.g., "School Regulations → Teaching Management → Scholarship Management"). For example, when the domain label of the question is "scholarship application requirements," all FAQs related to the "Scholarship Management Regulations" can be quickly extracted from the first knowledge repository, ensuring that the search content is strictly aligned with the question's theme. The second knowledge repository (semantic vector repository) retrieval can input the semantic feature vector of the question (e.g., a 768-dimensional vector generated by BERT) into the vector database, and retrieve the historical question-and-answer units with the closest semantics through cosine similarity calculation. For example, the semantic vector of the question "Can a GPA of 3.2 apply for a national scholarship?" can match historical questions and answers such as "A GPA of 3.5 or higher is eligible to apply," achieving generalized retrieval through semantic association even if the keywords "GPA" and "grade point" are expressed differently.

[0078] S320, determine the first matching coefficient list of the first question-and-answer unit set and the second matching coefficient list of the second question-and-answer unit set respectively.

[0079] It is understandable that when determining the first matching coefficient list, for each unit in the first question-and-answer unit set, the coefficient is calculated from three aspects: domain tag matching degree, historical adoption rate, and historical follow-up question rate. The domain tag matching degree is determined according to the similarity of the tag tree level (for example, when the target tag is level four and the candidate tag is level three, the matching degree is level three divided by level four). The historical adoption rate is the number of times the question and answer has been adopted by the user divided by the total number of times it has been displayed. The historical follow-up question rate is the number of times the user continued to ask follow-up questions after adoption divided by the number of adoptions. Finally, the first matching coefficient is obtained by weighting the domain tag matching degree, historical adoption rate, and historical follow-up question rate. When determining the second matching coefficient list, for each unit in the second question-and-answer unit set, it is calculated by weighting the semantic vector cosine similarity and question-and-answer completeness. Cosine similarity reflects the semantic similarity between the question and the candidate question and answer. Question-and-answer completeness is the proportion of the answer containing the core elements of the question. The second matching coefficient is the cosine similarity multiplied by the question-and-answer completeness, thus forming a numerical list that quantifies the matching degree of each question-and-answer unit, providing a precise quantitative basis for subsequent cross-fusion.

[0080] Optionally, in step S320, a first matching coefficient list for the first question-and-answer unit set and a second matching coefficient list for the second question-and-answer unit set are determined, including:

[0081] S321, calculate the matching degree between each first question and answer unit in the first question and answer unit set and the target question information, obtain the first matching coefficient of each first question and answer unit, and form a first matching coefficient list of all first matching coefficients; wherein, the first matching coefficient is calculated by domain tag, historical adoption rate score and historical follow-up question rate.

[0082] Understandably, the matching degree of each first question-and-answer unit needs to be quantified from multiple dimensions. For example, firstly, extract its domain tags and perform semantic analysis with the domain tags of the target question (such as calculating tag path similarity based on WordNet or calculating overlap through co-occurrence frequency). Then, calculate the adoption rate of the unit in historical interactions (the ratio of the number of times it is confirmed as valid by the user to the number of times it is displayed). At the same time, convert the frequency of subsequent follow-up questions to users into an inverse score (such as 1 - follow-up question rate to reflect the sufficiency of the answer). Finally, the three indicators are weighted and summed according to preset weights (such as domain tags accounting for 40%, historical adoption rate accounting for 35%, and historical follow-up question rate accounting for 25%) to obtain the first matching coefficient of the first question-and-answer unit.

[0083] S322, calculate the matching degree between each second question and answer unit in the second question and answer unit set and the target question information, obtain the first matching coefficient of each second question and answer unit, and form a second matching coefficient list of all second matching coefficients; wherein, the second matching coefficient is calculated by cosine similarity and question and answer completeness.

[0084] It is understandable that each second question-and-answer unit can be evaluated based on semantic relevance and content completeness. The cosine similarity between its semantic feature vector (e.g., generated through BERT model encoding) and the semantic vector of the target question can be calculated to quantify the degree of semantic matching. Then, the coverage of the answer to the core elements of the question (such as whether it contains key information like arguments, steps, and conclusions) can be evaluated using metrics like information entropy and keyword coverage. Finally, the two metrics are weighted and fused according to preset weights (e.g., cosine similarity accounts for 60%, and question-and-answer completeness accounts for 40%) to obtain the second matching coefficient for that unit.

[0085] S330, based on the first matching coefficient list and the second matching coefficient list, the first question-and-answer unit set and the second question-and-answer unit set are cross-fused to obtain a cross-question-and-answer unit set; wherein, the cross-question-and-answer unit set includes at least one cross-question-and-answer unit.

[0086] It is understandable that the fusion of two sets can be achieved through multi-dimensional correlation calculation. The question-answering units in the two sets are combined by Cartesian product to generate a candidate set of unit pairs. Then, for each unit pair, the three-dimensional correlation degree is calculated, including domain label matching degree (such as semantic path similarity based on WordNet), semantic feature cosine similarity (such as calculated by BERT vector), and the product of the first matching coefficient and the second matching coefficient. Then, a comprehensive correlation score is generated according to preset weights (such as domain matching accounting for 30%, semantic similarity accounting for 40%, and coefficient product accounting for 30%). Finally, the semantically redundant unit pairs are eliminated by double threshold screening (such as basic threshold filtering of low correlation pairs, and then retaining the top k% according to the score) and combined with the maximum marginal relevance algorithm (MMR), ultimately forming a cross-question-answering unit set containing high-value combinations.

[0087] S340, each cross-question-answering unit in the cross-question-answering unit set is determined as a candidate question-answering unit.

[0088] It is understandable that all cross-question-answer units within the set can be traversed, and each cross-question-answer unit can be directly marked as a candidate state. This provides basic data units for the subsequent screening, evaluation, or application of candidate question-answer units, ensuring that each valid unit in the set is qualified as a candidate and laying the foundation for subsequent processing.

[0089] S400, generate at least one candidate answer scheme based on the candidate question-answer unit; wherein, any candidate answer scheme takes any candidate question-answer unit as the core content carrier.

[0090] It is understandable that answer schemes can be built based on candidate question-answer units. That is, for each candidate question-answer unit, a complete answer framework is generated by integrating relevant semantic information to ensure that the answer scheme closely revolves around the core of the question. At the same time, a multi-candidate mechanism provides diverse answer possibilities, providing sufficient options for subsequent scheme selection or optimization.

[0091] In one possible implementation, S400, at least one candidate answer scheme is generated based on the candidate question-answering unit, including:

[0092] S410, for each candidate question-answering unit, calculate the relevance between the candidate question-answering unit and the hierarchical nodes of the semantic subgraph dependency tree, and extract the node document fragments corresponding to the hierarchical nodes with a relevance greater than the threshold.

[0093] It is understandable that the semantic feature vector of the candidate question-answering unit can be compared with the feature vectors of each level of nodes in the semantic subgraph dependency tree (including semantic representations such as domain concepts and relation edges) by cosine similarity calculation, or the attention weight between nodes can be calculated to quantify the semantic correlation between the two. Then, by setting a relevance threshold (such as 0.6) in advance, the level nodes with relevance exceeding the threshold can be selected, and the document fragments (such as text blocks containing concept definitions, formula derivations, case descriptions, etc.) corresponding to these level nodes in the knowledge graph can be extracted.

[0094] S420: Based on the extracted node document fragments, candidate question-answering units are expanded and optimized through natural language generation to generate candidate answer schemes.

[0095] It is understandable that the original answer content of the candidate question-and-answer unit and the extracted node document fragments can be used as input. The key information in the answer and the supplementary knowledge in the document fragments can be aligned through the attention mechanism. Then, the content is reorganized according to the preset logical structure (such as question analysis - principle explanation - case evidence). At the same time, the professional terminology and key data in the document fragments are retained through the copy mechanism. Finally, the decoder generates fluent and natural text, forming a candidate answer scheme with in-depth background information, extended arguments or examples to support it, thereby improving the richness and professionalism of the answer.

[0096] S500 uses a pre-set deep reasoning filtering model to filter candidate answer schemes based on identity information, determines the target answer scheme, and displays the target answer scheme to the student asking the question; wherein, the deep reasoning filtering model is a pre-trained machine learning model.

[0097] It's understandable that a complete chain from user feature analysis to accurate answer recommendation can be established, breaking the "one-size-fits-all" model of traditional question-answering systems through deep reasoning technology. Different students exhibit significant differences in knowledge reserves, learning habits, and professional needs. For example, computer science students have a much deeper understanding of algorithmic problems than non-computer science students, while humanities students focus more on the logic and expression style of answers. Deep reasoning filtering models can integrate multi-dimensional features from student identity information, simulating human cognitive processes to select the most suitable answer from a set of candidate responses. These models can be trained on historical interaction data, using architectures such as Transformer and graph neural networks to learn the complex mapping relationship between user characteristics and answer quality. This enables intelligent and personalized filtering of candidate responses, ensuring that the answers presented to users satisfy both knowledge accuracy and their individual needs.

[0098] In one possible implementation, S500 uses a preset deep reasoning filtering model to filter candidate answer schemes based on identity information, determines the target answer scheme, and displays the target answer scheme to the student asking the question, including:

[0099] S510 extracts the student's professional background feature vector and historical question-and-answer preference feature vector based on the identity information.

[0100] It is understandable that representative features can be extracted from identity information. Student identity information typically includes multi-source heterogeneous data such as basic attributes (e.g., grade, major), learning records (course grades, electives), and interaction logs (historical questions, follow-up questions). When extracting professional background feature vectors, structured data such as major name and major courses can be encoded. For example, the major "Computer Science and Technology" can be mapped to a specific embedding vector. Combined with the prerequisite relationships and knowledge point associations between courses in the professional knowledge graph, a feature vector containing professional knowledge structure can be constructed. For historical question-and-answer preference feature vectors, natural language processing techniques can be used to analyze the semantic content, question type (factual questions, inference questions), and follow-up questioning patterns (questioning answers, seeking further explanations) of historical questions. Recurrent neural networks (RNNs) or Transformer models can be used to capture long-term dependencies in the question sequence, generating preference vectors that reflect students' knowledge weaknesses, areas of interest, and questioning habits, thereby comprehensively characterizing students' personalized features.

[0101] S520 calculates the first similarity between each candidate answer scheme and the professional background feature vector, and the second similarity between the candidate answer scheme and the historical question and answer preference feature vector.

[0102] It's understandable that for candidate answers, key information such as technical terms, knowledge points, and theoretical models can be extracted and encoded into semantic vectors. For example, for an answer about "the application of convolutional neural networks in image recognition," technical terms like "convolutional layer," "pooling layer," and "backpropagation" would be parsed out and converted into high-dimensional semantic vectors using a pre-trained language model (such as BERT). Then, metrics like cosine similarity and Euclidean distance are used to compare this semantic vector with the student's professional background feature vector. If the student is a senior computer science student and their professional background feature vector demonstrates in-depth learning in the field of deep learning, then an answer containing complex algorithm derivations and cutting-edge application examples will achieve a higher first similarity score. When calculating the second similarity score, the candidate answer's style of expression, level of detail, and type of examples can be analyzed to determine their alignment with the student's preferences, considering historical question-and-answer preferences. For example, if the student repeatedly asks follow-up questions about the practical application scenarios of concepts in past questions, then an answer containing rich practical examples will score higher in the second similarity calculation. The system also uses a text classification model to identify the type of response (such as popular science, academic, and general knowledge), and dynamically adjusts the weight of similarity calculation based on the student's adoption of different types of responses in their historical interactions, thereby more accurately assessing the degree of matching between candidate response schemes and student preferences.

[0103] S530 inputs the first and second similarities of each candidate answer scheme into a preset deep reasoning filtering model and outputs the comprehensive matching score of each candidate answer scheme.

[0104] This is understandable, and can be based on deep learning architectures, such as multi-layer Transformer networks or deep neural networks (DNNs). During model training, a large amount of labeled historical question-and-answer data, including student identity information, candidate answer schemes, and the final selected target answer scheme, is used to train a deep inference filtering model. The deep inference filtering model uses first and second similarity as input features and captures the correlation between the two dimensions through a multi-head attention mechanism. For example, it analyzes whether the influence weight of historical question-and-answer preferences on answer selection changes when the professional background matching degree is high. The features can be non-linearly transformed using a multi-layer perceptron (MLP) to learn the influence patterns of different feature combinations on answer quality. During the inference phase, when new candidate answer scheme similarity data is input, the model outputs a comprehensive matching score based on the knowledge learned during training. This comprehensive matching score not only considers the independent influence of professional background and historical question-and-answer preferences but also integrates the synergistic effect between the two. For example, for an answer that is difficult to answer due to professional knowledge but whose expression style matches students' preferences, the deep reasoning screening model will reasonably weigh the professional matching degree and the preference matching degree based on feedback from similar situations in the training data, and give an objective comprehensive score, thereby achieving a comprehensive evaluation of the candidate answer.

[0105] For example, the training of the deep reasoning screening model can be achieved by constructing a multi-source heterogeneous training dataset, including collecting historical question-and-answer interaction records, user identity information, and feedback data, integrating professional knowledge bases to construct domain knowledge data, and constructing positive and negative samples at a 1:3 ratio through key information replacement, etc., and then extracting user professional background and historical preference features, as well as semantic and structural features of candidate answer schemes by labeling the evaluation dimensions of high-quality answers. The model adopts a three-layer architecture of "feature fusion-deep reasoning-score output". The input layer concatenates user features and matching degree features to form a multi-dimensional tensor. The intermediate layer captures the correlation between features through multi-layer Transformer or deep neural network combined with multi-head attention mechanism, and learns the synergistic influence pattern of professional suitability and preference matching degree. The output layer generates a comprehensive matching score in the 0-1 interval through an activation function. During training, the cross-entropy loss function is used as the optimization objective, the AdamW optimizer is used and combined with an early stopping strategy to prevent overfitting, and batch normalization and gradient clipping are used to improve training stability. Finally, the hyperparameters are adjusted using the validation set, enabling the model to perform deep reasoning and accurate scoring of candidate answer schemes based on user personalized features.

[0106] S540 identifies the candidate answer with the highest overall matching score as the target answer and displays the target answer to the student who asked the question.

[0107] Understandably, by comparing the comprehensive matching scores of all candidate answer schemes, the target answer scheme with the highest score can be selected as the final answer, and the target answer scheme can be shown to the student asking the question, thereby improving the relevance and authority of the process from question understanding and knowledge retrieval to answer generation.

[0108] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0109] Corresponding to the intelligent service method based on deep reasoning and knowledge enhancement described in the above embodiments, this application also provides an intelligent service system based on deep reasoning and knowledge enhancement, wherein each unit of the system can implement each step of the intelligent service method based on deep reasoning and knowledge enhancement. Figure 2 The diagram shows a structural block diagram of an intelligent service system based on deep reasoning and knowledge enhancement provided in an embodiment of this application. For ease of explanation, only the parts related to the embodiments of this application are shown.

[0110] Reference Figure 2 This intelligent service system based on deep reasoning and knowledge enhancement includes:

[0111] The acquisition unit is used to acquire the identity information of the student asking the question and the target question information;

[0112] The parameter unit is used to construct the question-and-answer feature parameters corresponding to the questioning student based on the target question information;

[0113] The first filtering unit is used to filter at least one candidate question-answer unit from a preset question-answer knowledge base based on the question-answer feature parameters; wherein, the candidate question-answer unit includes historical similar questions and corresponding answers, and related document fragments;

[0114] A generation unit is used to generate at least one candidate answer scheme based on the candidate question-and-answer unit; wherein any candidate answer scheme takes any candidate question-and-answer unit as its core content carrier;

[0115] The second filtering unit is used to filter the candidate answer schemes using a preset deep reasoning filtering model based on the identity information, determine the target answer scheme, and display the target answer scheme to the student asking the question; wherein, the deep reasoning filtering model is a pre-trained machine learning model.

[0116] It should be noted that the information interaction and execution process between the above systems / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0117] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit module can exist physically separately, or two or more unit modules can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0118] This application also provides an electronic device. Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 3 As shown, the electronic device 6 of this embodiment includes: at least one processor 60 ( Figure 3 Only one is shown in the image), at least one memory 61 ( Figure 3 (Only one is shown in the diagram) and a computer program 62 stored in the at least one memory 61 and executable on the at least one processor 60, wherein when the processor 60 executes the computer program 62, it causes the electronic device 6 to implement the steps in any of the above embodiments of the intelligent service method based on deep reasoning and knowledge enhancement, or causes the electronic device 6 to implement the functions of each unit in the above system embodiments.

[0119] For example, the computer program 62 may be divided into one or more units, which are stored in the memory 61 and executed by the processor 60 to complete this application. The one or more units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program 62 in the electronic device 6.

[0120] Electronic device 6 can be a computing device or terminal device such as a desktop computer, laptop, handheld computer, or cloud server. This electronic device may include, but is not limited to, a processor 60 and a memory 61. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 6 and does not constitute a limitation on electronic device 6. It may include more or fewer components than shown, or combine certain components, or different components, such as input / output devices, network access devices, buses, etc.

[0121] The processor 60 can be a Central Processing Unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0122] In some embodiments, the memory 61 may be an internal storage unit of the electronic device 6, such as a hard disk or memory of the electronic device 6. In other embodiments, the memory 61 may be an external storage device of the electronic device 6, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 6. Furthermore, the memory 61 may include both internal and external storage units of the electronic device 6. The memory 61 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory 61 can also be used to temporarily store data that has been output or will be output.

[0123] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0124] This application provides a computer program product that, when run on an electronic device, causes the electronic device to perform the steps in any of the above method embodiments.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to an electronic device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0126] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0127] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0128] In the embodiments provided in this application, it should be understood that the disclosed intelligent service system / electronic device and method based on deep reasoning and knowledge enhancement can be implemented in other ways. For example, the embodiments of the intelligent service system / electronic device based on deep reasoning and knowledge enhancement described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0129] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0130] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An intelligent service method based on deep reasoning and knowledge enhancement, characterized in that, include: Obtain the identity information of the student asking the question and the target question information; Construct question-and-answer feature parameters corresponding to the student asking the question based on the target question information; Based on the question-and-answer feature parameters, at least one candidate question-and-answer unit is selected from a preset question-and-answer knowledge base; wherein, the candidate question-and-answer unit includes historical similar questions and corresponding answers, and related document fragments; At least one candidate answer scheme is generated based on the candidate question-and-answer unit; wherein, any candidate answer scheme uses any candidate question-and-answer unit as its core content carrier; Based on the identity information, a preset deep reasoning filtering model is used to filter the candidate answer schemes, determine the target answer scheme, and display the target answer scheme to the student who asked the question; wherein, the deep reasoning filtering model is a pre-trained machine learning model; Before obtaining the identity information of the student asking the question and the target question information, the method further includes: Obtain documents related to the question-and-answer content, and construct a question-and-answer knowledge base based on the documents related to the question-and-answer content; wherein, the question-and-answer knowledge base includes a first knowledge repository constructed based on domain tags, and a second knowledge repository constructed based on semantic feature vectors; The step of obtaining documents related to the question-and-answer content and constructing a question-and-answer knowledge base based on these documents includes: Obtain relevant documents for question and answer content, and extract seed question and answer units with labeled domain tags from the relevant documents for question and answer content; Semantic analysis is performed on each seed question-answering unit to generate a feature vector corresponding to the seed question-answering unit; A semantic index database is constructed based on the feature vectors, and the semantic index database is used to automatically assign domain labels to question-answering units that are not labeled with domain labels. Based on the domain tags, all seed question-and-answer units are clustered to form the first knowledge repository; The step of constructing the question-and-answer feature parameters corresponding to the questioning student based on the target question information includes: Semantic parsing is performed based on the target question information to determine the semantic knowledge subgraph corresponding to the target question information and the semantic index fragment of the semantic knowledge subgraph; Construct the semantic adjacency tensor corresponding to the semantic knowledge subgraph based on the semantic index fragment; The semantic adjacency tensor is topologically sorted to generate a semantic subgraph dependency tree; wherein, the semantic subgraph dependency tree explicitly expresses the hierarchical dependency relationship between semantic knowledge subgraphs in a tree structure; The semantic subgraph dependency tree is hierarchically decomposed to determine the question-and-answer feature parameters corresponding to the student asking the question.

2. The intelligent service method based on deep reasoning and knowledge enhancement as described in claim 1, characterized in that, Automatically assigning domain labels to unlabeled question-answering units using the semantic index database includes: Semantic analysis is performed on question-answering units without domain labels to generate target feature vectors; Calculate the similarity between the target feature vector and each feature vector in the semantic index database, and filter out candidate seed question-answering units whose similarity exceeds a threshold; The candidate seed question-answering units are subjected to semantic consistency verification to determine the target seed question-answering unit with the highest matching degree; Assign the domain label of the target seed question-and-answer unit to the question-and-answer unit.

3. The intelligent service method based on deep reasoning and knowledge enhancement as described in claim 1, characterized in that, The question-answering feature parameters include domain labels and semantic feature vectors; The step of selecting at least one candidate question-answer unit from a preset question-answer knowledge base based on the question-answer feature parameters includes: Match the corresponding first question-answering unit set from the first knowledge warehouse based on the domain label, and retrieve the corresponding second question-answering unit set from the second knowledge warehouse based on the semantic feature vector; Determine the first matching coefficient list of the first question-and-answer unit set and the second matching coefficient list of the second question-and-answer unit set, respectively; Based on the first matching coefficient list and the second matching coefficient list, the first question-and-answer unit set and the second question-and-answer unit set are cross-fused to obtain a cross-question-and-answer unit set; wherein, the cross-question-and-answer unit set includes at least one cross-question-and-answer unit; Each of the cross-question-answering units in the set of cross-question-answering units is determined as a candidate question-answering unit.

4. The intelligent service method based on deep reasoning and knowledge enhancement as described in claim 3, characterized in that, The step of determining the first matching coefficient list of the first question-and-answer unit set and the second matching coefficient list of the second question-and-answer unit set includes: Calculate the degree of matching between each first question and answer unit in the first question and answer unit set and the target question information to obtain the first matching coefficient of each first question and answer unit, and form a first matching coefficient list by combining all the first matching coefficients; wherein, the first matching coefficient is calculated by domain tag, historical adoption rate score and historical follow-up question rate; Calculate the degree of matching between each second question-and-answer unit in the second question-and-answer unit set and the target question information to obtain the first matching coefficient of each second question-and-answer unit, and form a second matching coefficient list by combining all the second matching coefficients; wherein, the second matching coefficient is calculated by cosine similarity and question-and-answer completeness.

5. The intelligent service method based on deep reasoning and knowledge enhancement as described in claim 4, characterized in that, The generation of at least one candidate answer scheme based on the candidate question-answering unit includes: For each candidate question-answering unit, the relevance between the candidate question-answering unit and the hierarchical nodes of the semantic subgraph dependency tree is calculated, and the node document fragments corresponding to the hierarchical nodes with a relevance greater than a threshold are extracted. Based on the extracted node document fragments, the candidate question-answering units are expanded and optimized using natural language generation to generate candidate answer schemes.

6. The intelligent service method based on deep reasoning and knowledge enhancement as described in claim 1, characterized in that, The candidate answer schemes are filtered using a preset deep reasoning filtering model based on the identity information to determine the target answer scheme, and the target answer scheme is displayed to the student asking the question; wherein, the deep reasoning filtering model is a pre-trained machine learning model, including: Based on the identity information, extract the student's professional background feature vector and historical question-and-answer preference feature vector; Calculate the first similarity between each candidate answer scheme and the professional background feature vector, and the second similarity between each candidate answer scheme and the historical question-and-answer preference feature vector; The first similarity and the second similarity of each candidate answer scheme are input into a preset deep reasoning filtering model, and the comprehensive matching score of each candidate answer scheme is output. The candidate answer with the highest comprehensive matching score is selected as the target answer and displayed to the student who asked the question.

7. An intelligent service system based on deep reasoning and knowledge enhancement, characterized in that, For implementing the method of any one of claims 1 to 6, the intelligent service system based on deep reasoning and knowledge enhancement comprises: The acquisition unit is used to acquire the identity information of the student asking the question and the target question information; The parameter unit is used to construct the question-and-answer feature parameters corresponding to the questioning student based on the target question information; The first filtering unit is used to filter at least one candidate question-and-answer unit from a preset question-and-answer knowledge base based on the question-and-answer feature parameters; wherein, the candidate question-and-answer unit includes historical similar questions and corresponding answers, and related document fragments; A generation unit is used to generate at least one candidate answer scheme based on the candidate question-and-answer unit; wherein, any candidate answer scheme takes any candidate question-and-answer unit as its core content carrier; The second filtering unit is used to filter the candidate answer schemes using a preset deep reasoning filtering model based on the identity information, determine the target answer scheme, and display the target answer scheme to the student asking the question; wherein, the deep reasoning filtering model is a pre-trained machine learning model.

Images