Game resource intelligent classification and retrieval system based on semantic network
By constructing a multidimensional semantic network and cross-modal feature extraction, combined with deep reinforcement learning and user feedback, the problems of flexibility and accuracy in game resource classification and retrieval are solved, and maintenance costs are reduced.
Patent Information
- Application Number
- CN202511108763.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-11-21
AI Technical Summary
Existing technologies for classifying and retrieving game resources lack flexibility, cannot adapt to diverse needs, face difficulties in cross-type resource retrieval, and lack adaptability, resulting in high maintenance costs.
A multidimensional semantic network is constructed, and cross-modal feature extraction and deep reinforcement learning are adopted to achieve automatic resource classification and multimodal query, and dynamic optimization is carried out in combination with user feedback.
It has improved the flexibility and accuracy of game resource classification, enhanced the precision of cross-type resource retrieval, and reduced manual maintenance costs.
Smart Images
Figure CN120994886A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of game development technology, specifically to a game resource intelligent classification and retrieval system based on semantic networks. Background Technology
[0002] Game development generates a large number of game resources, such as 3D models, images, animations, special effects, and code resources. With the rapid development of the game industry, the quantity and variety of game resources have increased dramatically. Therefore, efficient and accurate classification and retrieval of these resources has become crucial for improving game development efficiency.
[0003] Current technologies for classifying and retrieving game resources have several shortcomings. In terms of classification, most rely on fixed, manually pre-defined systems, dividing resources based solely on single dimensions such as type and format. This fails to adapt to the diverse needs of different development scenarios, resulting in a lack of flexibility and specificity in the classification results. Regarding retrieval, it primarily depends on keyword matching, making it difficult to deeply understand the semantic information of resources and the relationships between them. This leads to low relevance in search results, especially when searching across different resource types, where the results are even more unsatisfactory.
[0004] Furthermore, different types of game resources have different feature representations, making it difficult for existing technologies to unify them into a common semantic space for processing. This results in weak cross-modal processing capabilities, making cross-type resource association retrieval quite challenging. At the same time, most existing systems lack adaptability, failing to automatically optimize classification and retrieval strategies based on developers' usage habits and feedback, requiring continuous manual adjustments and incurring high maintenance costs. Summary of the Invention
[0005] The purpose of this invention is to provide a game resource intelligent classification and retrieval system based on semantic networks to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a game resource intelligent classification and retrieval system based on semantic networks, comprising:
[0007] The resource semantic modeling module is used to construct a multi-dimensional semantic network for game resources. This semantic network includes a resource entity layer, an attribute feature layer, an association rule layer, and a contextual scene layer. The resource entity layer covers entity nodes of 3D models, images, animations, special effects, and code resources. The attribute feature layer includes the visual features, functional features, format features, and creative features of the resources. The association rule layer stores the semantic association weights between entities, which are dynamically adjusted based on co-occurrence frequency and manual annotation. The contextual scene layer records the application context of resources in the game development scenario.
[0008] A cross-modal feature extraction module is used to extract features from different types of game resources and map them to a unified semantic space, including:
[0009] For 3D models, extract geometric topological features, material texture features, and skeletal animation features;
[0010] For image resources, extract color histograms, texture features, and content tag features;
[0011] Extract time series features, keyframe features, and particle system parameter features from animation and special effects resources;
[0012] For code resources, functional semantic features and call relationship features are extracted through abstract syntax tree analysis;
[0013] The dynamic classification module, based on deep reinforcement learning, enables automatic resource classification, including:
[0014] The classification strategy network takes the association rules of the semantic network as input and outputs resource classification paths;
[0015] The reward function calculates the reward value based on the consistency of the classification results, retrieval efficiency, and adaptability to the development scenario.
[0016] The classification results optimization unit optimizes classification strategies through continuous interaction and feedback with developers.
[0017] The intelligent search module supports multimodal queries and semantically expanded searches, including:
[0018] The multimodal query parsing unit converts text descriptions, sketch inputs, and example resource inputs into semantic query vectors;
[0019] The semantic extension unit generates extended sets of synonyms, hyponyms, and related words for query keywords based on the association rule layer of the semantic network;
[0020] The retrieval and ranking unit sorts resources based on semantic similarity to the query, resource usage frequency, and matching degree with development scenarios.
[0021] The self-evolution module dynamically updates the association rules and feature weights of the semantic network by monitoring resource usage logs and developer feedback, thereby achieving continuous system optimization.
[0022] Furthermore, the cross-modal feature extraction module also includes a feature mapping unit, which uses an adversarial generative network to map features of different types of resources to a unified semantic space. The adversarial generative network includes:
[0023] Multiple domain feature encoders, each corresponding to different types of game resources;
[0024] A shared semantic decoder maps domain-specific feature codes to a unified semantic vector;
[0025] Domain discriminators are used to distinguish which domain a feature vector comes from. Through adversarial training, feature vectors from different domains can be compared in a shared semantic space.
[0026] Furthermore, the classification strategy network of the dynamic classification module adopts a deep Q-network, whose state space consists of the feature vector of the current resource and the existing classification path, and whose action space consists of possible next-level classification labels. The optimal classification strategy is learned through interaction with the development scenario.
[0027] Furthermore, the multimodal query parsing unit of the intelligent retrieval module uses an attention mechanism to fuse input features from different modalities. For text description input, a pre-trained language model is used to extract semantic vectors; for sketch input, contour features and structural features are extracted and converted into semantic vectors; for example resource input, the cross-modal feature extraction module is called to obtain its semantic vector, and then the semantic vectors from different modalities are fused into a unified query vector.
[0028] Furthermore, the self-evolution module includes:
[0029] The log analysis unit records the frequency of resource retrieval, download frequency, usage scenarios, and associated resource combinations;
[0030] The feedback collection unit obtains feedback information through developers' rating of search results and correction of classification results;
[0031] The network update unit updates the association rules and feature weights of the semantic network using an incremental learning approach based on log analysis results and feedback information, without needing to retrain the entire model.
[0032] Furthermore, the method is as follows:
[0033] Includes the following steps:
[0034] S1: Construct a multidimensional semantic network for game resources, the semantic network comprising a resource entity layer, an attribute feature layer, an association rule layer, and a context scene layer;
[0035] S2: Extract cross-modal features of different types of game resources and map them to a unified semantic space;
[0036] S3: Based on deep reinforcement learning, game resources are dynamically classified according to the association rules of the semantic network and development scenario information;
[0037] S4: Receives multimodal queries from users, generates extended query sets through semantic expansion, and retrieves and sorts relevant resources in the semantic network;
[0038] S5: Dynamically update the semantic network based on resource usage logs and user feedback to enable the system to self-advance.
[0039] Furthermore, in step S3, the specific process of dynamic classification is as follows:
[0040] S31: Initialize the classification policy network and set the initial reward function;
[0041] S32: Input the semantic vector of the resource and the current classification state into the classification policy network to obtain classification path suggestions;
[0042] S33: Calculate the reward value based on the consistency of the classification results, retrieval efficiency, and adaptability to the development scenario;
[0043] S34: Update the parameters of the classification policy network based on the reward value;
[0044] S35: Receive feedback from developers regarding the classification results and further optimize the classification strategy.
[0045] This invention provides a game resource intelligent classification and retrieval system based on semantic networks, which has the following beneficial effects: This invention constructs a multi-dimensional semantic network containing a resource entity layer, attribute feature layer, association rule layer and context scene layer through a resource semantic modeling module, which can comprehensively capture the deep semantic associations and context information between game resources, breaking through the limitations of the single classification system of the existing technology, and making the classification more comprehensive and flexible.
[0046] The cross-modal feature extraction module uses an adversarial generative network to map the features of different types of resources to a unified semantic space, which effectively solves the problem of difficult cross-type resource association retrieval in existing technologies and improves the accuracy of cross-modal retrieval.
[0047] The dynamic classification module uses deep reinforcement learning to automatically classify resources. It can dynamically adjust the classification strategy according to the development scenario and user feedback. Compared with fixed classification system and traditional machine learning classification methods, it has better adaptability and flexibility.
[0048] The intelligent retrieval module supports multimodal queries and semantic extended retrieval, and can deeply understand the user's semantic intent to generate relevant extended queries, thereby improving the accuracy and recall of retrieval.
[0049] The self-evolution module can dynamically update the semantic network based on resource usage logs and user feedback, enabling the system to continuously optimize and adapt to ever-changing game development needs, thus reducing manual maintenance costs. Attached Figure Description
[0050] Figure 1 This is a flowchart of a game resource intelligent classification and retrieval system based on semantic networks according to the present invention. Detailed Implementation
[0051] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.
[0052] The intelligent classification and retrieval system for game resources based on semantic networks provided by this invention enables intelligent classification and retrieval of game resources through the collaborative work of its various modules in practical applications.
[0053] When constructing a multidimensional semantic network, the resource semantic modeling module subdivides 3D models in the resource entity layer into character models, scene models, prop models, etc.; image resources include UI images, textures, posters, etc. In the attribute feature layer, various features are added to each entity node in detail, such as the number of polygons, vertices, and material type of 3D models, and the resolution and color mode of image resources. In the association rule layer, semantic association weights are calculated based on co-occurrence frequency, manual annotation, and scene co-occurrence. For example, character models and weapon models that are frequently used together in combat scenes will have correspondingly higher association weights. The context scene layer records in detail the application of resources in different game development stages and different game types.
[0054] When the cross-modal feature extraction module is working, it extracts specific features for different types of resources. The feature mapping unit then maps these features to a unified semantic space through an adversarial generative network. Multiple domain feature encoders process features of different types of resources respectively, and a shared semantic decoder maps them to a unified semantic vector. The domain discriminator optimizes the mapping results through adversarial training, making the semantic vectors of different types of resources comparable.
[0055] In the dynamic classification module, the classification strategy network employs a deep Q-network, taking the semantic vector of the resource and the current classification state as input, and outputting classification path suggestions. The reward function calculates a reward value based on the consistency of the classification results, retrieval efficiency, and adaptability to the development scenario. The network parameters are updated using the reward value, and the classification strategy is continuously optimized by incorporating feedback from developers.
[0056] When processing user queries, the intelligent retrieval module employs an attention mechanism to fuse input features from different modalities, generating a query semantic vector. The semantic expansion unit generates an expanded set of query keywords based on a semantic network, and the retrieval ranking unit comprehensively considers semantic similarity, resource usage frequency, and development scenario matching to rank the search results and return the most relevant resources.
[0057] The self-evolution module's log analysis unit continuously records resource usage, the feedback collection unit collects developers' ratings and corrections, and the network update unit uses incremental learning to periodically update the semantic network's association rules and feature weights, enabling the system to continuously adapt to new requirements.
[0058] For example, during the development of a role-playing game, developers need to retrieve resources related to "fire magic." After receiving the text query, the intelligent retrieval module generates extended terms such as "fire magic" and "fire skills," retrieving relevant special effects and animation resources from the semantic network. Through a cross-modal feature extraction module, the features of these different types of resources are mapped to a unified semantic space, calculating the semantic similarity to the query. This is then combined with the resource's usage frequency and its matching degree in the magic skill development scenario to rank the resources and return the most suitable ones. Simultaneously, the self-evolution module updates the association rules and feature weights of relevant resources based on developer feedback on the search results, improving the accuracy of subsequent searches.
[0059] In summary, the intelligent classification and retrieval system for game resources based on semantic networks provided by this invention can effectively improve the efficiency and accuracy of game resource classification and retrieval, providing strong support for game development.
[0060] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.
Claims
1. A game resource intelligent classification and retrieval system based on semantic networks, characterized in that, include: The resource semantic modeling module is used to construct a multi-dimensional semantic network for game resources. This semantic network includes a resource entity layer, an attribute feature layer, an association rule layer, and a contextual scene layer. The resource entity layer covers entity nodes of 3D models, images, animations, special effects, and code resources. The attribute feature layer includes the visual features, functional features, format features, and creative features of the resources. The association rule layer stores the semantic association weights between entities, which are dynamically adjusted based on co-occurrence frequency and manual annotation. The contextual scene layer records the application context of resources in the game development scenario. A cross-modal feature extraction module is used to extract features from different types of game resources and map them to a unified semantic space, including: For 3D models, extract geometric topological features, material texture features, and skeletal animation features; For image resources, extract color histograms, texture features, and content tag features; Extract time series features, keyframe features, and particle system parameter features from animation and special effects resources; For code resources, functional semantic features and call relationship features are extracted through abstract syntax tree analysis; The dynamic classification module, based on deep reinforcement learning, enables automatic resource classification, including: The classification strategy network takes the association rules of the semantic network as input and outputs resource classification paths; The reward function calculates the reward value based on the consistency of the classification results, retrieval efficiency, and adaptability to the development scenario. The classification results optimization unit optimizes classification strategies through continuous interaction and feedback with developers. The intelligent search module supports multimodal queries and semantically expanded searches, including: The multimodal query parsing unit converts text descriptions, sketch inputs, and example resource inputs into semantic query vectors; The semantic extension unit generates extended sets of synonyms, hyponyms, and related words for query keywords based on the association rule layer of the semantic network; The retrieval and ranking unit sorts resources based on semantic similarity to the query, resource usage frequency, and matching degree with development scenarios. The self-evolution module dynamically updates the association rules and feature weights of the semantic network by monitoring resource usage logs and developer feedback, thereby achieving continuous system optimization.
2. The intelligent classification and retrieval system for game resources based on semantic networks according to claim 1, characterized in that, The cross-modal feature extraction module further includes a feature mapping unit, which uses an adversarial generative network to map features of different types of resources to a unified semantic space. The adversarial generative network includes: Multiple domain feature encoders, each corresponding to different types of game resources; A shared semantic decoder maps domain-specific feature codes to a unified semantic vector; Domain discriminators are used to distinguish which domain a feature vector comes from. Through adversarial training, feature vectors from different domains can be compared in a shared semantic space.
3. The intelligent classification and retrieval system for game resources based on semantic networks according to claim 2, characterized in that, The classification strategy network of the dynamic classification module adopts a deep Q-network. Its state space consists of the feature vector of the current resource and the existing classification path, and its action space consists of possible next-level classification labels. It learns the optimal classification strategy through interaction with the development scenario.
4. The intelligent classification and retrieval system for game resources based on semantic networks according to claim 3, characterized in that, The multimodal query parsing unit of the intelligent retrieval module uses an attention mechanism to fuse input features from different modalities. For text description input, a pre-trained language model is used to extract semantic vectors; for sketch input, contour features and structural features are extracted and converted into semantic vectors; for example resource input, the cross-modal feature extraction module is called to obtain its semantic vector, and then the semantic vectors from different modalities are fused into a unified query vector.
5. The intelligent classification and retrieval system for game resources based on semantic networks according to claim 4, characterized in that, The self-evolution module includes: The log analysis unit records the frequency of resource retrieval, download frequency, usage scenarios, and associated resource combinations; The feedback collection unit obtains feedback information through developers' rating of search results and correction of classification results; The network update unit updates the association rules and feature weights of the semantic network using an incremental learning approach based on log analysis results and feedback information, without needing to retrain the entire model.
6. The intelligent classification and retrieval system for game resources based on semantic networks according to claim 5, characterized in that, The method is as follows: Includes the following steps: S1: Construct a multidimensional semantic network for game resources, the semantic network comprising a resource entity layer, an attribute feature layer, an association rule layer, and a context scene layer; S2: Extract cross-modal features of different types of game resources and map them to a unified semantic space; S3: Based on deep reinforcement learning, game resources are dynamically classified according to the association rules of the semantic network and development scenario information; S4: Receives multimodal queries from users, generates extended query sets through semantic expansion, and retrieves and sorts relevant resources in the semantic network; S5: Dynamically update the semantic network based on resource usage logs and user feedback to enable the system to self-advance.
7. The intelligent classification and retrieval system for game resources based on semantic networks according to claim 6, characterized in that, In step S3, the specific process of dynamic classification is as follows: S31: Initialize the classification policy network and set the initial reward function; S32: Input the semantic vector of the resource and the current classification state into the classification policy network to obtain classification path suggestions; S33: Calculate the reward value based on the consistency of the classification results, retrieval efficiency, and adaptability to the development scenario; S34: Update the parameters of the classification policy network based on the reward value; S35: Receive feedback from developers regarding the classification results and further optimize the classification strategy.