Content generation method based on scene memory map

Abstract

The invention relates to the technical field of computer games, and particularly discloses a content generation method based on a scene memory atlas, which comprises the steps of multi-modal context data acquisition, scene memory atlas management, content adjustment parameter generation and self-adaptive game content generation. The method comprises the following steps: firstly, constructing a dynamic memory graph containing event nodes and associated edges, quantifying player emotion and associating historical experiences, secondly, generating personalized context interpretation and pre-judging emotion influence in combination with a current situation and an activated memory fragment, and formulating a content adjustment strategy, and finally, converting the strategy into specific parameters so as to adjust the content of the player. Game objects, environment, and narrative are reconstructed, generating adaptive content that highly resonates with player emotional states and historical experiences. According to the method, the limitation of traditional single difficulty adjustment is broken through, and the game immersion, the narrative continuity and the individuation depth are remarkably improved through logic association and emotion prediction of the memory graph.

Description

Technical Field

[0001] This invention relates to the field of computer game technology, and more specifically, to a content generation method based on contextual memory graphs. Background Technology

[0002] In the realm of interactive electronic entertainment and computer games, providing players with personalized and engaging gaming experiences is one of the core goals of technological development. Dynamic content generation technology, as a key means to achieve this goal, aims to adjust the game world in real time based on player behavior and state to enhance playability, replay value, and immersion. By procedurally generating or adaptively adjusting challenges, narratives, and environments within the game, developers strive to create a unique gaming journey for each player.

[0003] In existing technologies, methods for achieving adaptive adjustment of game content mainly rely on dynamic difficulty adjustment mechanisms. These methods typically quantify a player's skill level by tracking direct performance data within the game, such as scores, completion times, number of failures, or resource availability. Based on these quantified metrics, the system adjusts enemy attack power and numbers, level layout complexity, or available resources in real time using preset rules or algorithms. The core objective is to match the game difficulty with the player's real-time operational ability to maintain the player's flow experience.

[0004] However, existing technological solutions have significant limitations. First, their perception dimension is singular, focusing only on the player's in-game performance and ignoring the player's real-world emotional reactions, such as tension, excitement, or frustration. Therefore, they cannot comprehensively and accurately understand the player's overall experience. Second, the decision-making basis for content adjustments is isolated and immediate, lacking memory and understanding of the player's long-term gaming experience. The system cannot connect the current situation with the player's specific past successes or failures, resulting in adaptive adjustments that are often mechanical and lack contextual relevance, making it difficult to generate deeply personalized content that resonates with the player's emotions and personal gaming narrative. Summary of the Invention

[0005] In view of this, in order to solve the problems mentioned in the background technology, a content generation method based on contextual memory graph is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides a content generation method based on contextual memory graph, including the following steps: S1, multimodal context data acquisition: acquire the player's real-time multimodal context data.

[0007] S2. Contextual Memory Graph Management: Based on the multimodal context data, update a contextual memory graph containing event nodes and associated edges, and retrieve activated related memory fragments from the contextual memory graph.

[0008] S3. Content adjustment parameter generation: The activated relevant memory fragments and the multimodal context data are fused to generate content adjustment parameters.

[0009] S4. Adaptive Game Content Generation: Obtain basic game materials, and reconstruct the basic game materials according to the content adjustment parameters to generate adaptive game content.

[0010] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention constructs and dynamically maintains a contextual memory map, deeply binding and establishing logical connections between the player's discrete game events and real emotional reactions, thereby achieving a profound modeling of the player's individual game journey. This enables the content generation system to go beyond simple adjustments based on short-term behavior, and instead make decisions based on the player's unique experience, behavioral patterns, and emotional triggers, greatly enhancing the personalization depth of the game content.

[0011] (2) This invention integrates the player's in-game state with external environmental sensor data, and quantifies the player's emotional state in real time by analyzing information such as facial expressions and voice. This combination of internal and external perception enables the system to grasp the player's real experience state more accurately and comprehensively, avoiding misjudgments caused by relying solely on in-game data. As a result, it can generate game content that highly resonates with the player's current emotional state, significantly enhancing the immersion of the game and the authenticity of emotional interaction.

[0012] (3) By comparing and analyzing the current situation with activated historical memory fragments, the system can predict the potential emotional impact of the current event on the player, thereby making proactive and preventative content adjustments rather than passive and delayed responses. This makes the evolution of the entire game experience more logical and narratively coherent, as if the game world can remember and understand the player's past, creating a dynamic evolutionary experience that grows together with the player. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0015] Figure 2 A schematic diagram showing parameter adjustments for the content of this invention.

[0016] Figure 3 This is a schematic diagram illustrating the adaptive game content generation of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 The present invention provides a content generation method based on contextual memory graph, including: S1, multimodal context data acquisition: acquiring the player's real-time multimodal context data.

[0019] In a specific embodiment of the present invention, the specific steps for obtaining the player's real-time multimodal context data include: obtaining external environment sensor data from the external environment.

[0020] Obtain in-game state data from within the game.

[0021] It should be noted that the multimodal context data acquisition step in this method aims to construct a comprehensive and dynamic player state profile. Its execution begins with the parallel acquisition of two heterogeneous data sources. The first data source is external environmental sensor data. Through various sensor devices deployed in the player's physical environment, such as high-definition cameras, microphone arrays, heart rate monitors, or brain-computer interface devices, raw signals reflecting the player's physiological and emotional state are captured in real time. For example, video streams captured by cameras are used for subsequent facial expression analysis, and audio streams captured by microphones are used for speech emotion recognition. The second data source is in-game state data. Through the game engine's built-in application programming interface or logging system, all interactive information of the player in the virtual world is accurately recorded, including but not limited to the character's health, location coordinates, equipment status, current quest progress, and dialogue history with non-player characters.

[0022] The multimodal context data is generated by fusing the external environment sensor data with the in-game state data.

[0023] It should be noted that after acquiring the above two types of data, the data fusion stage is entered to generate unified multimodal context data. This stage is not a simple information splicing, but rather a feature-level fusion model that maps heterogeneous data to a unified semantic space. First, the original external environment sensor data and in-game state data are preprocessed and feature extracted. For example, facial motion unit sequences are automatically extracted from the video stream, Mel-frequency cepstral coefficients are automatically extracted from the audio stream, and quantified key indicators such as game difficulty and player score are extracted from the game data to form their respective feature vectors. Subsequently, a weighted fusion strategy is used to integrate these feature vectors into a high-dimensional context vector, i.e., multimodal context data. This process can be represented by the following formula: ,in, This represents the final generated multimodal context data. and These represent feature vectors extracted from external environment sensor data and in-game state data, respectively. and These are the weighting coefficients for the fusion of external environment sensor data and in-game state data. These two coefficients can be dynamically adjusted according to the current game context to balance the importance of the player's real-world reactions and virtual-world states within the context. For example, at key plot decision points... The weight of this function may be increased to give more weight to the player's genuine emotional response. The fusion function, in one implementation, can be a vector concatenation operation that connects weighted feature vectors into a longer vector, thus preserving information from all dimensions and forming structured multimodal contextual data, providing comprehensive input for subsequent contextual memory graph management.

[0024] This invention acquires real-time multimodal contextual data of players, enabling comprehensive and high-dimensional perception of player states. It not only overcomes the limitations of traditional games that rely solely on in-game input to determine player intent, but also incorporates players' real-world physiological and emotional responses into the analysis. This deep integration of internal and external data sources generates a far more accurate and richer contextual environment than a single data source, allowing subsequent content generation to deeply understand the player's immediate experience. This lays a solid data foundation for truly personalized and emotionally resonant adaptive game content, significantly enhancing the game system's dynamic adaptability and immersive experience.

[0025] S2. Contextual Memory Graph Management: Based on the multimodal context data, update a contextual memory graph containing event nodes and associated edges, and retrieve activated related memory fragments from the contextual memory graph.

[0026] In a specific embodiment of the present invention, the specific steps of updating a context memory graph containing event nodes and associated edges include: identifying the current game event from the multimodal context data.

[0027] It should be noted that the contextual memory graph management step in this method involves in-depth processing and organization of the acquired multimodal contextual data, aiming to construct a dynamic knowledge base reflecting the individual player's experiences and emotional journey. This step first parses the input multimodal contextual data through an event recognition module. This module pre-sets a series of game event templates or uses machine learning classifiers to accurately identify the current game event by analyzing changes in key parameters of in-game state data, such as task status updates, player location entering a specific area, or sudden changes in character attributes.

[0028] The player's emotional response is quantified based on the multimodal context data to obtain an emotional quantification value.

[0029] In a specific embodiment of the present invention, the specific steps of quantifying the player's emotional response based on the multimodal context data to obtain an emotional quantification value include: separating the player's facial expression data from the multimodal context data and analyzing it to generate an emotion index.

[0030] It should be noted that the process of quantifying player emotional responses in this method aims to accurately convert the player's internal emotional state into a numerical form that the system can process, i.e., an emotional quantification value. This process begins by separating a specific data stream from the multimodal context data. First, a facial expression analysis module extracts video sequences containing player facial information from the multimodal context data, i.e., player facial expression data. This module uses a deep learning model, such as a convolutional neural network, to process each frame of the image, identify and track facial key points, and then analyze the combination and intensity of facial action units, ultimately generating a quantified emotion index.

[0031] Player voice data is extracted from the multimodal context data and analyzed to generate intonation features.

[0032] It should be noted that, in parallel, a speech sentiment analysis module separates the player's speech signal, i.e., player speech data, from the multimodal context data. This module uses signal processing techniques to extract a series of acoustic features reflecting emotional states from the speech data, such as the mean and range of the fundamental frequency, the intensity and distribution of energy, speech rate, and Mel-frequency cepstral coefficients, which together constitute intonation features. Similar to facial expression analysis, these intonation features are then input into a classification or regression model, outputting a sentiment vector consistent with the dimension of the sentiment index.

[0033] The emotion index and the tone feature are combined to generate the emotion quantification value.

[0034] It should be noted that, ultimately, to obtain a stable and comprehensive emotional quantification value, the system employs a multimodal fusion model to integrate emotional information from both facial expressions and speech. This fusion process is not a simple numerical addition, but rather uses a weighted decision mechanism to balance the contributions of different modalities. This process can be represented by the following formula: In this formula, This refers to the final generated sentiment quantification value, which is a multi-dimensional vector. This represents a normalized emotion index vector generated by analyzing players' facial expression data. This represents the emotion vector corresponding to the intonation features generated by analyzing player voice data and also after normalization; the normalization operation ensures that the two vectors are in the same numerical range and scale, and are additive; and These represent the fusion weights for facial expression modality and speech modality, respectively, and their sum is 1. These weights can be dynamically adjusted based on signal quality or contextual environment, for example, by decreasing them in low light conditions. Reduce noise in noisy environments The emotional quantification value calculated using this formula accurately captures the overall emotional state of the player.

[0035] The current game event is added as a new event node to the context memory graph, and the emotion quantification value is associated with the new event node, thereby updating the context memory graph.

[0036] It's important to note that, next, the system encapsulates the identified current game event and its corresponding emotional quantification value into a new event node and adds it to the contextual memory graph. The contextual memory graph is a graph-structured data model where nodes represent experienced events and edges represent relationships between events. In this step, the new event node not only contains the semantic information of the event itself but also marks the player's emotional quantification value during the experience, enriching the memory record with emotional depth.

[0037] After updating the aforementioned contextual memory map, the structure of the updated contextual memory map also needs to be dynamically evolved.

[0038] In a specific embodiment of the present invention, the specific steps for dynamically evolving the structure of the updated context memory map include: after adding the new event node, analyzing the causal and temporal relationships between the new event node and the existing event nodes.

[0039] It's important to note that the dynamic evolution of the episodic memory graph structure in this method is a process of organically integrating new information into the existing memory network and revealing its deep connections. When a new event node is added to the episodic memory graph, the system does not treat it as an isolated data point but immediately launches an association analysis engine. This engine first performs temporal relationship analysis, automatically establishing a basic association edge between the new event node and the event node that occurred immediately before it by comparing the timestamps of the new event node with those of existing event nodes in the graph. This constitutes the linear narrative skeleton of the player's experience. More importantly, the system performs in-depth causal relationship analysis. This analysis employs a hybrid strategy. On the one hand, it utilizes a knowledge rule base pre-set in the game logic to identify explicit causal chains, such as the player using a specific key inevitably leading to a door being opened. On the other hand, it uses a statistical learning model to uncover potential, non-explicit causal relationships by analyzing the co-occurrence frequency, time intervals, and accompanying changes in emotional quantification values ​​of events. For example, if a player's emotional quantification value shows a significant negative shift after encountering a certain type of enemy multiple times, the system will infer that there is a strong correlation between that type of enemy and the player's negative emotions.

[0040] Based on the causal and temporal relationships, establish or adjust the associated edges and their weights between the event nodes to generate updated association relationships.

[0041] Based on the updated associations, the structure of the contextual memory map is dynamically evolved.

[0042] It should be noted that, based on the analysis results of temporal and causal relationships, the system will establish or adjust the associated edges and their weights between event nodes, generating updated associations. The edge weight is a quantitative indicator reflecting the tightness of the association between two events, and its calculation can be represented by a comprehensive model: In this formula, Represents from event node To the event node The weight of the associated edges, It is an event With the event The causal relationship strength score between the two relationships is output by the causal analysis engine, and its value range has been normalized. It is an event With the event The temporal proximity score between the intervals is calculated, with shorter time intervals resulting in higher scores. This score is also normalized. Representative from the event To the event The change in the emotional quantification value, function By taking the absolute value through mathematical transformation, this change is mapped to a non-negative value, reflecting the impact of emotional fluctuations on the strength of the correlation between events. , , It is an adjustable hyperparameter used to balance the strength of causality, temporal proximity, and the proportion of sentiment factors in the weights of event association edges. + + =1. In this way, each associated edge acquires a precise weight that reflects its inherent logic and emotional importance. The entire structure of the episodic memory graph dynamically evolves through this cycle of continuously adding nodes, analyzing relationships, and finely adjusting edge weights.

[0043] It should also be noted that the normalization process for the causal relationship strength score is as follows: The causal analysis engine analyzes events... With the event The causal relationship between them is determined, and a raw causal relationship strength score is output. The range of the raw causal relationship strength score extracted from the database is... It can be mapped to the interval [0,1] through a linear transformation, and the specific formula is as follows: The normalization process for temporal proximity scores is as follows: the time interval has a definite minimum value. and maximum value You can first set the time interval Through formula Mapping to the [0,1] interval, the longer the time interval, the closer the normalized value is to 1; to satisfy the requirement that the shorter the time interval, the higher the score, a linear transformation is then applied. This way, the score is close to 1 when the time interval is at its minimum and close to 0 when the time interval is at its maximum.

[0044] In a specific embodiment of the present invention, hyperparameters , , These parameters are used to balance the strength of causal relationships, temporal proximity, and the proportion of emotional factors in the weights of event-related edges. Their specific values ​​need to be determined through experimental analysis. By comparing indicators such as player immersion and replay rate under different parameter combinations through A / B testing, the fixed value combination that can make the game content personalized is optimal can be obtained: α=0.5, β=0.3, γ=0.2.

[0045] This invention achieves deep modeling of personalized gaming experiences by constructing and dynamically maintaining a contextual memory graph with accompanying emotional information. It combines discrete game events with the player's emotional responses and further organizes them into a memory structure with internal logical connections. This technology allows the system to understand the player by utilizing a long-term, dynamic memory bank rich in emotional detail, rather than simply reacting to the current stimulus. This provides crucial decision-making basis for subsequently generating adaptive game content that truly touches the player's heart and resonates with their past experiences, greatly enhancing the coherence, personalization, and emotional immersion of the gaming experience.

[0046] In a specific embodiment of the present invention, the specific steps of retrieving activated relevant memory fragments from the contextual memory map include: mapping the multimodal context data into semantic context signals.

[0047] It should be noted that the process of retrieving activated memory fragments from the episodic memory map in this method aims to use the current context as a clue to awaken the player's most relevant past experiences. This process first transforms the real-time acquired multimodal context data through a pre-trained semantic encoding model, generating a compact and highly semantically generalized vector—the semantic context signal. This model, such as a deep neural network, can understand the deeper meanings of different feature combinations in the multimodal context data and map them into a unified semantic vector space, thus abstracting the raw, high-dimensional, heterogeneous data into a "meaning" vector representing the current core context.

[0048] The semantic contextual signals are used as retrieval clues to query the contextual memory map.

[0049] It should be noted that, next, the system uses this semantic context signal as a retrieval clue to perform an efficient query on the entire context memory graph. This query is not a simple traversal matching, but is based on an activation diffusion mechanism. First, the semantic similarity between the semantic context signal and each event node in the graph is calculated to obtain the initial activation value of each node. This similarity calculation can be expressed by the following formula: In this formula, Represents event nodes The initial activation value, It is a semantic context signal vector generated in the current context. It is an event node The semantic vectors generated synchronously during storage ensure and Within the same semantic space, they are comparable. It is a similarity calculation function, such as cosine similarity, whose output value reflects the current context and historical events. The degree of semantic fit. After obtaining the initial activation values, these activation energies diffuse along the weighted association edges in the episodic memory graph, activating other nodes closely connected to the highly activated nodes. This diffusion process simulates the characteristics of human memory association, enabling the discovery of historical events that are indirectly but significantly related to the current context.

[0050] Extract memory information that matches the retrieval clues to form the activated relevant memory fragments.

[0051] It should be noted that after the activation diffusion process stabilizes or reaches a preset number of iterations, the system extracts the group of event nodes with the highest activation values ​​from the graph. These nodes represent the core memory information that best matches the current context. However, the activated related memory fragments are not merely these isolated nodes, but rather local subgraphs constructed around these core nodes. The system extracts these high-activation event nodes themselves, including the event content they record, the associated emotional quantification values, and the high-weighted association edges between them. This series of information collectively constitutes one or more structured memory sets rich in context and emotional detail, namely the activated related memory fragments, providing in-depth background interpretation for subsequent content adjustments.

[0052] This invention achieves deep memory retrieval that goes beyond surface information matching by semanticizing the current context and combining it with associative retrieval based on the topological structure of the memory graph. It can accurately recall past experiences that are most relevant semantically, causally, or emotionally, based on the player's current behavior and emotional state. This technical effect allows the system to understand the unspoken implications of the current context, linking the player's immediate experience with their personal history. This provides a crucial, personalized historical reference frame for generating adaptive content with continuity and emotional resonance, thereby greatly enhancing the depth and personalization of the gaming experience.

[0053] Please see Figure 2 S3. Content adjustment parameter generation: The activated relevant memory fragments and the multimodal context data are fused to generate content adjustment parameters.

[0054] In a specific embodiment of the present invention, the specific steps of fusing the activated relevant memory fragments with the multimodal context data to generate content adjustment parameters include: using the activated relevant memory fragments as the interpretation background, analyzing the multimodal context data, and generating personalized context interpretation.

[0055] In a specific embodiment of the present invention, the specific steps of analyzing the multimodal context data to generate personalized context interpretations include: comparing the contextual features in the multimodal context data with the historical context in the activated relevant memory fragments.

[0056] It's important to note that the process of generating personalized contextual interpretations in this method is a reasoning process that deeply correlates the current situation with the player's personal history to predict their emotional responses. The core of this process begins with a meticulous comparison of contextual features in multimodal contextual data with historical contexts in activated relevant memory fragments. This comparison is not a simple text or numerical match, but rather performed at a semantic level. The system calculates the similarity between key elements in the current situation, such as scene type, encountered non-player characters, and task objectives, and historical event elements recorded in activated memory fragments. This comparison aims to identify the extent to which the current situation reproduces one or more important past experiences of the player.

[0057] It should also be noted that in the content generation method based on contextual memory graphs, when the system performs similarity calculations to identify the degree of matching between the current context and the player's historical experience, it first maps the multimodal features of the current context, such as scene type, non-player character attributes, and task objective descriptions, and the historical event features in the activated memory fragments to a unified semantic vector space through a semantic encoding model, generating structured semantic vectors. Subsequently, a cosine similarity algorithm is used to calculate the cosine value of the angle between the current context vector and the historical event vector in space. The closer this value is to 1, the higher the semantic matching degree. For example, if the current scene type is "cave exploration" and the task objective is "treasure hunting," the system will extract the semantic features of these key elements and compare them with the historical event vectors in the memory fragments that also involve "cave" and "treasure hunting." The cosine value is used to quantify the degree of similarity between the two in deep semantic aspects such as narrative logic and emotional connection, rather than relying solely on surface text or numerical matching.

[0058] By combining the emotional information associated with the activated relevant memory fragments, the potential emotional impact of the situational features on the player is assessed, and an emotional impact prediction is formed.

[0059] It's important to note that after identifying similar historical contexts, the system further integrates the emotional information associated with activated memory fragments—that is, the emotional quantification values ​​stored alongside those historical events—to assess the potential emotional impact of the current context on the player. This step is crucial for forming an emotional impact prediction. The system analyzes what types and intensities of emotions the player primarily experienced in similar past contexts. For example, if activated memory fragments show that the player's emotional quantification values ​​were consistently high arousal and negative valence in all historical contexts involving battles against a specific boss, the system predicts that facing that boss again is highly likely to trigger anxiety or frustration in the player. This prediction process can be modeled as follows: In this formula, It generates a predictive sentiment vector, which is used to predict the impact of emotions. This represents a contextual feature vector extracted from the current multimodal context data. The first of the relevant memory fragments representing activation Feature vectors of historical contexts Is the current situation and the first The semantic similarity scores of each historical context are used for weighting. It is the first A vector of sentiment quantification values ​​associated with each historical context, a function It is an aggregation function that represents a weighted average calculation. It integrates the emotional contributions of all relevant historical memories to form a comprehensive prediction of the player's possible emotional state in the current situation.

[0060] It should also be noted that the current situation is different from the first... The semantic similarity score of each historical context needs to be simplified through structured feature matching or lightweight semantic modeling: first, the semantic similarity score between the current context and the first historical context needs to be calculated. Key features are extracted from historical contexts, such as scene type, core NPCs, task objectives, and emotional tags. These features are then discretized into numerical or label forms. For example, the scene "forest" is encoded as 0, and "cave" as 1. Core NPCs are divided into limited categories based on profession, faction, or function, such as guards and merchants, with "guards" encoded as 0 and "merchants" as 1. Task objectives are extracted based on core actions, such as defeating and collecting, with "defeat" encoded as 0 and "collect" as 1. Emotional tags are based on basic emotions, such as anger and calmness, with "anger" encoded as 0 and "calm" as 1. Next, a weighted cosine similarity method is used to sum the similarities of each feature, with the weights adjustable according to the game type. Finally, a normalized scalar score is output to quantify the degree of semantic matching between the two.

[0061] Based on the predicted emotional impact, the personalized contextual interpretation is generated.

[0062] It's important to note that, ultimately, based on this accurate prediction of emotional impact, the system integrates and compares the situational similarities discovered during the comparison process to generate a structured or textual description—the personalized contextual interpretation. This interpretation not only includes an objective description of the current situation, but more importantly, it clearly identifies which past memories the situation might trigger in the player, and what emotional responses these memories might induce. For example, one interpretation might be: "The current environment is highly similar to the scenario of the player's previous failed mission, which is expected to trigger mild tension. However, given the player's swift and successful revenge after the previous failure and the positive emotional feedback received, there is also the potential to motivate success in this challenge."

[0063] Based on the personalized context interpretation, a content adjustment strategy is determined.

[0064] It's important to note that based on this personalized contextual interpretation rich in historical insights, the system then activates a decision engine to determine an optimal content adjustment strategy. This decision engine can be based on a predefined rule base or a policy network trained through reinforcement learning. For example, if the personalized contextual interpretation shows that the current game situation is highly similar to a past memory that caused the player strong frustration, and the current multimodal contextual data also shows signs of negative emotion, the decision engine will determine a content adjustment strategy aimed at reducing difficulty, providing encouragement, or diverting attention. Conversely, if the interpretation finds that the player has shown positive emotions of enjoying challenges in similar situations, the strategy might be to moderately increase the challenge to maintain the flow experience.

[0065] The content adjustment strategy is converted into specific numerical values ​​to generate the content adjustment parameters.

[0066] It's important to note that the final step is to transform this abstract content adjustment strategy into a set of specific, quantifiable content adjustment parameters. This transformation process is handled by a parameter mapping module. Based on the selected strategy, this module generates a series of numerical values ​​that directly correspond to various controllable variables within the game engine. For example, the strategy of "reducing difficulty" might be translated into a series of specific numerical changes, such as reducing the attack power of non-player characters, increasing the drop rate of healing items, or extending quest time limits. This process can be represented by a function: In this formula, This represents the final set of content adjustment parameters; it may be a structured data set containing multiple key-value pairs. It is the identifier for the content adjustment strategy determined by the decision engine. This represents the current personalized context interpretation, providing a basis for fine-tuning the parameters. It is a mapping function that encapsulates the logic for translating abstract strategies into concrete numerical values. For example, it will... The intensity of negative emotions experienced by players determines the specific percentage by which non-player characters' attack power needs to be reduced, thus achieving fine-grained adjustments. This set of content adjustment parameters is ultimately output to guide the next step of adaptive game content generation.

[0067] It should also be noted that the identifier is a unique label generated by the decision engine from a predefined strategy library or dynamic mapping after discretization encoding or rule tree matching based on multi-dimensional features such as player status, environmental characteristics and historical behavior data in the game context. It is used to quickly index the corresponding content adjustment strategy and realize real-time differentiated adaptation of the game experience.

[0068] This invention achieves a complete closed loop from perception to decision-making to command execution by fusing activated memory fragments with current context data. This allows game content adjustments to no longer be based on generalized player models or simple instantaneous reactions, but rather on a deep understanding of the unique experiences and emotional patterns of individual players. This technological effect ensures that the generated content adjustment parameters are highly personalized and context-adaptive, capable of anticipating and proactively regulating the player's emotional trajectory, avoiding the repetition of negative experiences, and reinforcing positive experiences, thereby creating a game experience that deeply resonates with the player and evolves dynamically.

[0069] Please see Figure 3 S4. Adaptive game content generation: Obtain basic game materials, and reconstruct the basic game materials according to the content adjustment parameters to generate adaptive game content.

[0070] In a specific embodiment of the present invention, the specific steps of reconstructing the basic game materials according to the content adjustment parameters to generate adaptive game content include: parsing the content adjustment parameters to obtain adjustment instructions for game objects, game environment and game narrative.

[0071] It's important to note that the step of generating adaptive game content in this method is the final execution stage of the entire adaptive loop. Its purpose is to materialize abstract content adjustment parameters into changes in the player's perceptible game experience. This process is driven by a content reconstruction engine, which first receives and parses the content adjustment parameters generated in the previous step. These parameters are not single values, but a structured set of instructions that clearly indicates which basic game assets need to be modified to what extent. The parsing module then breaks down the parameters into specific adjustment instructions for three core dimensions: game objects, game environment, and game narrative, according to a predefined format.

[0072] According to the adjustment instructions, modify the attribute parameters of the game objects, adjust the rendering parameters of the game environment, and change the script parameters of the game narrative.

[0073] It's important to note that after receiving the decomposed adjustment instructions, the content reconstruction engine will operate on the game's basic assets in parallel or sequentially. For adjustment instructions targeting game objects, the engine directly accesses the corresponding object's database or real-time instance within the game, modifying its attribute parameters. For example, an instruction might require reducing the health attribute parameter of a specific non-player character by 20%, or increasing its attack speed attribute parameter by 10%. For adjustment instructions targeting the game environment, the engine interacts with the game's rendering and audio subsystems to adjust rendering parameters. For example, an instruction might require increasing the global illumination intensity of the scene to create a brighter, more positive atmosphere, or changing the background music's script parameters to switch to a more soothing track. For adjustment instructions targeting the game's narrative, the engine intervenes in the game's scripting system, altering the narrative's script parameters. This might include activating a hidden dialogue branch to provide players with additional clues, or relaxing the time limit for a task to reduce player stress.

[0074] The modified, adjusted, and altered game objects, game environment, and game narrative are combined to generate the adaptive game content.

[0075] It's important to note that, ultimately, all modified, adjusted, and altered game objects, game environments, and narrative elements will be uniformly loaded and processed in the game engine's next rendering or logic update cycle. The game engine seamlessly combines these updated basic assets to generate a new, complete game world state. This entirely new state presented to the player is the adaptive game content mentioned above. This process is continuous and dynamic, ensuring that the game content can be reconstructed in real time based on constantly updated parameters.

[0076] This invention provides a specific and efficient execution path that transforms the system's deep understanding of the player's state into comprehensive and coordinated changes in the game world. It transcends the single-dimensional difficulty adjustments of traditional games, achieving a three-dimensional and refined shaping of the player experience through synchronized manipulation of game objects, environmental atmosphere, and narrative rhythm. This systematic reconstruction capability ensures the naturalness and immersion of adaptive adjustments, avoiding abrupt or jarring changes, ultimately realizing a truly dynamic and personalized game world that resonates with the player's emotional state and the game's rhythm.

[0077] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A content generation method based on contextual memory graphs, characterized in that, Includes the following steps: S1. Multimodal context data acquisition: Acquire the player's real-time multimodal context data; S2. Contextual memory graph management: Based on the multimodal context data, update a contextual memory graph containing event nodes and associated edges, and retrieve the activated related memory fragments from the contextual memory graph; S3. Content adjustment parameter generation: The activated relevant memory fragments and the multimodal context data are fused to generate content adjustment parameters; S4. Adaptive Game Content Generation: Obtain basic game materials, and reconstruct the basic game materials according to the content adjustment parameters to generate adaptive game content.

2. The content generation method based on contextual memory graphs according to claim 1, characterized in that: The specific steps for obtaining the player's real-time multimodal context data include: Acquire external environment sensor data from the external environment; Obtain in-game state data from within the game; The multimodal context data is generated by fusing the external environment sensor data with the in-game state data.

3. The content generation method based on contextual memory graphs according to claim 1, characterized in that: The specific steps for updating a context memory graph containing event nodes and associated edges include: Identify the current game event from the multimodal context data; The player's emotional response is quantified based on the multimodal context data to obtain an emotional quantification value; The current game event is added as a new event node to the context memory graph, and the emotion quantification value is associated with the new event node, thereby updating the context memory graph; After updating the aforementioned contextual memory map, the structure of the updated contextual memory map also needs to be dynamically evolved.

4. The content generation method based on contextual memory graphs according to claim 3, characterized in that: The specific steps for quantifying the player's emotional response based on the multimodal context data to obtain the emotional quantification value include: Player facial expression data is extracted from the multimodal context data and analyzed to generate an emotion index; Player voice data is extracted from the multimodal context data and analyzed to generate intonation features; The emotion index and the tone feature are combined to generate the emotion quantification value.

5. The content generation method based on contextual memory graphs according to claim 3, characterized in that: The specific steps for dynamically evolving the structure of the updated episodic memory map include: After adding the new event node, analyze the causal and temporal relationships between the new event node and the existing event nodes; Based on the causal and temporal relationships, establish or adjust the associated edges and their weights between the event nodes to generate updated association relationships; Based on the updated associations, the structure of the contextual memory map is dynamically evolved.

6. The content generation method based on contextual memory graphs according to claim 1, characterized in that: The specific steps for retrieving activated relevant memory fragments from the episodic memory map include: The multimodal context data is mapped into semantic context signals; The semantic contextual signals are used as retrieval clues to query the contextual memory map; Extract memory information that matches the retrieval clues to form the activated relevant memory fragments.

7. The content generation method based on contextual memory graphs according to claim 1, characterized in that: The specific steps for fusing the activated relevant memory fragments with the multimodal context data to generate content adjustment parameters include: Using the activated relevant memory fragments as the interpretation background, the multimodal context data is analyzed to generate personalized context interpretations; Based on the personalized contextual interpretation, a content adjustment strategy is determined; The content adjustment strategy is converted into specific numerical values ​​to generate the content adjustment parameters.

8. The content generation method based on contextual memory graphs according to claim 7, characterized in that: The specific steps for analyzing the multimodal context data and generating personalized context interpretations include: The contextual features in the multimodal contextual data are compared with the historical context in the activated relevant memory fragments; By combining the emotional information associated with the activated relevant memory fragments, the potential emotional impact of the situational features on the player is assessed, and an emotional impact prediction is formed; Based on the predicted emotional impact, the personalized contextual interpretation is generated.

9. The content generation method based on contextual memory graphs according to claim 1, characterized in that: The specific steps for reconstructing the basic game assets based on the adjusted parameters to generate adaptive game content include: The parameters of the content are analyzed to obtain adjustment instructions for game objects, game environment, and game narrative; According to the adjustment instructions, modify the attribute parameters of the game objects, adjust the rendering parameters of the game environment, and change the script parameters of the game narrative. The modified, adjusted, and altered game objects, game environment, and game narrative are combined to generate the adaptive game content.

Images