Knowledge graph and deep reinforcement learning-based cultural and tourism activity interaction method and system

Abstract

The embodiment of the application relates to the field of artificial intelligence, and provides a cultural and travel activity interaction method and system based on a knowledge graph and deep reinforcement learning, which comprises the following steps: acquiring local cultural resource data and constructing a cultural and travel knowledge graph, and performing knowledge representation learning on the cultural and travel knowledge graph; a candidate activity interaction script is constructed by a teacher-end intelligent agent based on the association relationship between cultural elements and teaching tasks, and teaching constraints, strategy correction and / or manual intervention are performed; real-time state information is collected by a student-end intelligent agent in an interaction process; a target activity interaction scheme is obtained by a deep reinforcement learning intelligent agent in combination with the candidate activity interaction script; and an activity interaction guide result and a teaching monitoring result are pushed based on the target activity interaction scheme, and matched teaching content is generated. Thus, the association between local culture and cultural and travel activities is deeply mined to dynamically optimize the activity scheme, the difficulty of manual planning and classroom guidance is reduced, and the generation efficiency of teaching content and dissemination content is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and more specifically to a method and system for interactive cultural and tourism activities based on knowledge graphs and deep reinforcement learning. Background Technology

[0002] With the continuous iteration of large-scale modeling and multimodal generation technologies, the cultural and tourism industry is placing higher demands on intelligentization in areas such as event planning, content marketing, and tourist services. Existing technologies often break down the cultural and tourism event planning process into multiple stages, including resource organization, theme extraction, process arrangement, and promotional material production. Planning primarily relies on human experience and case reuse, supplemented by rule-based or tag-based recommendation methods to match tourist preferences. In terms of knowledge organization, some solutions utilize knowledge graphs to structure and describe elements such as attractions, transportation, dining, accommodation, and activities for retrieval, question answering, or recommendation explanations. Regarding content production, AIGC tools are used to generate materials such as attraction introductions, travel guides, promotional articles, poster illustrations, short video scripts, speech synthesis, and digital avatars to improve production efficiency.

[0003] However, existing technologies rely on templates and reuse of historical cases in the planning process, lacking in-depth exploration of local characteristics, resulting in homogenization of event themes, processes, and interactive formats. Furthermore, AIGC (AI-generated content) often generates copy, posters, or video scripts independently using single tools and manual prompts, easily leading to mismatches between event content and processes, requiring repeated manual proofreading and revisions, which fails to meet the actual needs of cultural tourism operation training or event planning. Therefore, a new technological solution is urgently needed to address the technical problems of homogenized solutions and mismatched event content and processes in existing technologies. Summary of the Invention

[0004] This application provides a method and system for interactive cultural and tourism activities based on knowledge graphs and deep reinforcement learning. It can solve the technical problems existing in the prior art, such as reliance on human experience and template reuse in the planning of cultural and tourism activities, insufficient exploration of local cultural characteristics, serious homogenization of activity themes and processes, poor matching between generated content and activity processes, and difficulty in adapting to classroom teaching and operational training scenarios.

[0005] In a first aspect, embodiments of this application provide a method for interactive cultural and tourism activities based on knowledge graphs and deep reinforcement learning. This method is executed collaboratively by a teacher-side intelligent agent and a student-side intelligent agent, and includes: Acquire local cultural resource data related to cultural and tourism teaching activities, and construct a cultural and tourism knowledge graph based on the local cultural resource data; Knowledge representation learning is performed on the cultural and tourism knowledge graph to obtain vectorized representations of each entity in the cultural and tourism knowledge graph and the relationships between them; The teacher-side intelligent agent constructs candidate activity interaction scripts adapted to the current teaching scenario based on the relationship between the Chinese cultural elements and teaching tasks in the vectorized representation. The candidate activity interaction scripts are used to represent the activity organization logic, task arrangement logic, interaction triggering logic, and content output logic. Teaching constraints, strategy corrections, and / or manual interventions are performed on the candidate activity interaction scripts. The student-side intelligent agent collects real-time status information during the interaction process based on the candidate activity interaction script to characterize user participation, teaching execution, and / or scene changes. A deep reinforcement learning agent, combined with the candidate activity interaction script, makes a strategy decision on the real-time state information to obtain the target activity interaction scheme. Based on the target activity interaction scheme, the student-side intelligent agent pushes the activity interaction guide results to the student terminal, the teacher-side intelligent agent pushes the teaching monitoring results to the teacher terminal, and calls the content distribution module integrated with the AIGC toolchain to generate teaching content that matches the target activity interaction scheme.

[0006] Compared to existing technologies, the above method, by constructing a cultural tourism knowledge graph and performing knowledge representation learning, can uncover potential semantic relationships between local cultural elements and between cultural elements and teaching tasks. Through the collaborative interaction of teacher-side and student-side intelligent agents, collaborative control of teaching constraints, interactive guidance, and state acquisition is achieved. A deep reinforcement learning intelligent agent makes dynamic strategy decisions on candidate activity interaction scripts based on real-time state information, achieving adaptive optimization of the target activity interaction scheme. Furthermore, a content distribution module integrating an AIGC toolchain generates teaching content that matches the target activity interaction scheme. Therefore, the embodiments of this application can effectively solve the technical problems of homogenization in cultural tourism activity planning, poor matching between activity content and activity process, and high costs of manual intervention in existing technologies, thereby improving the depth of local cultural exploration, the flexibility of teaching organization, the interactive experience of activities, and the efficiency of content generation.

[0007] Secondly, embodiments of this application provide a cultural and tourism activity interaction system based on knowledge graphs and deep reinforcement learning, which has the function of implementing the cultural and tourism activity interaction method based on knowledge graphs and deep reinforcement learning provided in the first aspect above. The function can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above functions, and the modules can be software and / or hardware.

[0008] In one embodiment, the cultural tourism activity interaction system based on knowledge graphs and deep reinforcement learning includes a server, a teacher-side intelligent agent, and a student-side intelligent agent. The system further comprises: The server is configured to acquire local cultural resource data related to cultural tourism teaching activities, and construct a cultural tourism knowledge graph based on the local cultural resource data; and perform knowledge representation learning on the cultural tourism knowledge graph to obtain vectorized representations of each entity in the cultural tourism knowledge graph and the relationships between each entity. The teacher-side intelligent agent is configured to construct candidate activity interaction scripts adapted to the current teaching scenario based on the correlation between cultural elements and teaching tasks in the vectorized representation. The candidate activity interaction scripts are used to represent activity organization logic, task arrangement logic, interaction triggering logic, and content output logic. The agent also performs teaching constraints, strategy corrections, and / or manual interventions on the candidate activity interaction scripts. Based on the target activity interaction scheme, the agent pushes teaching monitoring results to the teacher terminal. The student-side intelligent agent is configured to collect real-time status information during the interaction process based on the candidate activity interaction script to characterize user participation, teaching execution, and / or scene changes; and to push activity interaction guide results to the student terminal based on the target activity interaction scheme. The server is also configured to call a deep reinforcement learning agent to make policy decisions based on the real-time state information in conjunction with the candidate activity interaction script, thereby obtaining the target activity interaction scheme; and to call a content distribution module integrated with the AIGC toolchain to generate teaching content that matches the target activity interaction scheme.

[0009] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the cultural and tourism activity interaction method based on knowledge graphs and deep reinforcement learning as described in the first aspect.

[0010] Fourthly, embodiments of this application provide a computing 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 cultural and tourism activity interaction method based on knowledge graphs and deep reinforcement learning described in the first aspect.

[0011] Fifthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, cause the computer to execute the cultural and tourism activity interaction method based on knowledge graphs and deep reinforcement learning provided in the first aspect.

[0012] Compared to existing technologies, this application embodiment, by constructing a cultural tourism knowledge graph and performing knowledge representation learning on the knowledge graph, can uncover potential semantic relationships between local cultural elements and between cultural elements and teaching tasks. Since this application embodiment uses a semantic association modeling mechanism based on knowledge graphs to structurally organize local cultural resources and identify deep relationships, rather than using template reuse, case splicing, or rule-based shallow matching in existing technologies, it can more accurately extract local cultural characteristics and generate more targeted activity themes, task arrangements, and interaction logic. Therefore, this application embodiment can effectively reduce the homogenization problem in cultural tourism activity planning.

[0013] Because this application embodiment introduces a collaborative interaction mechanism between teacher-side and student-side intelligent agents, with the teacher-side agent responsible for teaching constraints, strategy correction, and manual intervention, and the student-side agent responsible for interactive guidance, task prompts, and real-time status collection, this application embodiment can simultaneously balance teaching standardization, activity flexibility, and user participation, thereby improving the intelligence level of cultural and tourism activity organization and execution in classroom teaching scenarios. Furthermore, because this application embodiment further introduces a deep reinforcement learning intelligent agent and makes dynamic strategy decisions on candidate activity interaction scripts based on real-time status information, this application embodiment can adaptively adjust the activity flow, task triggering method, and content output strategy according to user participation, teaching execution, and scenario changes, obtaining a target activity interaction scheme that is more suitable for the current teaching scenario. This can achieve ideal dynamic optimization effects and significantly improve the personalized adaptability and interactive experience quality of teaching activities. In addition, since this application embodiment also calls the content distribution module integrated with the AIGC toolchain to generate teaching content that matches the target activity interaction scheme, instead of having multiple AIGC tools independently generate text, posters or scripts and then manually splicing and correcting them repeatedly as in the prior art, this application embodiment can improve the semantic consistency and execution correspondence between activity content and activity process, reduce manual modification costs, and improve the generation efficiency of teaching content and dissemination content.

[0014] In summary, the embodiments of this application, through the organic combination of knowledge graphs, dual-end intelligent agent collaboration, deep reinforcement learning, and AIGC toolchain, achieve in-depth mining of local cultural characteristics, dynamic optimization of classroom interaction processes, and collaborative generation of teaching content. Therefore, they can meet the application needs in scenarios such as cultural tourism operation teaching, classroom training, and cultural tourism activity planning assistance. Attached Figure Description

[0015] The objectives, features, and advantages of the embodiments of this application will become readily understood by referring to the accompanying drawings and reading the detailed description of the embodiments.

[0016] Figure 1This is a schematic diagram of a system for a cultural and tourism activity interaction method based on knowledge graphs and deep reinforcement learning, as described in this application. Figure 2 This is a flowchart illustrating a cultural tourism activity interaction method based on knowledge graphs and deep reinforcement learning, according to an embodiment of this application. Figure 3 This is a schematic diagram of the structure of a cultural tourism activity interaction system based on knowledge graphs and deep reinforcement learning, according to an embodiment of this application. Detailed Implementation

[0017] This application provides a method and system for interactive cultural and tourism activities based on knowledge graphs and deep reinforcement learning. It can be applied to intelligent interactive systems in scenarios such as cultural and tourism operation teaching, classroom training, study tour design, local culture course teaching, and cultural and tourism activity planning assistance. The intelligent interactive system may include a server, a teacher terminal, and a student terminal. The teacher-side intelligent agent can be deployed on the teacher terminal or the server, and the student-side intelligent agent can be deployed on the student terminal or the server. The deep reinforcement learning intelligent agent, knowledge graph construction module, knowledge representation learning module, and content distribution module integrated with the AIGC toolchain can be deployed on the server. The above modules and intelligent agents can be deployed in an integrated manner or in a distributed collaborative manner.

[0018] In this embodiment, the server is at least used to acquire local cultural resource data related to cultural tourism teaching activities, construct a cultural tourism knowledge graph, and perform knowledge representation learning on the cultural tourism knowledge graph to obtain vectorized representations of each entity and the relationships between entities in the cultural tourism knowledge graph. The teacher-side agent is at least used to construct candidate activity interaction scripts adapted to the current teaching scenario based on the association between cultural elements and teaching tasks in the vectorized representations, and to execute teaching constraints, strategy corrections, and / or manual interventions on the candidate activity interaction scripts. The student-side agent is at least used to collect real-time status information representing user participation, teaching execution, and / or scenario changes during the interaction process based on the candidate activity interaction scripts, and to push activity interaction guide results to student terminals. The server can also call a deep reinforcement learning agent to make strategy decisions based on the real-time status information in conjunction with the candidate activity interaction scripts to obtain a target activity interaction scheme, and call a content distribution module integrated with the AIGC toolchain to generate teaching content matching the target activity interaction scheme.

[0019] This application's embodiments involve artificial intelligence technology, knowledge graph technology, machine learning technology, deep reinforcement learning technology, natural language processing technology, and AIGC generation technology. Specifically, artificial intelligence technology is used to realize knowledge organization, semantic understanding, strategy decision-making, and content generation in the interactive process of cultural and tourism activities; knowledge graph technology is used to perform structured modeling of local cultural resources, teaching tasks, scene elements, and resource elements; machine learning and deep reinforcement learning technologies are used to achieve dynamic optimization of target activity interaction schemes; natural language processing technology can be used for semantic parsing of local cultural resource information and teaching task information; AIGC generation technology can be used to generate teaching materials, promotional materials, visual material prompts, courseware display content, and short video scripts that match the activity theme, activity process, and interactive nodes.

[0020] Knowledge graphs are a method of organizing knowledge by structuring heterogeneous data through entities, relationships, and attributes. In this application's embodiments, a cultural tourism knowledge graph can be used to describe the relationships between cultural objects, historical events, folk activities, performing arts forms, architectural styles, regional products, teaching task elements, activity scene elements, and activity resource elements. Deep reinforcement learning is a technology that combines the representational capabilities of deep neural networks with the decision-making capabilities of reinforcement learning. It can dynamically adjust the strategies of candidate activity interaction scripts based on real-time state information, thereby obtaining a target activity interaction scheme adapted to the current teaching scenario. AIGC technology can generate multimodal teaching content consistent with the target activity interaction scheme to improve the matching degree between activity flow and content output.

[0021] In existing technologies, cultural and tourism activity planning typically relies on human experience, template reuse, or piecing together historical cases, resulting in insufficient depth in exploring local cultural characteristics and a tendency for activity flow design and interactive formats to become homogenized. Furthermore, most existing AIGC tools exist as independent tools, usually requiring manual input of prompts to generate copy, posters, or video scripts. The generated content lacks a unified logical constraint, easily leading to mismatches between activity content and flow, and failing to meet the requirements for dynamic interaction, teaching constraints, and content consistency in cultural and tourism teaching activities, classroom training, and operational support scenarios.

[0022] Compared to existing technologies, this application's embodiments, by constructing a cultural tourism knowledge graph and performing knowledge representation learning, can identify potential semantic relationships between local cultural elements and between cultural elements and teaching tasks; through the collaborative execution of teacher-side and student-side intelligent agents, it can achieve dual-end linkage of teaching constraints, interactive guidance, status collection, and process monitoring; through deep reinforcement learning intelligent agents making strategy decisions based on real-time status information, it can dynamically generate or adjust target activity interaction schemes; and by generating teaching content matching the target activity interaction scheme through a content distribution module integrated with the AIGC toolchain, it can improve the semantic consistency between activity flow and content output. Therefore, this application's embodiments can effectively solve the technical problems in existing technologies such as homogenized activity schemes, insufficient exploration of local cultural characteristics, and poor matching between activity content and flow.

[0023] In some implementations, the server, teacher terminal, and student terminal can be configured, with reference to the accompanying drawings, into a cultural tourism activity interactive system based on knowledge graphs and deep reinforcement learning. For example, referring to... Figure 1 The system may include a server 01, a teacher terminal 02, and a student terminal 03. The server 01 may deploy a knowledge graph construction module, a knowledge representation learning module, a deep reinforcement learning agent, and a content distribution module; the teacher terminal 02 may deploy a teacher-side agent; and the student terminal 03 may deploy a student-side agent. Here, besides… Figure 1 In addition to the desktop computer shown, teacher terminal 02 and student terminal 03 can also be other types of terminal devices. This is just an example and does not limit the device type.

[0024] In one embodiment, the teacher terminal 02 can be used to input local cultural resource information, activity scene information, teaching objective information, teaching task information, and activity resource information related to cultural tourism teaching activities, and send the information to the server 01. The server 01 can preprocess the information, perform entity recognition, relationship extraction, and attribute alignment to construct a cultural tourism knowledge graph, and further learn the knowledge representation of each entity and the relationship between entities in the cultural tourism knowledge graph through a knowledge graph embedding model to obtain the corresponding vectorized representation. The teacher-side intelligent agent can construct candidate activity interaction scripts adapted to the current teaching scene based on the association between cultural elements and teaching tasks in the vectorized representation, and execute teaching constraints, strategy corrections, and / or manual interventions on the candidate activity interaction scripts according to teaching needs.

[0025] In one embodiment, the student-side agent can collect real-time status information representing user participation, teaching execution, and / or scene changes during the interaction process, based on the candidate activity interaction script. This information includes user engagement, task completion status, environmental adaptability, achievement of teaching objectives, and other information reflecting the execution status of classroom activities. The real-time status information can be sent to server 01, which then invokes a deep reinforcement learning agent to make strategy decisions based on the real-time status information and the candidate activity interaction script, thereby obtaining the target activity interaction scheme.

[0026] In one embodiment, server 01 can also, based on the target activity interaction scheme, drive the student-side agent to push activity interaction guidance results to student terminal 03. These results may include node identifiers, participation paths, task prompts, interaction criteria, and recommendation criteria. Simultaneously, server 01 can also drive the teacher-side agent to push teaching monitoring results to teacher terminal 02. These results may include student participation status, task completion status, teaching progress matching status, and strategy adjustment criteria. Furthermore, server 01 can call a content distribution module integrated with the AIGC toolchain to generate teaching content matching the target activity interaction scheme and send this content to teacher terminal 02 and / or student terminal 03 for classroom demonstration, activity execution, and teaching review.

[0027] It should be noted that the server involved in the embodiments of this application can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, middleware services, and artificial intelligence platform services. The teacher terminal and student terminal can be a desktop computer, laptop computer, tablet device, smartphone, interactive teaching all-in-one machine, wearable device, or other terminal device capable of data interaction, interface display, and information collection.

[0028] It should be understood that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit the scope of protection of this application. For those skilled in the art, various modifications, substitutions and improvements can be made to the above embodiments without departing from the spirit and substance of this application, and all such modifications, substitutions and improvements should fall within the scope of protection of this application.

[0029] Reference Figure 2 , Figure 2 This is a flowchart illustrating a cultural tourism activity interaction method based on knowledge graphs and deep reinforcement learning, provided in an embodiment of this application. This method can be executed by a cultural tourism activity interaction system based on knowledge graphs and deep reinforcement learning. The method is executed collaboratively by a teacher-side intelligent agent and a student-side intelligent agent. The method includes steps 101-106: Step 101: Obtain local cultural resource data related to cultural and tourism teaching activities, and construct a cultural and tourism knowledge graph based on the local cultural resource data.

[0030] In this embodiment of the application, the local cultural resource data includes, but is not limited to, intangible cultural heritage archives, local chronicles, folk literature, regional architectural data, local product data, cultural performance data, and activity scene information, teaching objective information, teaching task information, and activity resource information related to classroom teaching, which are entered by teachers on the terminal or obtained by the server.

[0031] In this embodiment, the acquisition of local cultural resource data related to cultural and tourism teaching activities can be achieved through methods such as teacher terminal input, server-side collection, and access via third-party data interfaces. Specifically, for local cultural resource data such as intangible cultural heritage archives, local chronicles, folk literature, regional architectural data, local product data, and cultural performance data, teachers can manually input relevant text, images, audio and video descriptions, tag information, and structured form content through their teacher terminals. Alternatively, the server can automatically retrieve, synchronize, or import data in batches from school resource databases, local cultural databases, public cultural and museum open platforms, government cultural and tourism data platforms, digital libraries, digital museums, or publicly available internet databases. For activity scene information, teaching objective information, teaching task information, and activity resource information related to classroom teaching, teachers can input them through a preset information input interface according to the teaching plan, classroom task arrangement, and practical training activity requirements, or extract and synchronize them from the teaching management system, course resource system, and practical training platform. Furthermore, student characteristic information, learning behavior information, and activity participation intention information can be obtained through methods such as active filling in by students on their terminals, account authorization, extraction of historical learning records, or real-time interactive collection in the classroom. Through the above-mentioned multi-source heterogeneous data collection methods, it is possible to achieve unified aggregation of local cultural resources, teaching organization information, and individual student characteristics, providing a complete data foundation for the subsequent construction of a cultural tourism knowledge graph.

[0032] In some implementations, to ensure the availability and consistency of the acquired data, the server can also perform preprocessing operations on data from different sources. These preprocessing operations may include data cleaning, deduplication, format conversion, missing value completion, time field normalization, unified geographic location encoding, and semantic standardization. For example, for unstructured text from local gazetteer documents, the server can use natural language processing techniques to extract information on historical figures, traditional architecture, folk activities, and regional products. For images, videos, and courseware materials uploaded by teachers, the server can extract their corresponding titles, descriptions, time tags, location tags, and course tags. For learning behavior information from students' terminals, the server can extract click frequency, dwell time, answer results, task completion progress, and interest preferences. Through these processes, the original, scattered, and inconsistently formatted data can be transformed into standardized input data suitable for subsequent entity recognition, relation extraction, and attribute alignment.

[0033] In step 101, the server can preprocess, standardize, semantically parse, identify entities, extract relationships, and align attributes of the local cultural resource data to extract cultural objects, historical events, folk activities, performance forms, architectural styles, local products, teaching task elements, scene elements, and resource elements. Further, these elements are used as entity nodes, and the semantic, scene, task, and resource relationships between entities are used as relation edges to construct a cultural tourism knowledge graph.

[0034] This step allows for the structured organization of scattered local cultural information and teaching activity information, providing a data foundation for subsequent knowledge representation learning and activity interaction script generation.

[0035] In this embodiment, the cultural tourism knowledge graph is a knowledge representation model that structures and semantically relates local cultural resources, teaching tasks, scene information, activity resources, and user profile information involved in cultural tourism teaching activities. Specifically, the cultural tourism knowledge graph uses entities, relationships, and attributes as basic building blocks. Entities may include cultural objects, historical events, folk activities, performing arts forms, architectural styles, regional products, teaching task elements, activity scene elements, activity resource elements, and student profile tags, etc. Relationships can be used to represent semantic, historical, scene, task, resource, and adaptation relationships between entities. Attributes can be used to describe information such as the name, category, time, location, style, applicable audience, teaching difficulty, resource quantity, and topic tags of each entity. By unifying and associating information that was originally scattered in local cultural materials, teaching plans, and activity resources in a graph format, it supports subsequent knowledge reasoning, relationship mining, candidate activity script generation, and dynamic strategy optimization.

[0036] For example, in a Minnan cultural experience teaching scenario, the cultural tourism knowledge graph can include cultural entities such as Minnan ancient houses, Gaojia opera, the mooncake gambling custom, Minnan red brick architecture, and local tea culture, as well as teaching task entities such as cultural awareness explanations, situational interactive Q&A, route guidance tasks, and short video creation tasks. Specifically, an architectural feature system can be established between Minnan ancient houses and red brick architecture; a cultural affiliation relationship can be established between Gaojia opera and local opera culture; a festival association relationship can be established between the mooncake gambling custom and Mid-Autumn Festival folk activities; and a teaching adaptation relationship can be established between Gaojia opera and role-playing tasks. Through this graph-based representation, not only can individual cultural elements be identified, but also the combination logic and adaptation relationships between cultural elements and their relationship with teaching tasks can be identified, thus providing knowledge support for the organization and interaction of cultural tourism teaching activities.

[0037] As an optional embodiment, in step 101, local cultural resource information, activity scene information, teaching objective information, teaching task information, and activity resource information entered by the teacher's terminal are obtained, along with student characteristic information, learning behavior information, and activity participation intention information input or authorized by the student's terminal. Then, the local cultural resource information is preprocessed and semantically parsed to extract cultural elements corresponding to cultural objects, historical events, folk activities, performing arts forms, architectural styles, and regional products. The activity scene information, teaching objective information, teaching task information, and activity resource information are structurally parsed to extract scene elements, teaching task elements, teaching objective elements, and resource elements. Next, entity recognition, relationship extraction, and attribute alignment are performed on the extracted cultural elements, scene elements, teaching task elements, teaching objective elements, and resource elements to establish entities corresponding to each element in the cultural tourism knowledge graph and the relationships between entities. Finally, student characteristic information, learning behavior information, and activity participation intention information are used as user profile tags and associated with the corresponding entities in the cultural tourism knowledge graph to obtain the cultural tourism knowledge graph.

[0038] First, the system acquires local cultural resource information, activity scenario information, teaching objective information, teaching task information, and activity resource information entered by teachers on their terminals, and student characteristic information, learning behavior information, and activity participation intention information entered or authorized by students on their terminals. Local cultural resource information may include introductions to intangible cultural heritage projects in a specific region, excerpts from local chronicles, records of folk activities, descriptions of traditional architecture, information on local specialties, and introductions to cultural performances. Activity scenario information may include classroom teaching scenarios, on-campus training scenarios, scenic area study tour scenarios, or online virtual teaching scenarios. Teaching objective information may include knowledge comprehension objectives, cultural awareness objectives, practical operation objectives, and dissemination and expression objectives; teaching task information may include thematic explanation tasks, route design tasks, interactive experience tasks, and content creation tasks. Activity resource information may include venue resources, explanation materials, graphic materials, video materials, digital devices, and available AIGC tools. Student characteristic information may include grade level, major, interests, and cognitive foundation; learning behavior information may include historical participation records, task completion status, and interaction preferences; and activity participation intention information may include preferences for a particular cultural theme, interactive format, or content creation method.

[0039] After obtaining the aforementioned data, the server can preprocess and semantically analyze local cultural resource information to extract cultural elements corresponding to cultural objects, historical events, folk activities, performance forms, architectural styles, and regional products. For example, for a descriptive text about ancient Minnan houses entered by a teacher, the server can extract the cultural object entity of "ancient Minnan houses," the architectural style attribute of "red brick and white stone double-slope curves," and the cultural connotation attribute of "clan settlement culture." For information about Gaojia opera, the server can extract the performance form entity of "Gaojia opera," the performance characteristic of "clown performance," and the cultural category information of "Minnan local opera." Simultaneously, the server can also perform structured analysis on activity scene information, teaching objective information, teaching task information, and activity resource information to extract scene elements, teaching task elements, teaching objective elements, and resource elements. For example, in a cultural tourism planning course, teaching scene and task elements such as classroom presentations, group collaboration, parent-child study tour theme design, and short video script generation can be extracted, along with resource elements such as PPT materials, historical images, explanatory videos, and prompt templates.

[0040] Next, the server performs entity recognition, relation extraction, and attribute alignment on the extracted cultural elements, scene elements, teaching task elements, teaching objective elements, and resource elements to establish entities corresponding to each element in the cultural tourism knowledge graph and the relationships between entities. Specifically, a named entity recognition model can be used to identify cultural entities such as Minnan ancient houses, Gaojia opera, the Bo Bing custom, Nanyin music, and tea mountains, while a relation extraction model can be used to identify relationships such as belonging to, originating from, applicable to, associated with, and usable with. For example, an "applicable" relationship can be established between "Minnan ancient houses" and "architectural culture cognition task," an "adaptable" relationship can be established between "Gaojia opera" and "role-playing interactive task," a "scene association" relationship can be established between "Bo Bing custom" and "festival folk experience scene," and a "resource support" relationship can be established between "tea mountain resources" and "parent-child study tour activities." Through attribute alignment, multiple expressions pointing to the same cultural object from different sources can be unified, for example, mapping Minnan traditional ancient dwellings, Minnan ancient houses, and Minnan red brick ancient houses to the same entity node.

[0041] Finally, student characteristic information, learning behavior information, and activity participation intention information are used as user profile tags and associated with corresponding entities in the cultural tourism knowledge graph to obtain a complete cultural tourism knowledge graph for subsequent activity script generation and strategy optimization. For example, for a student user, their profile tags may include a preference for interactive experiences, a preference for short video creation, a high interest in architectural culture, and a fast task execution speed. These tags can be associated with entities such as interactive tasks, short video script tasks, Minnan ancient houses, and medium-to-high difficulty tasks, respectively. Based on this, when the teacher's intelligent agent needs to generate candidate activity interaction scripts for the student's class, it can prioritize cultural themes and task paths that better match the student's interests and learning characteristics. For example, in a Xiamen local cultural and creative planning class, the knowledge graph can identify a potential semantic association between "Minnan ancient houses" and "Gaojia opera"—ancient architectural scenes and opera performances. This can be further combined with the "short video creation" teaching task to generate a script for an activity themed around ancient houses and opera, including node tasks such as ancient house cultural tours, opera element identification, theme poster generation, and short video storyboard creation. In this way, not only is an effective match between local cultural resources and teaching tasks achieved, but a structured knowledge foundation is also provided for subsequent knowledge representation learning, candidate activity interaction script construction, and dynamic decision-making in deep reinforcement learning.

[0042] Through the above implementation methods, step 101 not only achieves the unified acquisition of local cultural resource data, teaching activity data, and student profile data, but also realizes the transformation of various heterogeneous data into a cultural tourism knowledge graph. The constructed cultural tourism knowledge graph can incorporate cultural elements, teaching tasks, activity scenarios, resource conditions, and student characteristics into a unified knowledge network, thereby providing computable, inferable, and scalable data support for knowledge representation learning, candidate activity interaction script generation, and target activity interaction scheme optimization in subsequent steps.

[0043] Step 102: Perform knowledge representation learning on the cultural tourism knowledge graph to obtain vectorized representations of each entity in the cultural tourism knowledge graph and the relationships between them.

[0044] In the cultural tourism knowledge graph, each entity refers to a knowledge unit that has a clear semantic orientation and can be independently modeled within the context of cultural tourism teaching activities. These entities may include cultural element entities, teaching task entities, scene element entities, resource element entities, and user profile entities. Specifically, cultural element entities may include cultural objects, historical events, folk activities, performance forms, architectural styles, regional products, local figures, and cultural symbols. Teaching task entities may include knowledge explanation tasks, interactive experience tasks, route guidance tasks, study tour tasks, content creation tasks, and evaluation and feedback tasks. Scene element entities may include classroom teaching scenes, on-campus training scenes, scenic area study tour scenes, online virtual teaching scenes, and activity node scenes. Resource element entities may include image resources, text materials, video materials, explanation materials, equipment resources, venue resources, and AIGC tool resources. User profile entities may include student interest tags, learning ability tags, preference tags, behavioral characteristic tags, and participation intention tags. For example, traditional Minnan houses, Gaojia opera, the Bo Bing custom, and Nanyin music can be considered as cultural element entities; cultural knowledge explanation tasks and short video script creation tasks can be considered as teaching task entities; classroom display scenes and ancient building tour scenes can be considered as scene element entities; courseware material packages and digital human explanation resources can be considered as resource element entities; and preferences for interactive experiences and short video creation can be considered as user profile entities.

[0045] In this embodiment, the relationships between entities in the cultural tourism knowledge graph refer to semantic connections, logical connections, or application associations between different entities. These relationships may include cultural affiliation, historical evolution, form of expression, semantic similarity, teaching adaptation, scene applicability, resource support, task dependence, and user preference associations. For example, there can be a "cultural affiliation" relationship between Gaojia opera and Minnan local opera; a "architectural feature" relationship between Minnan ancient houses and red brick architectural styles; a "festival association" relationship between the mooncake gambling custom and Mid-Autumn Festival folk activities; a "teaching adaptation" relationship between Gaojia opera and role-playing tasks; a "scene applicability" relationship between ancient house cultural tour tasks and ancient building tour scenes; a "resource support" relationship between short video script creation tasks and video generation tools; and a "user preference association" relationship between preferred interactive experiences and role-playing tasks. Through these relationships, different types of entities can be connected into a knowledge network with semantic structure and application logic.

[0046] Furthermore, the corresponding vectorized representation refers to the numerical representation of entities and relationships in the cultural tourism knowledge graph mapped to a low-dimensional continuous vector space through knowledge representation learning methods. In other words, the system no longer represents entities such as Minnan ancient houses, Gaojia opera, and classroom presentation tasks solely in a symbolic manner, nor does it represent relationships such as adaptation, association, and support solely in a discrete label manner. Instead, each entity and relationship is represented as a computable vector. These vectors can preserve the semantic features of the entities themselves as well as the association features between entities, making entities with similar semantics or close relationships closer in distance and more related in direction in the vector space. For example, if "Minnan ancient houses" and "red brick architectural style" are semantically closely related, their corresponding entity vectors can show a high degree of similarity in the vector space. If "Gaojia opera" and "role-playing experience task" have a strong teaching adaptation relationship, the combination of the two with their corresponding relationship vectors can show a high association score. Through the above vectorized representation, entities and relationships in the cultural tourism knowledge graph can be transformed from symbolic knowledge structures into numerical forms suitable for machine learning model calculation and reasoning.

[0047] In this embodiment, after performing knowledge representation learning on the cultural tourism knowledge graph in step 102, the potential semantic relationships between local cultural elements and the matching relationships between local cultural elements and teaching tasks can be identified at the vector space level. The so-called potential semantic relationships between local cultural elements refer to the similarity or association between different cultural entities, even though they are not explicitly labeled in the original data, through their overall connection patterns with other entities and relationships. For example, although "Minnan ancient houses" and "Gaojia opera" belong to architectural culture and opera culture respectively, they may exhibit a strong potential semantic relationship after knowledge representation learning because they are both associated with intermediate entities such as "Minnan traditional life scenes and local folk performance activities," thus supporting the system in generating composite theme activities such as "Ancient House Opera Charm." The matching relationship between local cultural elements and teaching tasks refers to whether a cultural entity is suitable for combination with a certain type of teaching task. For example, "Nanyin" (a type of traditional Chinese music) has a high degree of matching with "audio-visual appreciation tasks," "the mooncake gambling custom" has a high degree of matching with "interactive experience tasks," and "traditional Minnan houses" has a high degree of matching with "route guidance tasks." Based on these vectorized representations, the teacher-side intelligent agent can automatically filter cultural themes, task nodes, and activity resources that are more suitable for the current teaching scenario in a data-driven manner, thus providing a reliable basis for the generation of subsequent candidate activity interaction scripts.

[0048] Step 102 can identify the potential semantic relationships between local cultural elements and the matching relationship between local cultural elements and teaching tasks, providing a data-driven basis for the subsequent generation of candidate activity interaction scripts by the teacher-side intelligent agent.

[0049] As an optional embodiment, in step 102, the cultural tourism knowledge graph is input into a preset knowledge graph embedding model to perform low-dimensional vector space mapping on the cultural element entities, teaching task element entities, scene element entities, and resource element entities in the cultural tourism knowledge graph. Then, based on the triplet relationships in the cultural tourism knowledge graph, positive and negative sample pairs are constructed. By optimizing the goal of maximizing the score of positive sample triplets and minimizing the score of negative sample triplets, the knowledge graph embedding model is iteratively trained to ensure that semantically related entities and relationships maintain corresponding geometric constraints in the vector space. After model training is completed, the entity embedding vectors corresponding to each entity in the cultural tourism knowledge graph, as well as the relationship embedding vectors representing the relationships between entities, are extracted and output. Finally, using the entity embedding vectors and relationship embedding vectors, the semantic similarity between cultural element entities and the association weights between cultural element entities and teaching task element entities, scene element entities, and resource element entities are calculated using a vector space metric method.

[0050] In practical applications, the knowledge graph embedding model can be one or a combination of RotatE model, TransE model, ComplEx model, and DistMult model.

[0051] The following describes the specific implementation process of the above optional embodiments using the ComplEx model as an example. In step 102, when performing knowledge representation learning on the cultural tourism knowledge graph, the ComplEx model can be used to perform low-dimensional vector space mapping on the cultural element entities, teaching task element entities, scene element entities, and resource element entities in the cultural tourism knowledge graph. In this way, entities and relationships originally represented in symbolic form can be transformed into continuous vector representations, thereby enabling the semantic relationships in the cultural tourism knowledge graph to have computable, comparable, and inferable characteristics.

[0052] In one implementation, the cultural tourism knowledge graph constructed in step 101 can first be represented as a set of triples, where each triple can be represented as (head entity, relation, tail entity). The head and tail entities can be cultural element entities, teaching task element entities, scene element entities, or resource element entities, and the relations can be cultural affiliation relations, teaching adaptation relations, scene applicability relations, resource support relations, semantic association relations, etc. For example, triples such as "Minnan ancient houses - applicable to - architectural culture cognition tasks," "Gaojia opera - associated with - role experience tasks," "Bo Bing custom - applicable to - festival interaction scenarios," and "short video creation tasks - dependent on - video generation resources" can be constructed. Based on the above set of triples, an original training sample set for knowledge representation learning can be formed.

[0053] Furthermore, the entities and relations in the cultural tourism knowledge graph can be mapped to a complex vector space. Taking the ComplEx model as an example, each entity and each relation corresponds to a complex embedding vector, and the dimensions of the entity vector and relation vector can be set according to the actual application scenario. In some implementations, the embedding dimension can be set to 128, 200, or 256 dimensions. Optionally, it can be set to 200 or 256 dimensions to balance expressive power and training efficiency. Compared with the traditional real vector representation, the complex vector representation can simultaneously utilize amplitude and phase information to express the deep semantic features of entities and relations, and is therefore more suitable for describing the numerous asymmetric relations, composite relations, and implicit semantic relations that exist in the cultural tourism knowledge graph. For example, mapping a cultural origin to a cultural heritage site, and mapping a cultural heritage site to a cultural origin site, although involving the same entity, have different relational directions and semantic connotations. The phase difference in the complex space can more accurately characterize such differences.

[0054] In this embodiment, a scoring function based on Hermitian bilinear form can be used to model triplet relationships. Taking the ComplEx model as an example, its scoring function can be represented as the real part of the complex inner product between the head entity vector, the relation vector, and the tail entity vector's conjugate. This scoring function allows for a quantitative evaluation of the validity of any triplet. If the semantic relationship corresponding to a triplet is true and reasonable, its score should be relatively high. If the semantic relationship corresponding to a triplet is not true or reasonable, its score should be relatively low. Compared to models with strong symmetry such as DistMult, this scoring method can better distinguish asymmetric semantic relationships in cultural and tourism teaching scenarios and is compatible with various relationship patterns such as one-to-many, one-to-one, many-to-one, and many-to-many, thus making it more suitable for knowledge representation learning in cultural and tourism knowledge graphs.

[0055] During training, negative sample triples can be constructed based on positive sample triples in the cultural tourism knowledge graph. Specifically, a negative sampling strategy oriented towards complex vector spaces can be adopted to replace the head or tail entities in the positive sample triples to generate negative samples with stronger semantic adversarial power. To avoid generating negative samples with obvious errors but no training value, in some implementations, an entity type-constrained negative sampling strategy can be adopted. That is, when replacing the head or tail entity, sampling is only performed from the set of candidate entities of the same type as the original entity. For example, for the positive sample triple "Gaojia Opera - Applicable to - Role Experience Task", when constructing a negative sample, "Role Experience Task" can be replaced with another teaching task entity instead of randomly replacing it with a scene entity or resource entity. Through this type-constrained negative sampling method, the semantic adversarial power of negative samples can be improved, the semantic confusion problem that occurs during training can be reduced, and the ability to identify the adaptation relationship between cultural element entities and teaching task element entities can be improved.

[0056] In terms of loss function design, a logistic loss function based on complex space scores can be used, combined with a sigmoid mapping to optimize the probabilities of positive and negative samples. The goal is to maximize the score of positive triples and minimize the score of negative triples, ensuring that entities and relations with genuine semantic associations maintain tighter geometric constraints in the vector space, while irrelevant or incorrectly associated entities and relations maintain greater discriminative power. In some implementations, the training optimizer can be the Adam optimizer, with a learning rate set to 1e-3, 5e-4, or 3e-4. The batch size can be set to 256 or 512; the number of negative samples corresponding to each positive sample can be set to 5 to 20. The number of training epochs can be set to 100 to 300; and the regularization coefficient can be set to 1e-5 or 1e-4. In a preferred embodiment, the embedding dimension can be set to 256, the learning rate to 5e-4, the batch size to 512, the negative sample ratio to 1:10, and the number of training rounds to 200, in order to achieve a better balance between training stability and representation performance.

[0057] After model training is complete, entity embedding vectors corresponding to each entity in the cultural tourism knowledge graph, as well as relation embedding vectors corresponding to each relation, can be extracted and output. Furthermore, using these entity embedding vectors and relation embedding vectors, the semantic similarity between cultural element entities and the association weights between cultural element entities and teaching task element entities, scene element entities, and resource element entities can be calculated using vector space metric methods. For the semantic similarity calculation between cultural element entities, the magnitude and phase information of complex vectors can be comprehensively considered, and the calculation can be performed using the real part of the complex dot product, a normalized similarity function, or a phase difference constraint function. Compared to methods based solely on real distance or cosine similarity, this calculation method can more accurately capture the deep semantic connections between cultural elements, such as the cultural origin relationships between different folk activities and the stylistic inheritance relationships between different regional architectural styles.

[0058] Furthermore, for calculating the association weights between cultural element entities and teaching task element entities, scene element entities, and resource element entities, the distribution characteristics of entity embedding vectors and relation embedding vectors in the complex vector space can be combined to obtain the adaptation strength between different entity pairs. In some implementations, different weight calculation rules can be set for different relationship types. For example, for the "teaching adaptation" relationship, the phase consistency between cultural elements and task elements can be emphasized more. For the "scene applicability" relationship, the multi-hop path score between cultural elements and scene elements can be emphasized more. For the "resource support" relationship, the direct association score between task elements and resource elements can be emphasized more. In this way, the potential semantic associations between cultural elements and the adaptation relationships between cultural elements and teaching tasks, scenes, and resources can be evaluated simultaneously in the same vector space.

[0059] In some alternative implementations, multi-hop relational reasoning can also be performed based on the entity embedding vector and relation embedding vector. For example, the feasibility and resource accessibility of a set of cultural themes in a specific classroom scenario can be evaluated according to the combination path of cultural elements, teaching tasks, scene elements, and resource elements. Taking the ancient Minnan houses as an example, it can first be identified that they have a high degree of fit with the "architectural culture cognition task," then the task can be identified that it has a high degree of scene applicability with the "classroom presentation scene," and further, the "classroom presentation scene" can be identified that it has a high degree of resource support relationship with "text and image courseware resources and short video generation resources." Thus, it can not only determine that the "ancient Minnan houses" are suitable for the current classroom activity, but also further determine what task form and content generation method should be used to present them. This multi-hop reasoning capability is an important improvement of the ComplEx model in this embodiment, which can automatically discover the implicit and indirect relationships between local cultural elements and transform them into script generation basis that is meaningful to the teacher's agent.

[0060] Optionally, based on the calculated semantic similarity and association weights, the potential semantic relationships between local cultural elements and the matching degree between local cultural elements and teaching tasks can be extracted, providing a data-driven decision-making basis for the teacher-side agent to generate candidate activity interaction scripts that match the current teaching scenario. Specifically, based on the semantic similarity between cultural element entities, a set of core cultural elements with high thematic consistency can be selected from several cultural entities; then, based on the association weights between the core cultural element set and teaching task elements, the task node that best matches the current teaching objective can be selected; next, combining the adaptation weights of scene element entities and resource element entities, scene configurations and resource combinations that can be actually executed in the current classroom environment can be selected; finally, the above cultural elements, task nodes, scene configurations, and resource combinations are output as association basis that the teacher-side agent can call, for constructing candidate activity interaction scripts.

[0061] For example, in a Xiamen local cultural and creative planning course, cultural elements such as traditional Minnan houses, Gaojia opera, Nanyin music, and the Bo Bing (a traditional dice game) custom can be identified from a cultural tourism knowledge graph. After knowledge representation learning, it is found that traditional Minnan houses and Gaojia opera have a high latent semantic association in the complex vector space; Gaojia opera has a high pedagogical fit with role-playing tasks; and traditional Minnan houses have a high matching degree with cultural tour tasks. Furthermore, role-playing and cultural tour tasks also maintain a high degree of fit with classroom demonstration scenarios and short video creation resources, respectively. Based on these results, candidate activity interaction scripts for the theme of traditional houses and opera can be generated, including node tasks such as traditional house cultural tours, opera role-playing experiences, theme poster generation, and short video storyboard creation. Therefore, the vectorized representation obtained in step 102 can not only support the discovery of implicit semantics between local cultural elements but also support the automatic matching of cultural content with teaching tasks and scenario resources.

[0062] Further optionally, in the above embodiments, it is assumed that the knowledge graph embedding model adopts the ComplEx model architecture. Based on this, the low-dimensional vector space mapping of cultural element entities, teaching task element entities, scene element entities, and resource element entities in the cultural tourism knowledge graph includes: mapping each entity and each relation in the cultural tourism knowledge graph into a complex vector representation; using a scoring function based on Hermitian bilinear form to model the triple relations in the cultural tourism knowledge graph, wherein the scoring function is used to calculate the association score between the head entity, relation, and tail entity; and characterizing the asymmetric relations, multi-type relations, and multi-hop combination relations between entities based on the association score, so as to improve the ability to represent the implicit semantic associations between local cultural elements and the adaptation relationship between local cultural elements and teaching tasks.

[0063] Subsequently, in the above embodiments, the step of constructing positive and negative sample pairs based on the triplet relationships in the cultural and tourism knowledge graph, and iteratively training the knowledge graph embedding model by maximizing the score of the positive sample triplet and minimizing the score of the negative sample triplet, includes: constructing negative sample triplets corresponding to the positive sample triplets using a negative sampling method oriented towards complex vector space based on the positive sample triplets in the cultural and tourism knowledge graph; calculating the association scores of the positive sample triplets and negative sample triplets based on the ComplEx model architecture; and optimizing the association scores using the logistic loss function to increase the score of the positive sample triplets and decrease the score of the negative sample triplets, so as to obtain entity embedding vectors and relation embedding vectors that satisfy semantic constraints.

[0064] For example, suppose the knowledge graph embedding model described above adopts the ComplEx model architecture to perform low-dimensional vector space mapping on cultural element entities, teaching task element entities, scene element entities, and resource element entities in the cultural tourism knowledge graph. Specifically, the cultural tourism knowledge graph constructed in step 101 can first be represented as a set of triples. Each triple includes a head entity, a relation, and a tail entity. The head and tail entities can be cultural element entities, teaching task element entities, scene element entities, or resource element entities. The relations can be teaching adaptation relations, scene applicability relations, resource support relations, cultural affiliation relations, or semantic association relations, etc. For example, triples such as "Minnan ancient houses - applicable to - architectural culture cognition tasks," "Gaojia opera - adapted to - role experience tasks," "Bo Bing custom - applicable to - festival interaction scenarios," and "short video creation tasks - dependent on - video generation resources" can be constructed. Subsequently, each entity and each relation in the triples are input into the ComplEx model, and each entity and each relation are mapped to a set of complex vector representations. In this embodiment, the so-called complex vector representation means that each entity and relation is represented by both real and imaginary vectors, thus enabling it to represent not only semantic strength but also semantic direction and phase features. Compared with the traditional real vector model, this representation is more suitable for describing the numerous directional and implicit semantic relationships present in cultural and tourism knowledge graphs.

[0065] Furthermore, after initializing the complex vectors of entities and relations, a scoring mechanism based on Hermitian bilinear form can be used to model the relations of each triple. Specifically, the complex vectors of the head entity, relation, and tail entity can be combined to obtain an association score representing the degree of validity of the triple. In this embodiment, a higher association score indicates a stronger semantic match between the head entity, relation, and tail entity, suggesting that the triple is more likely to be a valid knowledge relation in the cultural tourism teaching scenario. Conversely, a lower score indicates that the semantic relation represented by the triple is unreasonable or invalid. For example, for the triple "Gaojia Opera - Adapted to - Role-playing Task," because Gaojia Opera itself has obvious performative and role-playing characteristics, it has a high teaching adaptability to the role-playing task, and the association score calculated by the model is usually high. However, for the triple "Gaojia Opera - Adapted to - Architectural Surveying Task," because the two are less logically related in teaching, its association score is usually low. Using the scoring method described above, the system is able to distinguish between strongly related and weakly related knowledge relationships in the vector space.

[0066] In some implementations, another important reason for adopting the ComplEx model is its ability to effectively represent asymmetric, multi-type, and multi-hop combination relationships in cultural tourism knowledge graphs. Specifically, not all knowledge relationships commonly found in cultural tourism teaching scenarios are symmetrical. For example, the relationship between "Minnan ancient houses - carrying on - clan culture" and "clan culture - carrying on - Minnan ancient houses" are not semantically equivalent. Similarly, the relationship between "cultural origin - influence - cultural inheritance form" and its reverse relationship also shows significant differences. Because the complex phase in the ComplEx model can describe the differences in relationship direction and type, it can more accurately represent the aforementioned asymmetric relationships. Furthermore, for multi-hop chain relationships composed of cultural elements - teaching tasks - scene elements - resource elements, the ComplEx model can also utilize the combination features of different entities and relationships in complex space to perform implicit association reasoning. For example, in the multi-hop chain of "Minnan ancient houses - applicable to - architectural culture cognition task", "architectural culture cognition task - applicable to - classroom presentation scenario", and "classroom presentation scenario - dependent on - courseware presentation resources", the model can not only identify the direct association between Minnan ancient houses and architectural culture cognition task, but also further identify the indirect adaptation relationship between Minnan ancient houses and classroom presentation scenario and courseware presentation resources. This can improve the ability to represent the implicit semantic associations between local cultural elements and the adaptation relationship between local cultural elements and teaching tasks, providing a more accurate knowledge base for the subsequent construction of candidate activity interaction scripts by the teacher-side intelligent agent.

[0067] For example, classroom activities designed for Xiamen's local cultural creativity can incorporate cultural entities such as traditional Minnan houses, Gaojia opera, Nanyin music, and the mooncake gambling custom, along with entities like cultural awareness tasks, role-playing tasks, short video creation tasks, classroom presentation scenarios, and video generation resources, into a cultural tourism knowledge graph. The system first performs complex vector mapping on these entities and relationships, then uses a Hermitian bilinear scoring mechanism to calculate the association scores for each triple. If the model finds that traditional Minnan houses and Gaojia opera have high latent semantic similarity in the complex space, and that Gaojia opera has high compatibility with role-playing tasks and traditional Minnan houses with cultural tour tasks, then it can be further inferred that the two can be combined into a theme of "Traditional Houses and Opera Charm." In this case, this theme can not only be mapped to classroom presentation scenarios but also to short video creation resources and graphic courseware resources, thus forming a basis for cultural theme combinations suitable for classroom teaching activities.

[0068] In some optional implementations, the process of constructing positive and negative sample pairs based on triple relationships in the cultural tourism knowledge graph, and iteratively training the knowledge graph embedding model by maximizing the score of positive sample triples and minimizing the score of negative sample triples, can be achieved using a negative sampling strategy oriented towards complex vector spaces and a logistic loss optimization method. Specifically, existing triples can first be read from the cultural tourism knowledge graph as positive sample triples. For example, "Minnan ancient houses - suitable for - architectural culture cognition tasks," "Gaojia opera - suitable for - role-playing tasks," and "Bo Bing custom - associated with - Mid-Autumn Festival scenes" can all be used as positive sample inputs to the model. Subsequently, for each positive sample triple, the system can use a negative sampling algorithm to generate corresponding negative sample triples. Negative sampling refers to replacing the head or tail entity while keeping the relationship unchanged, thereby constructing semantically unreasonable or weakly related erroneous triples. For example, the task of recognizing Minnan ancient houses as applicable to architectural culture can be replaced with the task of recognizing Minnan ancient houses as applicable to speech cloning, or the task of adapting the mooncake gambling custom to architectural surveying can be used as negative samples. To improve the quality of negative samples, the system preferably adopts a negative sampling method with entity type constraints. That is, when replacing the head or tail entity, the replacement object is selected from the candidate entity set of the same type as the original entity, so as to avoid generating obviously invalid but meaningless negative samples. In this way, the model can learn more about how to distinguish the true adaptation relationship between similar entities.

[0069] After constructing the positive and negative samples, the association scores for each positive and negative sample triple can be calculated based on the ComplEx model. Specifically, the head entity complex vector, relation complex vector, and tail entity conjugate complex vector in the positive sample triple are combined to obtain the positive sample score. Similarly, the same operation is performed on the negative sample triple to obtain the negative sample score. Generally, if the model parameters are set reasonably, genuine positive sample triples should obtain higher association scores, while fake negative sample triples should obtain lower association scores. For example, the Gaojia Opera-adapted-to-role experience task should obtain a higher score, while the Gaojia Opera-adapted-to-architectural surveying task should obtain a lower score. By continuously comparing the score differences between positive and negative samples, the system gradually learns the real cultural semantic relationships and teaching adaptation relationships that exist in the cultural tourism knowledge graph.

[0070] Furthermore, to continuously increase the scores of positive samples and continuously decrease the scores of negative samples, the logistic loss function can be used to optimize the model parameters. Specifically, the system can first perform a probabilistic mapping of the scores of positive and negative samples, then calculate the loss value based on the training objective that positive samples should approach high probability and negative samples should approach low probability, and use the backpropagation algorithm to update the parameters of the entity embedding vector and relation embedding vector. In one embodiment, the training process can use the Adam optimization algorithm, with a learning rate of 3e-4, 5e-4, or 1e-3, an embedding dimension of 128, 200, or 256, a batch size of 256 or 512, each positive sample corresponding to 5 to 10 negative samples, and 100 to 300 training rounds. In a preferred embodiment, the embedding dimension can be set to 256, the learning rate to 5e-4, the batch size to 512, and each positive sample corresponding to 10 negative samples, to balance training efficiency and relation discrimination accuracy. By setting the parameters as described above, the model can learn the continuous cultural semantics in the cultural and tourism knowledge graph well while maintaining the geometric constraints of the complex space.

[0071] In the actual training process, firstly, the complex embedding vectors of each entity and relation are initialized. Secondly, a batch of positive sample triples is read from the cultural tourism knowledge graph, and corresponding negative sample triples are generated based on a negative sampling strategy. Thirdly, the association scores of positive and negative samples are calculated using the ComplEx model, and the loss value of the current batch is obtained according to the logistic loss function. Then, the backpropagation algorithm is used to pass the loss value to each entity vector and relation vector, and the parameters are updated using the optimizer. After that, the above process is repeated until the loss function converges or the preset number of training rounds is reached. Finally, the trained entity embedding vectors and relation embedding vectors are output for subsequent calculation of semantic similarity and association weights between cultural elements. Through this iterative training process, the model can gradually enhance its ability to identify the real fit between local cultural elements and teaching tasks, and reduce the probability of misjudging semantically similar but misfit relationships.

[0072] In a specific example, if a classroom theme is the multimodal expression of Minnan culture, relationships such as Nanyin music suitable for audio-visual appreciation tasks, Minnan ancient houses suitable for route guidance tasks, and the Bo Bing custom suitable for interactive experience tasks can be input into the ComplEx model as positive sample triples. Simultaneously, negative sample triples can be constructed, such as Nanyin music suitable for architectural surveying tasks, Minnan ancient houses suitable for speech cloning tasks, and the Bo Bing custom suitable for spatial modeling tasks. After multiple rounds of training, the ComplEx model can achieve higher association scores for the former type of positive samples and lower association scores for the latter type of negative samples. Therefore, in subsequent applications, it can more accurately determine whether Nanyin music is more suitable for audio-visual appreciation teaching scripts, whether Minnan ancient houses are more suitable for route guidance or architectural culture cognition scripts, and whether the Bo Bing custom is more suitable for interactive experience scripts. Further combining the adaptation relationships of scene elements and resource elements, the teacher-side agent can then generate candidate activity interaction scripts that better meet the current classroom needs.

[0073] Through the aforementioned ComplEx mapping and iterative training process, not only can entities and relations in the cultural tourism knowledge graph be mapped into computable complex vector representations, but entity embedding vectors and relation embedding vectors that satisfy semantic constraints can also be obtained through positive and negative sample comparison learning. Based on these vectorized representations, the semantic similarity between cultural element entities and the association weights between cultural element entities and teaching task element entities, scene element entities, and resource element entities can be further calculated. This allows for the identification of potential semantic relationships between local cultural elements and the adaptation relationships between local cultural elements and teaching tasks, scenes, and resources, which can then serve as the data-driven basis for the teacher-side intelligent agent to construct candidate activity interaction scripts.

[0074] Optionally, based on the calculated semantic similarity and association weights, latent semantic associations between local cultural elements and the matching degree between local cultural elements and teaching tasks are extracted, providing a data-driven decision-making basis for the teacher-side agent to generate candidate activity interaction scripts that match the current teaching scenario. Specifically, the semantic similarity between cultural element entities can be calculated based on the amplitude and phase information of the entity embedding vectors. Based on the distribution characteristics of the entity embedding vectors and relation embedding vectors in the complex vector space, the association weights between cultural element entities and teaching task element entities, scene element entities, and resource element entities are calculated. According to the semantic similarity and the association weights, the latent semantic associations between local cultural elements and the adaptation relationships between local cultural elements and teaching task elements, scene elements, and resource elements are identified. The latent semantic associations and adaptation relationships are used as the association basis for the teacher-side agent to construct candidate activity interaction scripts.

[0075] Step 103: The teacher-side intelligent agent constructs candidate activity interaction scripts adapted to the current teaching scenario based on the correlation between cultural elements and teaching tasks in the vectorized representation, and performs teaching constraints, strategy corrections and / or manual interventions on the candidate activity interaction scripts.

[0076] The teacher-side intelligent agent refers to an intelligent decision-making unit deployed on the teacher's terminal and / or server, used for activity arrangement, teaching constraint control, script modification, and process monitoring for the teaching side. Based on the vectorized representation of the cultural tourism knowledge graph, the teacher-side intelligent agent can automatically understand the relationships between local cultural elements, teaching task elements, activity scene elements, and resource elements. It can also generate, adjust, or filter candidate activity interaction scripts suitable for the current teaching scenario, combining the current course objectives, classroom stage, student profiles, and teaching standard requirements. In other words, the teacher-side intelligent agent is equivalent to a teaching organization and scheduling assistant for the teacher. It can automatically complete tasks such as selecting cultural themes, arranging task nodes, and configuring resource call paths. It can also respond to the teacher's input teaching preferences, executing teaching constraints, strategy modifications, and manual intervention on candidate activity interaction scripts. For example, in a Minnan cultural creative planning course, the teacher-side intelligent agent can prioritize selecting cultural elements such as Minnan ancient houses, Gaojia opera, and Nanyin music from the cultural tourism knowledge graph based on the teaching objectives of local cultural cognition and creative expression training, and automatically generate candidate activity interaction scripts containing cultural explanation nodes, interactive experience nodes, and content creation nodes.

[0077] In this embodiment, the candidate activity interaction script refers to a structured script constructed by the teacher-side intelligent agent based on the relationship between cultural elements and teaching tasks, used to describe the complete execution process of a cultural tourism teaching activity. The candidate activity interaction script is used to represent the activity organization logic, task arrangement logic, interaction triggering logic, and content output logic. For example, the candidate activity interaction script may include a main theme, node tasks, interaction triggering conditions, task execution order, resource call rules, and content output positions, used to describe the complete execution process of a cultural tourism teaching activity in a classroom or training scenario.

[0078] The main theme describes the core cultural theme around which the current teaching activity revolves; the node tasks define the specific tasks of each teaching segment; the interaction trigger conditions define when to switch nodes, when to trigger interaction, or when to adjust the difficulty; the task execution order defines the sequential relationship between nodes; the resource call rules define what kind of graphics, videos, explanatory materials, or AIGC-generated resources to call at different nodes; and the content output positions define what kind of teaching materials, display pages, or creative content should be output at different nodes. For example, in a "Traditional Houses and Opera Charm" themed class, the candidate activity interaction script can set "Recognizing Traditional House Architecture" as the first node task, "Experiencing Gaojia Opera Roles" as the second node task, and "Creating Short Videos on the Theme of Traditional Houses" as the third node task, and further stipulate that when students' accuracy rate in answering questions at the previous node reaches a preset threshold, the next interactive node will be automatically triggered.

[0079] Specifically, in step 103, the teacher-side intelligent agent can select a set of core cultural elements that match the current teaching objectives, student characteristics, and activity scenarios based on the correlation between cultural elements learned from knowledge representation and teaching tasks, and generate multiple candidate cultural themes in conjunction with the requirements of the teaching tasks. Further, the teacher-side intelligent agent can arrange the activity flow, design interactive nodes, and configure resource combinations for the candidate cultural themes, based on the layout of the activity venue, activity route, activity nodes, and classroom tasks, forming multiple candidate activity interaction scripts. Afterwards, the teacher-side intelligent agent can also screen, modify, or manually adjust the candidate activity interaction scripts according to preset teaching constraint rules, which may include constraints on teaching time, content difficulty, safety, cultural suitability, and knowledge point coverage. Through this step, candidate activity interaction scripts that both meet classroom teaching standards and take into account local cultural expression and interactive experience can be generated.

[0080] As an optional embodiment, in step 103, the teacher-side intelligent agent constructs candidate activity interaction scripts adapted to the current teaching scenario based on the correlation between cultural elements and teaching tasks in the vectorized representation, including: Based on the entity embedding vectors and relation embedding vectors in the vectorized representation, the association weights between cultural element entities and teaching task element entities are calculated, along with the semantic similarity between cultural element entities. A set of core cultural elements matching the current teaching scenario is determined. Based on this set of core cultural elements, combined with activity scenario information, teaching objective information, and student characteristic information, multiple candidate activity interaction paths are generated using a knowledge graph-based path planning algorithm and a deep search algorithm. Each candidate activity interaction path includes the cultural content presentation order, task triggering conditions, interaction node settings, and content output strategies. Feasibility assessments and teaching effect predictions are performed on these multiple candidate activity interaction paths, and a set of candidate activity interaction scripts that meet the teaching constraints is selected. These candidate activity interaction scripts contain complete definitions of activity organization logic, task arrangement logic, interaction triggering logic, and content output logic.

[0081] Specifically, in step 103, the process of the teacher-side agent constructing candidate activity interaction scripts adapted to the current teaching scenario based on the association between cultural elements and teaching tasks in vectorized representations can be divided into a knowledge graph path planning stage, a deep generation optimization stage, a multi-objective optimization stage, and a context-aware refinement stage. In the above embodiment, the teacher-side agent can first calculate the association weight between cultural element entities and teaching task element entities based on entity embedding vectors and relation embedding vectors, and combine the semantic similarity between cultural element entities to select a set of core cultural elements that match the current teaching scenario, teaching objectives, and student characteristics. The set of core cultural elements can be a combination of several cultural entities with high cultural relevance and teaching adaptability. For example, in the Minnan culture teaching scenario, Minnan ancient houses, Gaojia opera, Nanyin music, and the Bo Bing custom can be identified as the set of core cultural elements.

[0082] During the knowledge graph path planning phase, the teacher-side agent can plan multiple semantically coherent cultural content presentation paths based on the entity relationships in the cultural tourism knowledge graph. Starting with the core cultural element entities and ending with the teaching objective element entities, the agent can employ heuristic search algorithms or depth-first search algorithms. Specifically, the teacher-side agent can first read the set of core cultural element entities, the set of teaching objective element entities, and the associated teaching task element entities, scene element entities, and resource element entities from the cultural tourism knowledge graph corresponding to the current teaching scenario. These entities and their relationships are then constructed into a searchable directed graph structure. Nodes in the graph represent entities such as cultural elements, teaching tasks, scenes, and resources, while edges represent teaching adaptation relationships, scene applicability relationships, resource support relationships, or semantic relationships between entities.

[0083] Furthermore, the teacher-side agent can perform constraint modeling on the edges and nodes in the directed graph before path planning. Specifically, the semantic similarity between cultural elements, the association weight between cultural elements and teaching tasks, and the matching degree between teaching objectives and task nodes can be comprehensively determined as the evaluation criteria for path search. Simultaneously, factors such as task difficulty, class duration, scene capacity, resource availability, and knowledge point coverage requirements are used as path selection constraints. In other words, during the path search process, it is necessary not only to ensure that the generated path gradually transitions from core cultural elements to teaching objective elements, but also to ensure that the sequence of nodes traversed is logically sound in teaching, culturally coherent, and feasible for classroom implementation.

[0084] In one implementation, when using a heuristic search algorithm for path planning, the teacher-side agent first adds the core cultural element entity as the starting node to the set of nodes to be expanded. Heuristic evaluation values ​​are then constructed based on the semantic distance, task suitability, and scenario applicability between each candidate node and the target teaching element. Subsequently, in each search iteration, the teacher-side agent prioritizes expanding the node with the best overall evaluation value and adds downstream nodes directly related to the current node and meeting the constraints to the path candidate set. Through continuous iterative expansion, the system can gradually generate multiple candidate paths from the core cultural element entity to the teaching target element entity. During this process, nodes with a closer semantic distance to the target teaching element, a higher suitability for the current teaching task, and easier-to-satisfy resource constraints can be given higher priority, thereby improving path planning efficiency and path quality.

[0085] In another implementation, when using a depth-first search algorithm for path planning, the teacher-side agent can start from a core cultural element entity and expand layer by layer along teaching task nodes or culturally related nodes with high relevance to it until it finds the endpoint node connected to the current teaching objective element entity. If a search branch encounters duplicate nodes, exceeds the path length limit, exceeds the preset threshold for teaching difficulty, or fails to meet resource conditions, the teacher-side agent can backtrack on the current branch and continue searching on other candidate branches. In this way, the system can obtain multiple candidate paths with different cultural organization methods and task arrangement logics based on a complete exploration of different path branches, thus providing more alternative solutions for the generation of subsequent candidate activity interaction scripts.

[0086] Furthermore, during path expansion, the teacher-side agent can set path length thresholds, node type ratio constraints, and task connection constraints for each path. The path length threshold limits the upper limit of the number of nodes in a single cultural content presentation path to avoid excessively long paths leading to redundant classroom execution. The node type ratio constraint balances the distribution of cultural explanation nodes, interactive experience nodes, and content creation nodes to improve the rhythmic rationality of classroom activities. The task connection constraint ensures a reasonable knowledge progression and cognitive transition between nodes. For example, in a cultural content presentation path starting with a traditional Minnan house, the path could first pass through an architectural culture recognition task node, then transition to a Gaojia opera role-playing node, and finally extend to a short video creation task node, rather than directly jumping from the architectural culture recognition node to a high-difficulty content generation node. These constraints effectively prevent logical breaks, sudden changes in difficulty, or imbalances in classroom rhythm in the path planning results.

[0087] After generating multiple candidate paths, the teacher-side agent can further sort and filter these paths. Specifically, each candidate path can be comprehensively scored based on indicators such as semantic coherence, matching degree of teaching objectives, resource feasibility, node complexity, and estimated participation, and several paths with higher scores can be retained as candidate interaction paths for subsequent activities. Furthermore, an initial sequence of interactive nodes and task triggering thresholds can be assigned to each retained path to form a basic path framework. For example, if a path includes three main nodes: recognition of ancient buildings, interactive experience of Gaojia opera, and creation of Minnan culture, the system can further assign it the required accuracy rate, dwell time, or interaction completion threshold for entering the next node, and define the basic content output format of each node, thereby providing structured input for the subsequent deep generative model optimization stage.

[0088] For example, in a Xiamen local cultural and creative planning class, if the teaching objective is to complete a short video planning task on a local cultural theme, the teacher's AI agent can use the Minnan ancient houses as the starting node and the short video creation task as the ending node, and search for intermediate nodes related to both in the cultural tourism knowledge graph. After path planning, the system can generate multiple cultural content presentation paths. For example, the first path is Minnan ancient houses, architectural culture recognition task, Gaojia opera experience task, and short video creation task; the second path is Minnan ancient houses, local folk custom explanation task, mooncake gambling interaction task, and short video creation task; and the third path is Minnan ancient houses, Nanyin appreciation task, cultural expression training task, and short video creation task. Afterwards, the teacher's AI agent can select the most suitable path based on the current class time, student interests, and available resources as the foundation for constructing subsequent candidate activity interaction scripts.

[0089] Through the aforementioned knowledge graph path planning process, the teacher-side intelligent agent can transform the entity relationships in the cultural tourism knowledge graph into multiple executable cultural content presentation paths, establishing a clear organizational logic and progressive relationship between cultural elements, teaching tasks, scene resources, and teaching objectives. This provides a structured, filterable, and optimizable path foundation for the generation of subsequent candidate activity interaction scripts.

[0090] During path planning, the semantic similarity between cultural elements can be used as edge weights, and the difficulty coefficient of teaching task elements can be used as node constraints. This ensures that the generated paths maintain semantic coherence between cultural content while meeting the cognitive level requirements of classroom teaching. For example, for a teaching objective that starts with traditional Minnan houses and ends with a short video creation task, the system can plan multiple different cultural content presentation paths, such as recognizing traditional houses, associating with opera elements, connecting local products, and generating creative scripts. Furthermore, the teacher-side agent can assign an initial sequence of interactive nodes and task triggering thresholds to each planned path, forming a basic path framework for subsequent optimization of node tasks, interaction depth, and output content.

[0091] In the deep generation optimization stage, the teacher-side agent can input the basic paths planned from the aforementioned knowledge graph into a pre-trained sequence generation model or an attention-based generation model to further refine and expand the basic paths. Specifically, the model input can include cultural element sequence vectors, student feature embedding vectors, and scene constraint parameter vectors. The cultural element sequence vectors represent the order of cultural content in the current path, the student feature embedding vectors represent the student group's interests, cognitive levels, and participation tendencies, and the scene constraint parameter vectors represent class duration, venue conditions, equipment availability, and teaching resource constraints. During the generation process, the model can dynamically monitor the matching degree between student features and cultural content based on an attention mechanism, thereby automatically generating more personalized task triggering conditions, interaction node parameters, and content output strategies for each basic path. For example, for students who prefer interactive experiences, role-playing nodes can be added to the original path, and the continuous length of text explanation nodes can be reduced. For students who prefer creative expression, the system can add short video script generation and visual material creation nodes.

[0092] Furthermore, the teacher-side agent can employ a candidate retention search strategy to generate multiple candidate path variants, allowing each variant to maintain consistency in core cultural logic while exhibiting different interaction depths, task sequences, and difficulty gradients. In specific implementation, the teacher-side agent can first extract the core cultural thread from one or more basic cultural content presentation paths obtained during the aforementioned knowledge graph path planning phase, using it as a fixed framework. This core cultural thread typically includes core cultural element nodes, key teaching task nodes, and endpoint nodes directly related to teaching objectives. This content ensures consistency in cultural theme expression and teaching objective orientation across different path variants.

[0093] Based on this, the teacher-side agent can use non-core nodes, connection methods between nodes, task triggering conditions, and interaction intensity parameters in the basic path as adjustable variables, and perform candidate retention search around the core cultural theme. Specifically, in each round of search, the teacher-side agent can start from the current path state and expand the adjustable variables to generate several candidate paths with local changes. For example, it can generate three types of path variants for the same core cultural theme: explanation-priority, interaction-priority, and creation-priority; it can also swap the execution order of two adjacent task nodes while keeping the theme unchanged, or configure different task difficulties and interaction triggering thresholds for the same node. For each candidate path obtained from expansion, the teacher-side agent can calculate its comprehensive score in dimensions such as teaching objective matching degree, cultural coherence, resource feasibility, student suitability, and execution complexity, and retain several paths with higher scores for the next round of expansion to avoid an excessively large search space and improve path generation efficiency.

[0094] Furthermore, to ensure that multiple candidate path variants share a consistent cultural theme while possessing sufficient differentiation, the teacher-side agent can set variant generation rules. These rules can include node replacement rules, task rearrangement rules, difficulty adjustment rules, and interaction depth adjustment rules. Node replacement rules introduce alternative nodes with high semantic similarity to the current node into the path; for example, replacing a local product explanation node with a folk tale explanation node, thus creating variants with different content expression methods but a consistent theme. Task rearrangement rules change the order of some nodes while satisfying knowledge progression constraints; for example, moving the Gaojia opera interactive experience node before the architectural culture recognition node, thus creating an experience-based path. Difficulty adjustment rules stratify the complexity of node tasks based on the students' cognitive level; for example, setting the same cultural creation task into three versions: basic fill-in-the-blank expression, intermediate imitation creation, and advanced open-ended planning. Interaction depth adjustment rules control the number, trigger frequency, and feedback intensity of interactive nodes in each path, thereby creating different path variants such as light interaction, medium interaction, and deep interaction.

[0095] In one implementation, the teacher-side agent can generate path variants using a hierarchical candidate retention approach. Specifically, in the first layer, the core cultural theme is fixed, with only limited adjustments to the order of task nodes, retaining several candidate paths that meet the constraints of the teaching objectives. In the second layer, the number of interactive nodes and triggering conditions are further adjusted on the retained paths to form candidate variants with different levels of interaction. In the third layer, the task difficulty gradient of each candidate variant is adjusted to form path versions suitable for different student groups. Through this hierarchical search method, the system can gradually control variables and generate differentiated paths step by step, thereby avoiding unstable path quality caused by adjusting too many parameters simultaneously in a single round of search.

[0096] To ensure the comparability and selectability of candidate path variants, the teacher-side agent can generate corresponding evaluation tags for each variant. These tags can include information such as variant type, applicable student group, estimated execution time, estimated participation, cognitive load level, resource consumption level, and coverage of teaching objectives. For example, one path variant might be labeled as interaction-first, suitable for high-participation classes, with an execution time of 35 minutes, moderate cognitive load, and low resource requirements; another variant might be labeled as creation-first, suitable for older students, with an execution time of 45 minutes, high cognitive load, and moderate resource requirements. The teacher-side agent can then classify and further filter multiple candidate path variants based on these evaluation tags.

[0097] For example, in a Xiamen local cultural and creative planning class, if the basic path consists of Minnan ancient houses, architectural culture recognition tasks, Gaojia opera interactive experience tasks, and short video creation tasks, the teacher-side AI agent can generate multiple candidate path variations around this main theme. The first variation maintains the original order but reduces the frequency of interaction triggers, forming a lecture-enhanced path. The second variation brings forward the Gaojia opera interactive experience task, forming an experience-introduction path. The third variation adds a folk story explanation node after the Gaojia opera interactive experience task and appropriately reduces the difficulty of the short video creation task, forming a cultural extension path. The fourth variation retains the original nodes but breaks down the short video creation task into two sub-tasks: storyboard design and scriptwriting, forming a deep creation path. All these variations revolve around Minnan ancient houses and Gaojia opera in their core cultural logic, but they differ significantly in interaction depth, task order, and difficulty gradient, thus providing a richer candidate base for subsequent teacher selection and dynamic decision-making by the deep reinforcement learning AI agent.

[0098] Through the above implementation methods, the teacher-side intelligent agent can systematically generate multiple candidate path variants with different structures and focuses, while ensuring consistency between the core cultural logic and teaching objectives. This improves the diversity, adaptability, and selectivity of candidate activity interaction scripts and provides more strategic space for the dynamic optimization of subsequent target activity interaction schemes.

[0099] In the multi-objective optimization and diversity assurance stage, the teacher-side agent can comprehensively evaluate and optimize multiple generated candidate paths. Specifically, a multi-objective optimization function can be constructed to jointly evaluate multiple dimensions such as teaching effectiveness, student participation, time efficiency, and cultural depth. Teaching effectiveness measures the degree to which candidate paths support teaching objectives; student participation measures the effect of interactive design within the path on promoting active student participation; time efficiency measures the rationality of the path's execution within a given class time; and cultural depth measures the hierarchy and representativeness of the cultural content involved. To increase the diversity and selection space among candidate paths, cross-variant combination adjustments can be performed, such as swapping the order of some nodes, replacing interactive task types, or adjusting content output methods, to generate more candidate paths with significant differences. Furthermore, a minimum difference threshold can be set to detect differences in key nodes, task order, and interaction methods among multiple candidate paths, ensuring that the final retained candidate paths have sufficient differentiation in core node settings, thereby providing ample alternatives for subsequent teaching selection and dynamic decision-making.

[0100] In the context-aware path refinement stage, the teacher-side agent can further combine real-time activity scenario information and student profiles to refine the generated candidate paths for scenario adaptation and execution parameters. Specifically, the real-time activity scenario information may include venue capacity, equipment availability, network status, environmental noise level, and on-site resource layout; the student profile may include age distribution, cognitive level, cultural background, major, interests, and historical participation behavior. Based on the above contextual information, the teacher-side agent can adaptively adjust the complexity of interactive nodes, task presentation methods, resource call order, and prompting strategies. For example, when venue capacity is small and equipment resources are limited, nodes that originally relied on large-screen interaction can be adjusted to mobile terminal answering nodes; when the proportion of lower-grade students in the student group is high, the cognitive difficulty of some cultural explanation nodes can be reduced, and more visual and contextualized presentation methods can be added. Furthermore, the teacher-side agent can generate a detailed execution parameter set for each candidate path, which may include operational details such as time nodes, resource requirements, personnel configuration, and emergency plans to improve the actual executability of the candidate paths.

[0101] In one optional example, if the current teaching scenario is a Xiamen local cultural creative planning course, and the teaching objective is to guide students to understand the connections between local cultural elements and complete creative communication tasks, the teacher-side AI agent can first determine Minnan ancient houses, Gaojia opera, and Nanyin music as the core cultural element set based on the knowledge representation learning results. Subsequently, in the path planning stage, multiple candidate activity interaction paths are generated. For example, the first path involves recognizing ancient house architecture, interactive Gaojia opera experience, Nanyin culture explanation, and short video script creation; the second path involves Nanyin appreciation introduction, ancient house culture tour, Gaojia opera character association, and poster copy generation; and the third path involves interactive activities related to the mooncake gambling custom, explanation of local products, reconstruction of ancient house scenes, and creative short film generation. Afterwards, in the deep generation optimization stage, the system, combined with student preferences and classroom resource conditions, refines and adjusts the task trigger thresholds, interaction node order, and content output strategies in the above paths. Furthermore, a multi-objective optimization process is used to select candidate paths that align with class duration while also considering participation and cultural depth. During the context-aware refinement stage, each path is supplemented with specific resource call sequences, task execution durations, and emergency alternative nodes. Finally, the teacher-side agent organizes the selected and refined candidate activity interaction paths into a set of candidate activity interaction scripts. Each candidate activity interaction script contains a complete definition of activity organization logic, task arrangement logic, interaction triggering logic, and content output logic, thus providing an executable candidate foundation for subsequent strategy decisions by the deep reinforcement learning agent.

[0102] Through the above implementation method, the teacher-side intelligent agent can transform the semantic association results in the knowledge graph into multiple candidate activity interaction paths with cultural coherence, pedagogical rationality, and scenario-based feasibility, and further transform these candidate activity interaction paths into a complete set of candidate activity interaction scripts. Compared to methods that rely solely on human experience to arrange classroom activity processes, this implementation method can improve the automation level of activity script generation, the accuracy of content matching, and the diversity of candidate solutions, thereby providing a higher-quality input foundation for the dynamic generation and optimization of subsequent target activity interaction solutions.

[0103] As an optional embodiment, step 103, which involves implementing teaching constraints, strategy corrections, and / or manual interventions on the candidate activity interaction scripts, includes: verifying the compliance of the candidate activity interaction scripts according to preset teaching constraint rules, which include constraints on teaching duration, content difficulty, student safety, cultural suitability, and knowledge point coverage; triggering a strategy correction mechanism when the candidate activity interaction scripts are detected to violate the teaching constraint rules, readjusting the order of cultural content, task difficulty gradient, and interaction trigger threshold in the activity organization logic based on association weights and semantic similarity; and / or pushing a manual intervention interface to users so that teachers can manually adjust the candidate activity interaction scripts through their teacher terminals and feed the manual intervention results back to the knowledge graph embedding model for parameter fine-tuning.

[0104] For example, in a Minnan ancient house culture experience class, the candidate activity interaction scripts generated by the teacher's intelligent agent include nodes for recognizing ancient house architecture, interactive experience nodes for Gaojia opera, explanation nodes for local products, and short video creation nodes. If the system detects that the total execution time of the candidate activity interaction script exceeds the preset class time, or that the task difficulty of the Gaojia opera interactive experience node exceeds the cognitive level of the current student group, a strategy correction mechanism is triggered. The teacher's intelligent agent can adjust the order of cultural content based on the correlation weight between cultural elements and teaching tasks, as well as the semantic similarity between related cultural nodes, to reduce the task difficulty gradient and increase the trigger threshold of some interaction nodes. Furthermore, teachers can also manually intervene in the candidate activity interaction scripts through the teacher's terminal, such as deleting the explanation node for local products, replacing the Gaojia opera interactive experience node with a task for recognizing opera elements, or adjusting the short video creation node to the class summary stage, so that the candidate activity interaction scripts better meet the actual teaching needs.

[0105] Manual adjustments include, but are not limited to, modifying interactive node parameters, replacing cultural content resources, adding or removing task steps, and adjusting the timing of content output. For example, in a Minnan ancient house culture experience class, teachers can manually adjust the interactive scripts of candidate activities based on the current teaching progress and student participation. Specifically, teachers can adjust the trigger threshold of the Gaojia Opera interactive experience node from a high participation requirement to one that can be triggered with basic participation, thereby increasing the chances of more students entering subsequent tasks. The original local product graphic and textual explanation resources can be replaced with short video demonstration materials or contextualized image resources to enhance classroom appeal. Depending on the remaining class time, extended regional product explanation segments can be removed, or a knowledge point summary task can be added to ensure the completion of core teaching objectives. The short video creation node can also be moved from the middle of the class to the class summary stage, allowing the preceding cultural cognition and interactive experience content to lay the groundwork before entering the content creation stage. Through these manual adjustments, the interactive scripts of candidate activities can be made more closely aligned with the actual classroom environment, student status, and teacher's teaching arrangements.

[0106] Optionally, historical data on strategy corrections and manual interventions can be recorded to construct a knowledge base for optimizing teaching strategies. This knowledge base can then be used for experience-based guidance and dynamic updates of constraint rules during the generation of subsequent candidate activity interaction scripts. For example, during the execution of multiple local culture classroom activities, the following experience data can be gradually accumulated: For lower-grade students, continuous explanation nodes should be appropriately reduced and visual interactive tasks should be increased. For creative planning courses, content creation and result presentation nodes should be prioritized. For scenarios with insufficient equipment resources, task steps relying on terminal collaboration should be automatically reduced. Based on the above experience data, the teacher-side intelligent agent can automatically introduce corresponding constraints and optimization strategies when generating candidate activity interaction scripts, thereby reducing repetitive manual adjustments and improving script generation efficiency and classroom adaptability.

[0107] Step 104: The student-side intelligent agent collects real-time status information during the interaction process based on the candidate activity interaction script to characterize user participation, teaching execution, and / or scene changes.

[0108] The student-side intelligent agent refers to an intelligent interactive unit deployed on student terminals and / or servers for performing interactive guidance, task prompts, behavior collection, and status feedback to students. The student-side intelligent agent can push activity guidance information, learning prompts, interactive tasks, and feedback entry points to student terminals in real time according to the target activity interaction plan, and collect real-time status information to characterize user participation, teaching execution, and / or scene changes during student participation.

[0109] In practical applications, the student-side agent can be implemented as an interactive execution feedback agent facing the student. On one hand, it is responsible for presenting the activity plans generated by the teacher-side agent and the deep reinforcement learning agent to the students in a user-friendly interactive format. On the other hand, it is responsible for transmitting data such as students' click behavior, dwell time, answer status, interaction frequency, task completion status, and emotional feedback back to the system. For example, in the "Minnan Ancient House Culture Experience" classroom activity, the student-side agent can display prompts on the student's terminal such as entering the ancient house building identification task, completing the Gaojia opera character matching question and answer, and submitting a script explaining local products. It can also record the student's answering time, task completion progress, and changes in interest preferences in real time during the student's operation.

[0110] Specifically, in step 104, the student-side intelligent agent can interact with students through student terminals and collect multi-dimensional state information in real time during the interaction. This real-time state information may include user engagement information, task completion status information, classroom interaction status information, environmental context information, and information on the achievement of learning objectives. For example, user engagement information may include dwell time, click behavior, answering behavior, voice interaction behavior, or task response frequency. The environmental context information may include the current classroom stage, activity node location, time progress, device status, and changes in the on-site environment. The student-side intelligent agent can perform feature extraction, temporal alignment, and fusion processing on the collected data to form a real-time state information matrix that can be used by the deep reinforcement learning agent. Through this step, the real-time feedback and scene changes of students during the activity interaction process can be quantified, providing state input for subsequent dynamic strategy decisions.

[0111] As an optional embodiment, in step 104, the student-side intelligent agent collects real-time status information to characterize user participation, teaching execution, and / or scene changes during the interaction process based on the candidate activity interaction script. This includes: collecting user behavior data through the student terminal's behavior perception module, the user behavior data including location coordinate information, gaze focus trajectory, gesture operation sequence, voice interaction content, physiological state parameters, and task completion progress; performing feature extraction and temporal alignment on the collected user behavior data based on the interaction node definitions in the candidate activity interaction script to generate a multi-dimensional feature vector containing user participation depth, cognitive load level, and emotional state; and monitoring the current state of cultural and tourism activities through the student terminal's environment perception module. The physical environment parameters of the field are analyzed, including pedestrian density, noise level, light intensity, temperature and humidity variations, and spatial layout. The multi-dimensional feature vector is fused with these physical environment parameters, and combined with the expected execution path of the target activity interaction scheme, a weighted multi-dimensional dynamic time warping algorithm is used to calculate the deviation index between the actual execution state and the expected state. The weighting coefficients used in the calculation are dynamically allocated based on the priority of teaching objectives, the importance of cultural elements, and safety constraints. A real-time state information matrix, including user participation scores, task execution efficiency scores, environmental adaptability scores, and teaching objective achievement scores, is dynamically constructed based on the deviation index, serving as the state input for the deep reinforcement learning agent to make policy decisions. Furthermore, the weighting coefficients are used to quantify the comprehensive deviation between the actual state vector and the expected state vector in terms of temporal and feature dimensions.

[0112] For example, in a Minnan ancient house culture experience class, the student-side intelligent agent can collect user behavior data such as position changes, dwell time, click frequency, answering progress, voice response content, and facial expression changes at the ancient house building recognition node and the Gaojia opera character matching node through the student terminal's behavior perception module. Simultaneously, it can acquire physical environment parameters such as the current classroom traffic density, noise level, lighting conditions, and equipment operating status through the environment perception module. Subsequently, the student-side intelligent agent can perform feature extraction, temporal alignment, and fusion processing on the collected data according to the preset node execution order and target completion path in the candidate activity interaction script to generate a multi-dimensional feature vector reflecting the student's depth of participation, cognitive load level, and emotional state. Furthermore, the system can calculate user participation scores, task execution efficiency scores, environmental adaptability scores, and teaching goal achievement scores by combining the degree of deviation between the expected execution path and the actual execution state. For example, when students spend a short time at the ancient house building recognition node, have a high error rate in answering questions, and experience high noise levels, the system can determine that the current task execution efficiency and environmental adaptability are low. When students participate frequently, provide positive emotional feedback, and complete tasks quickly in the Gaojia Opera interactive nodes, the system can determine that their user engagement score is high. The real-time state information matrix constructed based on these score results can serve as the state input for the deep reinforcement learning agent to make subsequent policy decisions.

[0113] Step 105: The deep reinforcement learning agent, in conjunction with the candidate activity interaction script, makes a policy decision on the real-time state information to obtain the target activity interaction scheme.

[0114] The deep reinforcement learning agent refers to an intelligent decision-making unit that combines the representational capabilities of deep neural networks with the strategy optimization capabilities of reinforcement learning, and is used to dynamically decide on activity interaction schemes based on real-time state information. The deep reinforcement learning agent can use candidate activity interaction scripts as policy constraints, take the real-time state information collected by the student-side agent as the current environment state input, and select the optimal action in the current classroom context through a policy network, thereby generating or adjusting the target activity interaction scheme. Furthermore, the deep reinforcement learning agent can construct a reward function based on objectives such as teaching effectiveness, user participation, compliance with teaching constraints, and execution efficiency, and continuously optimize policy parameters through continuous interactive feedback, enabling the activity flow, task sequence, trigger thresholds, and content output methods to adapt to the needs of different teaching stages and different student groups. In other words, the deep reinforcement learning agent is equivalent to the dynamic scheduling core of the entire system, capable of adaptively optimizing activity details without changing the overall teaching objectives.

[0115] The target activity interaction scheme refers to the specific activity execution plan suitable for the current teaching scenario, ultimately determined by the deep reinforcement learning agent after comprehensively considering candidate activity interaction scripts and real-time status information. The target activity interaction scheme typically includes the target cultural theme, task node sequence, interaction triggering method, task difficulty configuration, resource call path, content push rhythm, and output format. The difference between it and candidate activity interaction scripts is that candidate activity interaction scripts are multiple pre-generated options by the teacher-side agent, while the target activity interaction scheme is the final execution plan formed by the deep reinforcement learning agent selecting, adjusting, or combining candidate schemes based on real-time interaction status. For example, in a Minnan ancient house culture experience class, the teacher-side agent initially generated three candidate activity interaction scripts, corresponding to three organizational paths: architectural cognition priority, opera experience priority, and creative expression priority, respectively. When the student-side AI agent reports that students in the current class have a high level of participation in the interactive performance but spend less time on the architectural knowledge explanation, the deep reinforcement learning AI agent can dynamically determine the target activity interaction plan that prioritizes the opera experience, advance the Gaojia opera interactive experience node, reduce the text density of the architectural style explanation node, and add the ancient house element short video creation task in the later stage to improve the overall teaching participation and teaching effect.

[0116] As an optional embodiment, in step 105, the step of the deep reinforcement learning agent making a policy decision on the real-time state information in conjunction with the candidate activity interaction script to obtain the target activity interaction scheme includes: Each candidate script in the candidate activity interaction script set is encoded as a policy constraint vector. The policy constraint vector includes a structured representation of activity organization logic constraints, task orchestration logic constraints, interaction triggering logic constraints, and content output logic constraints. A deep reinforcement learning policy network is constructed, which adopts a two-stream network architecture based on an attention mechanism. The first stream processes the real-time state information matrix, and the second stream processes the policy constraint vector. A cross-stream attention module is used to dynamically fuse state features and constraints. A multi-objective reward function is defined, which includes a teaching effectiveness reward, a user participation reward, a constraint compliance reward, and an execution efficiency reward. The constraint compliance reward is quantitatively evaluated based on the teaching constraint rules in the candidate activity interaction scripts. The deep reinforcement learning policy network is trained online using a proximal policy optimization algorithm. The real-time state information matrix is ​​used as the current state input, and the policy constraint vector is used as the action space constraint. The optimal activity interaction action in the current state is output. Based on the optimal activity interaction action, combined with the activity organization logic and task orchestration logic in the candidate activity interaction scripts, the target activity interaction scheme is dynamically generated.

[0117] In the above optional embodiments, step 105, the process of the deep reinforcement learning agent making policy decisions based on real-time state information in conjunction with candidate activity interaction scripts, can be divided into script encoding stage, state and constraint fusion stage, reward evaluation stage, policy training stage, and scheme generation stage.

[0118] Specifically, each candidate script in the candidate activity interaction script set is first structurally parsed to extract its activity organization logic, task orchestration logic, interaction triggering logic, and content output logic. This logical information is then converted into a policy constraint vector of uniform length. The policy constraint vector can be represented as a fixed-length multi-dimensional feature vector, which may include sub-vectors such as topic mainline encoding, node order encoding, task type encoding, trigger condition parameter encoding, resource call encoding, and content output rule encoding. Each sub-vector can be further uniformly mapped using one-hot encoding, sequential position encoding, rule parameter normalization encoding, or embedding representation, and then concatenated in a preset order to form a fixed-dimensional constraint representation. In this way, the originally heterogeneous candidate activity interaction scripts can be converted into standardized inputs that can be processed by a deep reinforcement learning agent. This allows the agent to characterize the differences between different candidate scripts in terms of topic mainline, node order, trigger conditions, resource call methods, and output rules, enabling the deep reinforcement learning agent to compare and select multiple candidate scripts within a unified decision-making framework.

[0119] Furthermore, a deep reinforcement learning policy network can be constructed to jointly model the real-time state information matrix and policy constraint vectors. Specifically, the policy network can adopt a hierarchical structure consisting of a state encoding layer, a constraint encoding layer, a feature fusion layer, and an action output layer. The state encoding layer extracts temporal features and multi-dimensional scoring features from the real-time state information matrix; the constraint encoding layer compresses and maps the script rule features in the policy constraint vector; the feature fusion layer jointly represents the state and constraint features; and the action output layer outputs the candidate script selection results and corresponding parameter adjustment actions. This hierarchical network structure enables the deep reinforcement learning agent to simultaneously possess the comprehensive modeling capability for both classroom dynamic states and script structure boundaries.

[0120] In one optional implementation, the policy network can adopt a two-stream network architecture based on an attention mechanism. The first stream processes the real-time state information matrix collected by the student agent to extract state features representing user participation, teaching execution, and scene changes. The second stream processes the policy constraint vectors corresponding to candidate activity interaction scripts to extract constraint features representing script structure rules and execution boundaries. Specifically, the state stream may include a temporal feature extraction unit and an attention weighting unit to extract key state change patterns from continuous time step state score changes; the constraint stream may include a multi-layer mapping unit and a rule compression unit to encode constraint information such as the main theme, node order, triggering conditions, and resource call rules into unified high-level semantic features. Then, the state features and constraint features are dynamically fused through a cross-stream feature fusion module, enabling the policy network to simultaneously consider the current classroom state and the structural constraints of the candidate scripts, thereby outputting activity interaction actions that are more closely matched to the current teaching scenario.

[0121] In the reward evaluation phase, a multi-objective reward function is defined to quantitatively evaluate the effectiveness of the current strategy action. Specifically, the multi-objective reward function can be in the form of a weighted sum, composed of teaching effectiveness reward items, user participation reward items, constraint compliance reward items, and execution efficiency reward items. The teaching effectiveness reward item can be calculated based on knowledge mastery, task completion quality, and cultural understanding depth. The user participation reward item can be calculated based on interaction frequency, dwell time, responsiveness in answering questions, and emotional feedback. The constraint compliance reward item can be quantitatively evaluated based on whether teaching time limits are exceeded, whether safety requirements are met, whether the content is appropriate, and whether knowledge points are fully covered. The execution efficiency reward item can be calculated based on time utilization rate, resource allocation efficiency, and task progress efficiency.

[0122] Furthermore, different weights can be assigned to each reward item based on the classroom stage, the students' cognitive level, and the current activity objectives. For example, in the early stages of class, the weights of user engagement rewards and content adaptability can be increased, while in the later stages of class, the weights of teaching effectiveness rewards and task completion efficiency can be increased, thus making the optimization direction of the strategy network more in line with actual teaching needs.

[0123] It is worth noting that the multi-objective reward function includes, but is not limited to, teaching effectiveness reward items, user engagement reward items, constraint compliance reward items, and execution efficiency reward items. Specifically, the teaching effectiveness reward item reflects the mastery of knowledge points and the degree of cultural understanding; the user engagement reward item reflects interaction frequency, dwell time, enthusiasm for answering questions, and emotional feedback; the constraint compliance reward item reflects the degree to which the current solution meets the requirements of teaching time, safety requirements, content suitability, and knowledge point coverage; and the execution efficiency reward item reflects time utilization efficiency and resource utilization efficiency. The system can assign different weights to the above reward items according to different classroom stages and different student group characteristics, thereby making the policy network more aligned with current teaching needs.

[0124] During the policy training phase, an online training algorithm based on policy optimization can be used to continuously update the deep reinforcement learning policy network. Specifically, the system can use a real-time state information matrix as the current state input, a policy constraint vector as the action selection constraint, and the policy network outputs candidate activity interaction actions in the current state. These activity interaction actions may include selecting a candidate script, adjusting the order of task nodes, changing the interaction trigger threshold, modifying the task difficulty coefficient, and changing the content output method. Subsequently, the system updates the policy network parameters based on the reward feedback results after the action is executed, making subsequent decisions more inclined to generate action selections that improve teaching effectiveness, enhance student participation, and meet teaching constraints. Through continuous online interaction and feedback updates, the deep reinforcement learning agent can gradually learn better activity interaction scheduling strategies for different teaching scenarios.

[0125] During the scheme generation phase, the system dynamically generates target activity interaction schemes based on the optimal activity interaction actions output by the policy network and combined with the activity organization logic and task orchestration logic in the corresponding candidate activity interaction scripts. Specifically, if the policy network determines that the current classroom is more suitable for an interaction-first approach, the system can select candidate scripts containing more interaction nodes and accelerate the pace of cultural content delivery while appropriately lowering the interaction trigger threshold. If the policy network determines that the current student group is more suited to a gradual cognitive path, the task order can be adjusted to prioritize the execution of basic explanation nodes and reduce the difficulty level of subsequent creative tasks. In this way, the target activity interaction scheme is no longer a fixed template, but an adaptive execution scheme dynamically generated by the deep reinforcement learning agent based on real-time state information and candidate script structure.

[0126] Furthermore, the target activity interaction scheme includes cultural content push instructions, interaction node trigger parameters, task difficulty adjustment coefficients, and content presentation mode configurations. Specifically, the cultural content push instructions control what cultural content should be displayed to student terminals and its push order; the interaction node trigger parameters control node switching conditions, question-answering trigger conditions, or interaction activation thresholds; the task difficulty adjustment coefficient controls task complexity, question level, or creative requirements; and the content presentation mode configuration controls whether teaching content is presented in text, image, video, audio, or a hybrid format. For example, in a classroom setting with high interactive participation and high cognitive load, the system can increase the proportion of image and video content, reduce the proportion of long text explanations, and appropriately relax the task trigger thresholds.

[0127] Optionally, during the strategy decision-making process, the execution effect of the target activity interaction scheme is monitored. When the deviation index in the real-time status information matrix exceeds a preset threshold, a strategy replanning mechanism is triggered to reselect a suitable candidate script from the candidate activity interaction script set and adjust the strategy network parameters. Specifically, the fit between the current target activity interaction scheme and the real-time classroom status can be reassessed, a more suitable candidate script can be selected from the candidate activity interaction script set, and the strategy network parameters can be adjusted simultaneously to correct the original decision direction.

[0128] For example, when the system detects that students' dwell time at the current explanation node is continuously decreasing, the task completion rate is significantly lower than expected, and the environmental noise level is rising, it can trigger a replanning mechanism to switch the original explanation-oriented activity path to an interaction-first path, while appropriately compressing the subsequent text explanation content to improve classroom recovery ability and activity execution stability.

[0129] Step 106: Based on the target activity interaction scheme, the student-side intelligent agent pushes the activity interaction guide results to the student terminal, the teacher-side intelligent agent pushes the teaching monitoring results to the teacher terminal, and calls the content distribution module integrated with the AIGC toolchain to generate teaching content that matches the target activity interaction scheme.

[0130] The interactive activity navigation results include at least node identifiers for the target activity content, participation paths, task prompts, interaction criteria, and recommendation criteria. The teaching monitoring results include at least student participation status, task completion status, teaching progress matching status, and strategy adjustment criteria. For example, in a Minnan ancient house culture experience class, student terminals can receive activity navigation results including nodes for identifying ancient houses, interactive experiences with Gaojia opera, and explanations of local products, and display corresponding participation paths and task prompts. Teacher terminals can simultaneously display teaching monitoring results such as student participation levels at each node, quiz completion status, achievement of teaching objectives, and reasons for current strategy adjustments.

[0131] The teaching content may include one or more of the following: teaching materials, promotional materials, visual material prompts, courseware presentation content, and short video scripts. Teaching materials may include cultural background explanations, task requirements, knowledge point summaries, and interactive prompts for classroom lectures. Promotional materials may be introductory texts for course presentations, activity promotions, or results announcements. Visual material prompts may be a set of prompts for generating posters, illustrations, character designs, scene diagrams, or storyboard sketches. Courseware presentation content may include PPT page text, graphic layout instructions, and page structure content adapted to classroom presentation logic. Short video scripts may include shot arrangements, narration, transition logic, and character action descriptions generated around a specific cultural theme. For example, in the "Ancient House Opera Charm" themed activity, the teaching materials can include explanations of the architectural features and cultural connotations of ancient Minnan houses; the promotional materials can include course promotional slogans such as "Explore Ancient House Opera Charm and Experience the Charm of Minnan Culture"; the visual material prompts can include image generation prompts such as "Minnan red brick ancient houses, opera characters, warm sunset, and traditional Chinese style illustrations"; the courseware display content can include pages such as ancient house architecture identification pages, Gaojia opera character comparison pages, and creative task release pages; and the short video scripts can include scene scripts such as students entering the ancient house scene, imitating opera movements, and summarizing and displaying cultural elements.

[0132] The content distribution module integrated with the AIGC toolchain refers to a content processing and output unit used to automatically generate, integrate, and distribute multimodal teaching content according to the target activity interaction plan. This content distribution module can integrate multiple AIGC capability units, such as text generation models, image generation models, video generation models, speech generation models, digital human generation modules, and script orchestration modules. Based on the main theme, task nodes, resource call rules, and content output positions defined in the target activity interaction plan, it automatically selects suitable generation tools and templates to generate teaching content consistent with the current activity flow. In other words, the content distribution module is not a single content generation tool, but a comprehensive module capable of collaboratively orchestrating, uniformly scheduling, and distributing different types of generated content to terminals according to a specific activity execution plan.

[0133] For example, in a Minnan traditional house culture experience class, the content distribution module can first call a text generation model to generate explanatory texts about the architectural culture of the traditional houses and interactive Q&A questions about Gaojia opera. Then, it can call an image generation model to generate illustrations of the traditional house scenes and poster prompts for opera characters. Next, it can call a short video script generation module to output short video scripts for the traditional house tour and opera experience. It can also call a voice generation or digital human module to generate audio explanations or digital human demonstrations with a Minnan-style spoken delivery. Afterwards, the content distribution module can push the generated results to the teacher's terminal for classroom presentation, or to the student's terminal for task execution, creative reference, and post-class review, depending on the content output positions at different nodes in the target activity interaction plan. This ensures a high degree of consistency and correspondence between the generated content and the activity theme, activity flow, and interactive nodes.

[0134] As an optional embodiment, in step 106, the step of pushing activity interaction guidance results from the student-side agent to the student terminal and pushing teaching monitoring results from the teacher-side agent to the teacher terminal based on the target activity interaction scheme includes: The student-side intelligent agent generates a multimedia guide package based on the interaction node definitions and current scene location information in the target activity interaction scheme. This package includes augmented reality navigation instructions, interactive task prompts, historical and cultural explanations, and personalized recommendations. The multimedia guide package undergoes adaptive compression and format conversion based on the student terminal's device performance, network status, and user interaction history. Based on the user participation score and cognitive load level in the real-time status information matrix, the push frequency, content complexity, and interaction depth of the multimedia guide package are dynamically adjusted. When the user's cognitive load exceeds a threshold, the content presentation format is automatically simplified and guidance prompts are added. The teacher-side intelligent agent summarizes the real-time data reported by each student terminal. The system generates a teaching monitoring panel that includes a heatmap of activity execution progress, a distribution map of student participation, a statistical chart of task completion rate, and early warning signals for abnormal behavior. The teaching monitoring panel supports multi-granularity data drilling based on time, space, student grouping, and cultural elements. It also pushes teaching intervention suggestions to teachers' terminals. These suggestions are generated based on the comparison between the deviation indicators and preset thresholds, and include suggestions for adjusting activity pace, reorganizing groups, adjusting content difficulty, and issuing safety risk warnings. Teachers can confirm, modify, or ignore these suggestions through their terminals, and the results are fed back to the deep reinforcement learning agent in real time for strategy optimization.

[0135] For example, in the implementation of pushing interactive activity guides and teaching monitoring results based on the target activity interaction scheme, assuming a Minnan ancient house culture experience class, the student-side intelligent agent can push a multimedia guide package to the student terminal based on the node definitions in the current target activity interaction scheme and the student's location. This package includes guidance on ancient house identification, prompts for Gaojia opera interactive tasks, explanations of local product backgrounds, and personalized learning suggestions. When a student terminal's user participation score is low or its cognitive load level is high, the complexity of the guide content can be automatically reduced, the amount of information displayed on the screen can be decreased, and step prompts and task breakdown explanations can be added. Simultaneously, the teacher-side intelligent agent can summarize the status information of each student terminal, generate a teaching monitoring panel reflecting the activity's progress, student participation status, task completion rate, and abnormal behavior, and push corresponding teaching intervention suggestions to the teacher terminal. For example, if a group lingers too long at a Gaojia opera interactive node, the teacher can be advised to appropriately reduce the task difficulty or readjust the groupings. If the simulated tourist gathering in a certain area is too high, suggestions for adjusting the activity pace or safety risk warnings can be sent to the teacher. After the teacher confirms, modifies, or ignores the suggestions, the relevant results can be fed back to the deep reinforcement learning agent for subsequent strategy optimization.

[0136] As an optional embodiment, in step 106, the step of calling the content distribution module integrated with the AIGC toolchain to generate teaching content matching the target activity interaction scheme includes: constructing a multimodal AIGC toolchain, which includes a text generation submodule, an image generation submodule, an audio generation submodule, and a video generation submodule. Each submodule performs semantic alignment with the knowledge graph entity embedding vector through a unified content description interface; according to the cultural content push instructions and content presentation configuration in the target activity interaction scheme, extracting the attribute information, relationships, and historical context of relevant cultural element entities from the cultural tourism knowledge graph to generate a structured content prompt template, which includes cultural connotation keywords, target audience characteristics, emotional expression tendencies, and media presentation requirements; and inputting the structured content prompt template into the AIGC toolchain to generate text explanation content, augmented reality visual materials, environmental sound effects materials, and interactive content through a multimodal collaborative generation mechanism. The video clips include a text generation submodule that performs style transfer based on the semantic similarity of cultural element entities, an image generation submodule that performs 3D reconstruction based on the geometric features of architectural features, and an audio generation submodule that performs timbre synthesis based on the acoustic characteristics of regional product entities. The generated multimodal teaching content undergoes quality assessment and consistency verification. The quality assessment includes cultural accuracy verification, teaching suitability scoring, and technical integrity detection. The consistency verification is achieved by calculating the cosine similarity between the feature vector of the generated content and the corresponding entity embedding vector in the knowledge graph. When the similarity is lower than a preset threshold, a content regeneration mechanism is triggered. The multimodal teaching content that has passed quality assessment and consistency verification is then spatiotemporally encapsulated to generate an adaptive content package that conforms to streaming media transmission protocols. Based on the environmental adaptability score in the real-time status information matrix of the student terminal, the resolution, bitrate, and interaction complexity of the content package are dynamically adjusted to ensure a smooth and immersive teaching experience under different network environments and device conditions.

[0137] For example, in a themed classroom activity on the theme of "Ancient Houses and Opera," the system can extract attribute information, relationships, and historical background information of cultural elements such as ancient Minnan houses, Gaojia opera, and local products from the cultural tourism knowledge graph, based on the cultural content push instructions and content presentation configuration in the target activity interaction scheme, and generate corresponding structured content prompt templates. Subsequently, the content distribution module can generate multimodal teaching content based on the prompt templates, including explanations of ancient house culture, classroom illustrations, background sound effects for interactive sessions, and short video teaching clips. After generation, the system can perform cultural accuracy verification, teaching suitability scoring, and consistency verification on the above content. If a significant semantic deviation is detected between a generated content and the corresponding cultural entity in the knowledge graph, a regeneration mechanism can be triggered. Finally, the system encapsulates the verified multimodal teaching content into a content package adapted to the current terminal and network conditions, and dynamically adjusts the display clarity, playback bitrate, and interaction complexity based on the student's terminal's environmental adaptability score to ensure smoothness and immersion during classroom presentations and student interactions.

[0138] As an optional embodiment, during and after the activity interaction, the student-side agent continuously collects multi-dimensional feedback information from students regarding the activity interaction guidance results. This multi-dimensional feedback information includes temporal data of interactive behavior, task completion quality indicators, emotional state change curves, and explicit evaluation text. Then, the multi-dimensional feedback information is spatiotemporally aligned with the corresponding real-time state information matrix to construct an experience replay buffer containing state-action-feedback triples. This experience replay buffer employs a priority sampling mechanism, dynamically adjusting sampling weights based on the information gain and teaching value of the feedback information. Next, the deep reinforcement learning agent periodically extracts high-value samples from the experience replay buffer and fine-tunes the policy network parameters online using a policy gradient update algorithm. This online fine-tuning process employs elastic weight consolidation regularization technology to prevent catastrophic forgetting of historical knowledge by new experiences. The teacher-side agent performs semantic annotation and importance classification on the multi-dimensional feedback information based on teaching monitoring results. The semantic annotation includes teaching effectiveness labels, cultural understanding depth labels, and safety compliance labels. The importance classification is based on the correlation between the feedback information and the teaching objectives. When a preset proportion of student feedback information deviates from the expected teaching trajectory, a local replanning mechanism is triggered. The local replanning mechanism includes: subgraph reconstruction based on knowledge graph, incremental generation of candidate activity interaction scripts, and targeted optimization of strategy network parameters, so as to dynamically adjust subsequent teaching strategies and generate updated target activity interaction schemes.

[0139] In practical applications, the student-side intelligent agent can continuously collect multi-dimensional feedback information during and after activities, including student click order, dwell time, task submission results, answer accuracy, voice interaction content, facial expression trends, and post-class evaluation text. This feedback information is then spatiotemporally aligned with a real-time state information matrix for the corresponding time period, forming state, action, and feedback correlation data reflecting the classroom evolution process. Subsequently, the system can write this correlation data into an experience replay buffer and assign different sampling priorities to different samples based on their contribution to improving teaching effectiveness, their reference value for strategy correction, and their relevance to teaching objectives. This prioritizes and extracts samples with high teaching value for subsequent online optimization.

[0140] Furthermore, the deep reinforcement learning agent can extract high-value samples from the experience replay buffer at preset intervals to fine-tune the policy network parameters online, continuously improving the adaptability and stability of subsequent activity interaction schemes. During the fine-tuning process, the system can simultaneously maintain the importance constraints of the parameters corresponding to historical high-quality samples to reduce the disruption of existing effective strategies after the introduction of new samples, avoiding the forgetting of historical teaching patterns or excessive policy fluctuations. At the same time, the teacher-side agent can also combine teaching monitoring results to semantically label and classify the importance of student feedback information. For example, it can mark a certain type of feedback as mainly reflecting insufficient teaching effectiveness, insufficient cultural understanding, or potential safety hazards in classroom implementation, and accordingly increase the reference weight of relevant samples in the subsequent optimization process.

[0141] For example, in a Minnan ancient house culture experience class, if most students show high participation in the Gaojia opera interactive nodes, but generally spend less time on the ancient house architecture explanation nodes, have a low accuracy rate in answering questions, and repeatedly report abstract content and difficulty in understanding in post-class evaluations, the system can determine that this part of the teaching path deviates from the expected teaching trajectory. When such deviation feedback reaches a preset proportion, a local replanning mechanism is triggered. At this time, the system can reorganize the cultural content and teaching task nodes around the ancient house architecture cognition-related subgraph, incrementally generating candidate activity interaction scripts more suitable for the current student group. For example, adding image recognition tasks, reducing the density of professional terminology, inserting contextualized interactive nodes in advance, and simultaneously optimizing the strategy network parameters. After the above adjustments, the system can output an updated target activity interaction plan to improve the teaching effect, participation experience, and content suitability of subsequent classroom activities.

[0142] In this embodiment, by acquiring local cultural resource data related to cultural tourism teaching activities and constructing a cultural tourism knowledge graph, knowledge representation learning is performed on the cultural tourism knowledge graph. Based on the correlation between cultural elements and teaching tasks, candidate activity interaction scripts are generated. Then, combined with the real-time status information collected by the student-side intelligent agent, a deep reinforcement learning intelligent agent makes strategy decisions. This can obtain a target activity interaction scheme that dynamically matches the current teaching scenario, thereby reducing the reliance on human experience and template reuse in cultural tourism activity planning, reducing the manual proofreading costs caused by the mismatch between activity content and activity process, bringing users a more personalized, interactive, and locally culturally distinctive learning experience, and reducing the workload of manual intervention and repeated modifications in the teaching organization and content generation process.

[0143] The above describes a cultural and tourism activity interaction method based on knowledge graphs and deep reinforcement learning in the embodiments of this application. The following describes the cultural and tourism activity interaction system based on knowledge graphs and deep reinforcement learning that implements the above-described cultural and tourism activity interaction method based on knowledge graphs and deep reinforcement learning.

[0144] See Figure 3 ,like Figure 3 The diagram shows a structural schematic of a cultural tourism activity interaction system based on knowledge graphs and deep reinforcement learning. The cultural tourism activity interaction system based on knowledge graphs and deep reinforcement learning in this embodiment can achieve the above-mentioned... Figure 2 The steps of the knowledge graph and deep reinforcement learning-based cultural tourism activity interaction method executed in the corresponding embodiments are not elaborated here, but similarities can be found in the previous embodiments. The functions implemented by the knowledge graph and deep reinforcement learning-based cultural tourism activity interaction system can be implemented in hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above functions, and the modules can be software and / or hardware. The system includes a server, a teacher-side intelligent agent, and a student-side intelligent agent. The system includes: The server is configured to acquire local cultural resource data related to cultural tourism teaching activities, and construct a cultural tourism knowledge graph based on the local cultural resource data; and perform knowledge representation learning on the cultural tourism knowledge graph to obtain vectorized representations of each entity in the cultural tourism knowledge graph and the relationships between each entity. The teacher-side intelligent agent is configured to construct candidate activity interaction scripts adapted to the current teaching scenario based on the correlation between cultural elements and teaching tasks in the vectorized representation. The candidate activity interaction scripts are used to represent activity organization logic, task arrangement logic, interaction triggering logic, and content output logic. The agent also performs teaching constraints, strategy corrections, and / or manual interventions on the candidate activity interaction scripts. Based on the target activity interaction scheme, the agent pushes teaching monitoring results to the teacher terminal. The student-side intelligent agent is configured to collect real-time status information during the interaction process based on the candidate activity interaction script to characterize user participation, teaching execution, and / or scene changes; and to push activity interaction guide results to the student terminal based on the target activity interaction scheme. The server is also configured to call a deep reinforcement learning agent to make policy decisions based on the real-time state information in conjunction with the candidate activity interaction script, thereby obtaining the target activity interaction scheme; and to call a content distribution module integrated with the AIGC toolchain to generate teaching content that matches the target activity interaction scheme.

[0145] In this embodiment, local cultural resource data related to cultural tourism teaching activities is acquired by the server and a cultural tourism knowledge graph is constructed. The knowledge graph is then subjected to knowledge representation learning to obtain vectorized representations of each entity and the relationships between entities. The teacher-side intelligent agent constructs candidate activity interaction scripts adapted to the current teaching scenario based on the correlation between cultural elements and teaching tasks. The student-side intelligent agent collects real-time status information during the interaction process. The server then calls a deep reinforcement learning intelligent agent to make strategy decisions based on the real-time status information in conjunction with the candidate activity interaction scripts. This results in a target activity interaction scheme that dynamically matches the current classroom context, student characteristics, and teaching objectives. Furthermore, teaching content that matches the target activity interaction scheme is generated. This significantly reduces the reliance on human experience, fixed templates, and repetitive proofreading in the planning and organization of cultural tourism activities and classrooms. It provides users with a more personalized, interactive, and culturally immersive teaching experience and reduces the manual adjustment costs for teachers in the process of activity arrangement, process monitoring, and content generation.

[0146] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0147] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0148] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules 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, apparatuses, or modules, and may be electrical, mechanical, or other forms.

[0149] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0150] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0151] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0152] The computer program product includes one or more computer instructions. When the computer program is loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0153] The technical solutions provided in the embodiments of this application have been described in detail above. Specific examples have been used in the embodiments of this application to illustrate the principles and implementation methods of the embodiments of this application. The description of the above embodiments is only for the purpose of helping to understand the methods and core ideas of the embodiments of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the embodiments of this application. Therefore, the content of this specification should not be construed as a limitation on the embodiments of this application.

Claims

1. A method for interactive cultural and tourism activities based on knowledge graphs and deep reinforcement learning, characterized in that, The method is executed collaboratively by a teacher-side intelligent agent and a student-side intelligent agent, and the method includes: Acquire local cultural resource data related to cultural and tourism teaching activities, and construct a cultural and tourism knowledge graph based on the local cultural resource data; Knowledge representation learning is performed on the cultural and tourism knowledge graph to obtain vectorized representations of each entity in the cultural and tourism knowledge graph and the relationships between them; The teacher-side intelligent agent constructs candidate activity interaction scripts adapted to the current teaching scenario based on the relationship between the Chinese cultural elements and teaching tasks in the vectorized representation. The candidate activity interaction scripts are used to represent the activity organization logic, task arrangement logic, interaction triggering logic, and content output logic. Teaching constraints, strategy corrections, and / or manual interventions are performed on the candidate activity interaction scripts. The student-side intelligent agent collects real-time status information during the interaction process based on the candidate activity interaction script to characterize user participation, teaching execution, and / or scene changes. A deep reinforcement learning agent, combined with the candidate activity interaction script, makes a strategy decision on the real-time state information to obtain the target activity interaction scheme. Based on the target activity interaction scheme, the student-side intelligent agent pushes the activity interaction guide results to the student terminal, the teacher-side intelligent agent pushes the teaching monitoring results to the teacher terminal, and calls the content distribution module integrated with the AIGC toolchain to generate teaching content that matches the target activity interaction scheme.

2. The interactive method for cultural and tourism activities based on knowledge graphs and deep reinforcement learning according to claim 1, characterized in that, The acquisition of local cultural resource data related to cultural tourism teaching activities, and the construction of a cultural tourism knowledge graph based on the local cultural resource data, includes: It can acquire local cultural resources, activity scenarios, teaching objectives, teaching tasks, and activity resources information entered by teachers on their terminals, and student characteristics, learning behaviors, and participation intentions information entered or authorized by students on their terminals. Preprocessing and semantic analysis of local cultural resource information to extract cultural elements corresponding to cultural objects, historical events, folk activities, performance forms, architectural styles and local products; The activity scenario information, teaching objective information, teaching task information, and activity resource information are structured and analyzed to extract scenario elements, teaching task elements, teaching objective elements, and resource elements. The extracted cultural elements, scene elements, teaching task elements, teaching objective elements and resource elements are subjected to entity recognition, relation extraction and attribute alignment to establish the entities corresponding to each element in the cultural tourism knowledge graph and the relationship between entities. By associating student characteristic information, learning behavior information, and activity participation intention information as user profile tags with the corresponding entities in the cultural and tourism knowledge graph, a cultural and tourism knowledge graph is obtained.

3. The interactive method for cultural and tourism activities based on knowledge graphs and deep reinforcement learning according to claim 2, characterized in that, The step of performing knowledge representation learning on the cultural tourism knowledge graph to obtain vectorized representations of each entity in the cultural tourism knowledge graph and the relationships between entities includes: The cultural tourism knowledge graph is input into a preset knowledge graph embedding model, and low-dimensional vector space mapping is performed on the cultural element entities, teaching task element entities, scene element entities and resource element entities in the cultural tourism knowledge graph. Based on the triplet relationships in the cultural tourism knowledge graph, positive and negative sample pairs are constructed. By optimizing the goal of maximizing the score of positive sample triplets and minimizing the score of negative sample triplets, the knowledge graph embedding model is iteratively trained to ensure that entities and relationships with semantic associations maintain the corresponding geometric constraints in the vector space. After completing the model training, extract and output the entity embedding vectors corresponding to each entity in the cultural tourism knowledge graph, as well as the relationship embedding vectors representing the relationships between entities; Using the entity embedding vector and relation embedding vector, the semantic similarity between cultural element entities and the association weight between cultural element entities and teaching task element entities, scene element entities and resource element entities are calculated by the vector space metric method.

4. The interactive method for cultural and tourism activities based on knowledge graphs and deep reinforcement learning according to claim 3, characterized in that, The knowledge graph embedding model adopts the ComplEx model architecture. The low-dimensional vector space mapping of cultural element entities, teaching task element entities, scene element entities, and resource element entities in the cultural tourism knowledge graph includes: Map each entity and relation in the cultural tourism knowledge graph into a complex vector representation; A scoring function based on Hermitian bilinear form is used to model the triplet relationship in the cultural tourism knowledge graph. The scoring function is used to calculate the association score between the head entity, the relationship, and the tail entity. The association score represents the asymmetric relationships, multi-type relationships, and multi-hop combination relationships between entities.

5. The interactive method for cultural and tourism activities based on knowledge graphs and deep reinforcement learning according to claim 4, characterized in that, The process involves constructing positive and negative sample pairs based on the triple relationships in the cultural tourism knowledge graph, and iteratively training the knowledge graph embedding model by optimizing the goal of maximizing the score of positive sample triples and minimizing the score of negative sample triples. This includes: Based on the positive sample triples in the cultural tourism knowledge graph, a negative sample triples corresponding to the positive sample triples are constructed using a negative sampling method oriented towards complex vector space. The association scores of the positive sample triples and the negative sample triples are calculated based on the ComplEx model architecture. The association scores are optimized using the logistic loss function to increase the scores of positive triples and decrease the scores of negative triples, so as to obtain entity embedding vectors and relation embedding vectors that satisfy semantic constraints.

6. The interactive method for cultural and tourism activities based on knowledge graphs and deep reinforcement learning according to claim 1, characterized in that, The process of constructing candidate activity interaction scripts adapted to the current teaching scenario by the teacher-side intelligent agent based on the vectorized representation of the relationship between cultural elements and teaching tasks includes: Based on the entity embedding vector and relation embedding vector in the vectorized representation, the association weight between cultural element entities and teaching task element entities is calculated, as well as the semantic similarity between cultural element entities, to determine the core cultural element set that matches the current teaching scenario. Based on the core cultural elements set, combined with activity scenario information, teaching objective information and student characteristic information, multiple candidate activity interaction paths are generated through knowledge graph-based path planning algorithm and deep search algorithm. Each candidate activity interaction path includes the cultural content presentation order, task triggering conditions, interaction node settings and content output strategy. Feasibility assessment and teaching effect prediction are performed on the multiple candidate activity interaction paths, and a set of candidate activity interaction scripts that meet the teaching constraints are selected.

7. The interactive method for cultural and tourism activities based on knowledge graphs and deep reinforcement learning according to claim 1, characterized in that, The real-time status information collected by the student-side intelligent agent during the interaction process based on the candidate activity interaction script, used to characterize user participation, teaching execution, and / or scene changes, includes: The student terminal's behavior perception module collects user behavior data, which includes location coordinate information, gaze focus trajectory, gesture operation sequence, voice interaction content, physiological state parameters, and task completion progress. Based on the interaction node definition in the candidate activity interaction script, feature extraction and temporal alignment are performed on the collected user behavior data to generate a multi-dimensional feature vector containing user participation depth, cognitive load level and emotional state. The physical environmental parameters of cultural and tourism activities are monitored through the environmental perception module on the student terminal. The multi-dimensional feature vectors are fused with the physical environment parameters, and combined with the expected execution path of the target activity interaction scheme, the deviation index between the actual execution state and the expected state is calculated by a weighted multi-dimensional dynamic time warping algorithm. Based on the aforementioned deviation index, a real-time state information matrix is ​​dynamically constructed, which includes user participation score, task execution efficiency score, environmental adaptability score, and teaching goal achievement score, serving as the state input for the deep reinforcement learning agent to make policy decisions.

8. The interactive method for cultural and tourism activities based on knowledge graphs and deep reinforcement learning according to claim 7, characterized in that, The process of obtaining a target activity interaction scheme by using a deep reinforcement learning agent to make policy decisions based on the real-time state information in conjunction with the candidate activity interaction scripts includes: Encode each candidate script in the candidate activity interaction script set into a policy constraint vector; A deep reinforcement learning policy network is constructed. The policy network adopts a two-stream network architecture based on an attention mechanism. The first stream processes the real-time state information matrix, and the second stream processes the policy constraint vector. The dynamic fusion of state features and constraints is achieved through a cross-stream attention module. Define a multi-objective reward function, wherein the constraint compliance reward item is quantitatively evaluated based on the teaching constraint rules in the candidate activity interaction script; The deep reinforcement learning policy network is trained online using a proximal policy optimization algorithm. The real-time state information matrix is ​​used as the current state input, and the policy constraint vector is used as the action space constraint condition. The optimal activity interaction action in the current state is output. Based on the optimal activity interaction action, and combined with the activity organization logic and task orchestration logic in the candidate activity interaction script, the target activity interaction scheme is dynamically generated.

9. The interactive method for cultural and tourism activities based on knowledge graphs and deep reinforcement learning according to claim 7, characterized in that, The process of pushing activity interaction guide results from the student-side agent to the student terminal and pushing teaching monitoring results from the teacher-side agent to the teacher terminal based on the target activity interaction scheme includes: The student-side intelligent agent generates a multimedia guide package containing augmented reality guide instructions, interactive task prompts, historical and cultural explanations, and personalized recommended content based on the interactive node definitions in the target activity interaction scheme and the current scene location information. Based on the user engagement score and cognitive load level in the real-time status information matrix, the push frequency, content complexity and interaction depth of the multimedia guide package are dynamically adjusted. When the user's cognitive load exceeds the threshold, the content presentation format is simplified and guidance prompts are added. The teacher-side intelligent agent aggregates the real-time status information reported by each student terminal and generates a teaching monitoring panel that includes a heat map of activity execution progress, a distribution map of student participation, a statistical chart of task completion rate, and early warning signals of abnormal behavior. The teaching monitoring panel supports multi-granularity data drilling by time dimension, space dimension, student group dimension, and cultural element dimension. Teaching intervention suggestions are pushed to the teacher's terminal. These suggestions are generated based on the comparison between the deviation index and a preset threshold. This allows the teacher to confirm, modify, or ignore the suggestions through the teacher's terminal, and the results are fed back to the deep reinforcement learning agent in real time for strategy optimization.

10. A cultural tourism activity interaction system based on knowledge graphs and deep reinforcement learning, characterized in that, The system includes a server, a teacher-side intelligent agent, and a student-side intelligent agent. The server is configured to acquire local cultural resource data related to cultural tourism teaching activities, and construct a cultural tourism knowledge graph based on the local cultural resource data; and perform knowledge representation learning on the cultural tourism knowledge graph to obtain vectorized representations of each entity in the cultural tourism knowledge graph and the relationships between each entity. The teacher-side intelligent agent is configured to construct candidate activity interaction scripts adapted to the current teaching scenario based on the correlation between cultural elements and teaching tasks in the vectorized representation. The candidate activity interaction scripts are used to represent activity organization logic, task arrangement logic, interaction triggering logic, and content output logic. The agent also performs teaching constraints, strategy corrections, and / or manual interventions on the candidate activity interaction scripts. Based on the target activity interaction scheme, the agent pushes teaching monitoring results to the teacher terminal. The student-side intelligent agent is configured to collect real-time status information during the interaction process based on the candidate activity interaction script to characterize user participation, teaching execution, and / or scene changes; and to push activity interaction guide results to the student terminal based on the target activity interaction scheme. The server is also configured to call a deep reinforcement learning agent to make policy decisions based on the real-time state information in conjunction with the candidate activity interaction script, thereby obtaining the target activity interaction scheme; and to call a content distribution module integrated with the AIGC toolchain to generate teaching content that matches the target activity interaction scheme.

Images