A situation-reasoning one-map-based character decision pattern analysis method and device

Abstract

The present application relates to the technical field of artificial intelligence decision analysis and behavior modeling, in particular to a kind of situation-reasoning one figure-based character decision mode analysis method and equipment, including the construction decision mode situation-reasoning one figure, the extraction and quantization of decision influencing factors and the extraction and quantization steps of decision influencing factors, in the execution above-mentioned steps, first, the structure of situation-reasoning one figure is used, and massive unstructured open source data is converted into computable, situation-reasoning one figure that can be deduced;Then, according to the decision theory, the internal psychological cognitive characteristics of the character are quantified, and they are used as the key constraint condition of reasoning network;Finally, the decision mode neural network is constructed, and the decision process of "perception stimulation-cognitive screening-logical reasoning" of the character under a certain situation is simulated.The present application realizes the deep integration and accurate analysis from external objective situation to internal subjective cognition.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence decision analysis and behavior modeling technology, specifically to a method and device for analyzing human decision-making patterns based on a situation-reasoning diagram. Background Technology

[0002] With the rapid development of information technology, a large amount of information related to the decision-making and activities of key figures has been accumulated in open-source channels on the Internet. In-depth mining and systematic analysis of this open-source information has become an important path to reveal the potential decision-making behavior patterns of key figures.

[0003] Currently, decision-making pattern analysis is mainly achieved through two types of methods: one is qualitative or semi-quantitative analysis based on expert experience and classical theories, and the other is data-driven intelligent analysis. The former relies on classical theoretical frameworks such as expected utility theory, bounded rationality theory, and game theory, attempting to formally describe the uncertainty in the decision-making process through mathematical models; the latter leverages the advantages of big data and artificial intelligence technologies, introducing neural networks, machine learning, and other technical means to conduct intelligent analysis of decision-making patterns. Essentially, the core of key personnel decision-making pattern analysis is to analyze the complex mapping relationship between decision-making patterns and open-source information. Its core objective is to systematically mine dispersed decision-making influencing factors from massive, sparse, and scattered open-source information, while also achieving a deep integration of external objective situations and internal subjective reasoning.

[0004] However, both existing approaches have significant limitations: Approaches based on expert experience and classical theories rely excessively on predefined rule frameworks and researchers' subjective judgments, making it difficult to efficiently process the massive, fragmented, and unstructured data in the internet environment. This results in a lag and bias in the perception of the decision-making situation, failing to meet the demands of the big data era for rapid, comprehensive perception of complex dynamic environments. Data-driven intelligent analysis methods often simplify decision analysis into end-to-end prediction tasks. While they may perform well in data fitting accuracy, they severely neglect modeling the internal reasoning mechanisms of the decision-making process. This "black box" model, lacking guidance from classical decision theory, struggles to explain the logical motivations behind decision-making behavior, ultimately leading to a disconnect between the decision-making situation and the reasoning process.

[0005] It is evident that existing technologies struggle to achieve a dynamic representation of the objective environment and interpretable reasoning of subjective thought processes from a unified perspective. Currently, research on the analysis of human decision-making patterns has undergone two major developmental stages, corresponding to the aforementioned qualitative or semi-quantitative analysis methods based on expert experience and classical theories, and data-driven intelligent analysis methods. The following section elaborates on the current research status of these two approaches.

[0006] Qualitative or semi-quantitative analysis methods based on expert experience and classical theories are employed, building upon the fundamental theories of behavioral economics, cognitive psychology, and management. The foundational work can be traced back to the expected utility theory proposed by Von Neumann and Morgenstern. This theory defines decision-makers as strictly "rational individuals" who pursue utility maximization under risk conditions, laying the core theoretical framework for subsequent research on decision-making models. Subsequently, Herbert Simon's theory of bounded rationality broke through the limitations of the "rational man" assumption, clarifying that decision-makers in real-world scenarios are constrained by both limited cognitive resources and environmental uncertainty, often pursuing a "satisfactory solution" that meets their aspirational level rather than the globally optimal solution. This theory provides more practical theoretical support for decision-making model analysis.

[0007] Based on this, research in this field has gradually diverged into two main directions:

[0008] Firstly, the formal construction and refinement of mathematical models. By introducing tools such as fuzzy mathematics and probability theory to address uncertainty in decision-making environments, representative achievements include fuzzy decision theory proposed by Bellman et al., and the sequential decision model proposed by Wald. These two models respectively improve the mathematical description of the decision-making process from the dimensions of fuzziness and dynamism. In recent years, Zhang et al. have further expanded this research direction, developing a multi-granularity three-branch decision model that integrates regret theory and incomplete T-spherical fuzzy sets, effectively solving the multi-criteria decision-making problem in complex management scenarios.

[0009] Secondly, the analysis of the psychological characteristics and behavioral logic of decision-makers. This direction focuses on analyzing the behavioral motivations and emotional preferences of decision-makers from a psychological perspective. Prospect theory proposed by Kahneman et al. is highly representative, systematically revealing the irrational behavioral characteristics of decision-makers under risk conditions and providing important theoretical support for explaining decision bias. In addition, satisfactory decision theory and case-based decision theory are also widely used in decision pattern analysis to accurately describe the internal logic and behavioral processes by which decision-makers make decisions based on past experience.

[0010] With the massive accumulation of open-source data on the internet and the iterative upgrades of artificial intelligence technology, methods for analyzing key figures' decision-making patterns are gradually shifting towards data-driven intelligent analysis. These methods do not require explicitly predefined rule frameworks; their core lies in utilizing intelligent algorithms such as neural networks and multi-agent systems to directly learn and extract the decision-making patterns of key figures from massive amounts of open-source data, thus possessing a stronger ability to adapt to large datasets.

[0011] Existing research findings include: Mahmud et al. used a multi-stage hybrid approach to study the collaborative mechanism between humans and robo-advisors in investment decision-making. By identifying different user types and advisor roles, they constructed a typology of user acceptance based on facilitating and inhibiting factors, providing a reference for pattern analysis in collaborative decision-making scenarios; Guleria and Sood proposed a career counseling framework integrating interpretable artificial intelligence and machine learning, recommending career paths by analyzing the academic attributes and employability characteristics of decision-makers, with the Naive Bayes algorithm demonstrating excellent performance in decision prediction accuracy; Meske and Bunde studied the interaction design principles of an AI-based hate speech detection decision support system. Empirical results show that the interpretability of the system has a significant impact on user trust, mental model construction, and information perception, providing a practical basis for the design of interpretable decision analysis systems; Sun et al. proposed an interpretable high-risk decision support system called CDFS for credit default prediction, which integrates "human-machine collaboration" (human-machine collaboration). The "loop" principle and advanced feature selection technology effectively adapt to the decision analysis needs of high-risk scenarios in the financial lending field.

[0012] In summary, existing methods for analyzing decision-making patterns can be categorized into two main types: qualitative or semi-quantitative analysis based on expert experience and classical theories, and data-driven intelligent analysis. However, both have significant shortcomings in practical applications. Methods based on expert experience and classical theories heavily rely on predefined knowledge rules and researchers' subjective judgments, making it difficult to efficiently mine and utilize large-scale open-source data scattered across the internet, and they are ill-suited to complex and dynamic decision-making environments. On the other hand, data-driven intelligent analysis methods often simplify decision-making pattern analysis into end-to-end prediction tasks, neglecting to model the internal reasoning mechanisms of the decision-making process. This "black box" approach, lacking guidance from classical decision theory, fails to reveal the core logical motivations behind key figures' decision-making behaviors and cannot meet the demands for precise and interpretable decision-making pattern analysis. Summary of the Invention

[0013] In view of this, this invention proposes a method and device for analyzing character decision-making patterns based on a situation-reasoning unified graph. First, it utilizes the structure of the situation-reasoning unified graph to transform massive amounts of unstructured open-source data into a computable and inferable situation-reasoning unified graph. Then, it quantifies the character's internal psychological cognitive characteristics based on decision theory and uses them as key constraints for the reasoning network. Finally, it constructs a decision-making pattern neural network to simulate the character's "perceptual stimulus-cognitive selection-logical reasoning" decision-making process under specific situations. This achieves deep integration and precise analysis from external objective situation to internal subjective cognition.

[0014] The specific technical solution of the present invention is as follows:

[0015] A method for analyzing character decision-making patterns based on a situation-reasoning diagram includes the following steps:

[0016] S1. Constructing a Decision-Making Pattern-Reasoning Diagram: From open-source text on the internet, entities and relations are extracted using a general information extraction model and prompt words; a time alignment algorithm is used to attach time information in the text to corresponding semantic triples, generating quadruples containing subject entity, object entity, relation, and timestamp. This allows for the construction of a dynamic knowledge graph; the extracted relationship types include affective relationships used to characterize attitude tendencies.

[0017] S2. Extraction and Quantification of Decision-Making Influencing Factors: Based on the aforementioned dynamic knowledge graph, environmental influencing factors and personal influencing factors are extracted and quantified respectively; wherein:

[0018] The extraction of environmental influencing factors includes: taking the decision event entity as the anchor point, backtracking a preset time window, extracting the nodes and edges connected to the anchor entity within the window to form an event background subgraph; filtering based on the frequency of the entity's occurrence in historical decision events to obtain a candidate entity set; and using a large language model to classify the candidate entities into stakeholder groups, others' attitudes, cultural backgrounds, or related event categories according to predefined prompts.

[0019] The aforementioned personal influencing factors include at least personal abilities, value orientations, and team relationships, and their quantification methods are as follows:

[0020] Individual ability: Determining the weight of corresponding events based on the structural complexity of the event background subgraph. , weight Attitude changes among key stakeholders that characterize the social effects of decisions The weighted sum is used to obtain the individual's overall ability score. ,in The number of events that trigger character decisions;

[0021] Value Orientation: By constructing and analyzing the individual value orientation tensor The calculations show that the tensor reflects the distribution of a person's emotional inclination towards the target entity;

[0022] Team relationships: The net semantic density of the network of team member interaction subgraphs with assigned emotional weights is calculated and normalized using the Sigmoid function to obtain a quantitative value that characterizes the closeness of team relationships.

[0023] S3. Character Decision-Making Pattern Modeling and Analysis: Construct a character decision-making pattern neural network containing a screening layer and a fusion reasoning layer; use the quantified personal influencing factors as the gate threshold of the screening layer to perform cognitive filtering on the input environmental influencing factors; the fusion reasoning layer performs nonlinear fusion on the filtered effective information and finally outputs the probability of the decision occurring.

[0024] Furthermore, in step S1, the emotional relationship includes at least one of support, opposition, condemnation, and appreciation.

[0025] Furthermore, in step S1, the time alignment algorithm specifically includes:

[0026] Extract time information as a relation type to obtain time triples;

[0027] Based on the principles of entity consistency and text proximity, the time value in the time triple is attached to the corresponding semantic triple to synthesize the quadruple.

[0028] Furthermore, in step S2, the predefined categories of environmental factors include: stakeholder groups, attitudes of others, cultural background, and related events.

[0029] Furthermore, in step S2: the event weights The calculation formula is:

[0030] ;

[0031] in, and Let N be the number of nodes and edges in the event background subgraph corresponding to the i-th decision time, respectively, and N be the total number of events. The change in attitude ratings towards individuals before and after a decision-making event.

[0032] The personal value orientation tensor Calculated using the following formula:

[0033] ;

[0034] in, For the physical form of a person, For the target entity, For emotional relationships.

[0035] In the quantification method of team relationships, the normalization process is performed according to the following formula:

[0036] ;

[0037] in, For the Sigmoid function, For the number of team members, Let represent the net interaction strength between members i and j. Positive relationships are denoted as positive weights, and negative relationships as negative weights.

[0038] Furthermore, the personal influencing factors also include leadership relationships, which are quantified as follows:

[0039] Construct an information propagation probability matrix P based on the team member interaction subgraph, where , For nodes The degree;

[0040] Simulate information through random walk within a finite number of steps. The internal dissemination process is characterized by the proportion of nodes covered by information and the speed of dissemination, which reflects the leadership style.

[0041] Furthermore, in S3, the filtering layer includes a fixed bias set according to the individual influencing factor vector to simulate the threshold psychological mechanism.

[0042] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements the aforementioned person decision-making pattern analysis method based on a situation-reasoning diagram.

[0043] This invention proposes a situation-reasoning single-map-based method for analyzing human decision-making patterns. Guided by decision theory, it extracts key elements from massive, sparse open-source information, analyzes the relationships between these elements, and simulates a person's cognitive filtering and deep reasoning process of external environmental information based on personal traits. Ultimately, it outputs decision prediction results, achieving accurate mining and representation of human decision-making patterns, and can achieve the following beneficial effects:

[0044] 1. This invention constructs a complete analytical framework from open source information to human decision-making patterns. In view of the characteristics of massive, sparse and scattered open source information on the Internet, it introduces a situation-inference diagram for representation, mines key decision-making elements, and analyzes human decision-making patterns from the perspectives of environmental factors and personal factors, thereby revealing the human decision-making logic and behavioral patterns hidden behind complex data.

[0045] 2. Guided by decision theory, this invention integrates the analytic hierarchy process (AHP) into a neural network structure. By freezing personal factors as biases, it forces the model to explain decision-making behavior based on the individual's real traits (such as historical preferences and abilities), which conforms to the stimulus-response threshold mechanism of cognitive psychology and enables the modeling of individual decision-making patterns.

[0046] 3. This invention utilizes the structural and semantic information of a situation-reasoning diagram to achieve a quantitative representation of environmental and personal influencing factors in a person's decision-making process, providing an objective basis for the quantitative analysis of a person's behavioral characteristics. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the human decision-making pattern analysis method of the present invention;

[0048] Figure 2 Schematic diagram of the personal value orientation tensor;

[0049] Figure 3 This is a diagram of the neural network architecture for decision-making patterns. Detailed Implementation

[0050] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0051] like Figure 1 As shown, the present invention provides a method for analyzing human decision-making patterns based on a situation-reasoning diagram. Its implementation process follows three steps: constructing a situation-reasoning diagram of the decision-making pattern, extracting and quantifying decision-influencing factors, and modeling the decision-making pattern. Each step is described in detail below:

[0052] S1. Constructing a Decision-Making Pattern Situation-Reasoning Graph. This step aims to transform massive, unstructured, sparse, and dynamically evolving open-source internet text information into a structured knowledge graph that can be processed by computers. The specific implementation steps are as follows:

[0053] S11. Defining the Knowledge Representation and Extraction Framework: This embodiment defines the knowledge structure as a quadruple structure containing timestamps, based on the traditional triplet. ;in, Represents the subject entity. Represents the object entity. This represents the relationship type (Relation) connecting the two. The timestamp (Time) represents the occurrence or validity of this relationship type. The four-tuple structure defined in this embodiment can accurately characterize the state of an entity relationship type at a specific moment, such as a person holding a certain position at a certain point in time, thereby capturing the dynamic changes in the decision-making environment.

[0054] To extract highly relevant information for decision analysis from complex open-source text, this embodiment pre-designs a constrained extraction template. Entity types cover people, organizations, locations, and specific events; in addition to general semantic relationships such as "belongs to" and "located in," sentiment relationships such as "support," "oppose," "condemn," and "appreciate" are specifically designed. The introduction of sentiment relationships aims to capture the attitudes of decision-makers and their stakeholders, providing a data foundation for subsequent preference analysis.

[0055] S12. Perform Structured Information Extraction: Employ a cue-based learning-based Universal Information Extraction (UIE) model to unify the entity and relation type extraction tasks into reading comprehension generation tasks. For each predefined entity or relation type, construct corresponding natural language cue words. For example, to extract ownership relations, the cue could be "[person entity]of[organization entity]" or "[r]of[e]is / are?", guiding the model to generate the corresponding relation triples from the text. .

[0056] S13. Time Information Alignment and Graph Construction: A specialized time alignment algorithm is designed to process time information that frequently appears as adverbial phrases in text; specifically:

[0057] By treating "time" as a special relation type and extracting it, we obtain time triples in the form of (entity, time, time value).

[0058] Based on the principles of "entity-time consistency" and "textual proximity," each semantic triple is searched for a time triple that is identical to its subject or object entity and is closest in the original text. The time value 't' is then attached to the semantic triple, thus synthesizing a complete quadruple. By aggregating all the quadruplets, a decision-making pattern situation-reasoning graph is constructed, also known as a structured dynamic knowledge graph.

[0059] S2. Extraction and Quantification of Factors Influencing Decision-Making. This step aims to extract and quantify two types of factors influencing decision-making based on the constructed situation-reasoning diagram: environmental factors from external stimuli and personal factors from internal constraints.

[0060] Extraction and Classification of Environmental Influencing Factors: Environmental influencing factors typically refer to the environmental state in which an individual makes a decision, constituting the external conditions that trigger the decision. These mainly include stakeholder groups, attitudes of others, cultural background, and related events. To focus on the context of a specific decision event, this embodiment uses the target decision event entity as the anchor point and its occurrence time as the baseline, tracing back a preset time window, such as 60 days before the decision. All nodes directly or indirectly connected to the anchor event entity (i.e., its multi-hop neighbors), as well as all edges between these nodes, are extracted to form the event background subgraph. To filter out significant noise in the open-source data, this embodiment statistically analyzes the frequency of occurrence of each entity in the subgraph across all historical decision events. Set frequency threshold Only entities that appear more frequently than a threshold are retained as the candidate entity set. This filters out occasional or irrelevant noise entities, as shown below:

[0061] ;

[0062] To address the problem of semantic ambiguity and difficulty in classifying entities in a graph, such as distinguishing whether an organization is a "stakeholder" or represents a "cultural background," this embodiment utilizes a large language model as a semantic classifier. This classifier constructs a prompt that includes candidate entities, the contextual relationships connected to those entities, and four predefined environmental dimensions. Leveraging the common-sense reasoning capabilities of the large language model, candidate entities are accurately mapped to four environmental subcategories: "stakeholder groups," "others' attitudes," "cultural background," or "related events."

[0063] Quantitative Calculation of Personal Influencing Factors: Personal influencing factors refer to an individual's intrinsic attributes, including personal abilities, value orientations, team relationships, and leadership relationships. Because these are difficult to observe directly, this embodiment employs a graph-based quantitative system method to transform abstract psychological characteristics into calculable numerical vectors. Specifically:

[0064] Individual competence: Individual competence is often reflected in an entity's professional skills in handling complex problems and the actual effectiveness of decision implementation. However, it is difficult to directly measure individual competence from open-source information. Therefore, this embodiment quantifies individual competence by calculating the weighted sum of the structural importance of historical decision events and the shift in external public opinion: For a specific decision event, the larger and more complex the event background subgraph, the more stakeholders involved and the higher the difficulty of handling the event, thus better reflecting the entity's competence. Based on this, the weight of the event is defined. for:

[0065]

[0066] in, and These represent the number of nodes and edges in the event background subgraph, respectively, where N is the total number of events. After obtaining the event weights, we combine this with the changes in the attitudes of key stakeholders towards the individuals before and after the decision-making event. The quantitative results of the individual's abilities were obtained: ;

[0067] in, Before the decision-making event ( )back( A positive shift in the attitudes of key stakeholders towards individuals indicates that the decision has achieved positive social effects, i.e., a change in attitude. It is a core indicator for characterizing the social effects of decision-making.

[0068] Value orientation: This is mainly reflected in an individual's attitude towards other organizations, individuals, etc. To quantify an individual's value orientation, this embodiment constructs a personal value orientation tensor. ,like Figure 2 As shown, the personal value orientation tensor The three dimensions represent the target entity e, the emotional relationship category, and so on. and time t. According to This calculation allows for the analysis of an individual's emotional preference distribution towards a specific entity (such as a business organization) over a period of time. Simultaneously, it calculates the interaction consistency between an individual and different target entities in the tensor space; higher interaction consistency indicates a stronger and more stable value preference for that target entity. (Personal value orientation tensor) The calculation formula is:

[0069] ;

[0070] in, For the physical form of a person, For the target entity, The numerator represents the number of samples within time t with character entity l as the head entity, target entity e as the tail entity, and emotional relationship category a as the association; the denominator represents the total number of samples within time t with all associations with character entity l as the head entity and target entity e as the tail entity.

[0071] Team relationships: This is reflected in the main members of the team a person belongs to, and their interactions and collaborations. Within a team, team cohesion is crucial for ensuring the effective execution of decisions. This embodiment calculates the net semantic density of the network based on the interaction subgraph of the team members; and introduces an sentiment weighting mechanism, assigning positive weights to edges representing cooperation and support in the knowledge graph, and negative weights to edges representing conflict and opposition. and Net interaction strength between The calculation formula is as follows:

[0072] ;

[0073] in, This is an indicator function; its value is 1 when the condition within the parentheses is true, and 0 otherwise. and These are the decision conditions, referring to the current relationship. It can be either a positive or negative relationship.

[0074] After calculating the net interaction strength between members, the weights of all edges in the interaction subgraph are summed, and the sum is divided by the total number of possible connections within the team. This calculates the team's relationship strength. Finally, the score is normalized to the [0,1] interval using the Sigmoid function. The closer the score is to 1, the more positive interactions within the team and the stronger the member relationships. The calculation formula is as follows:

[0075] ;

[0076] Leadership relationships: This is typically reflected in the degree to which a leader accepts information conveyed by subordinates. Therefore, this embodiment uses random walks to simulate the propagation process of information among team members and constructs a propagation probability matrix. The node propagation probability is inversely proportional to the node's degree, as shown in the following formula:

[0077] ;

[0078] in, For nodes The degree;

[0079] Assuming that information propagates uniformly among nodes, the simulation examines how information, after being sent from a team member node, travels through a finite number of steps. The coverage within the node is updated according to the following formula after k-step propagation, based on the node probability distribution:

[0080] ;

[0081] The higher the coverage and the faster the information spreads within the team, the more likely that the person has adopted an effective interactive or tightly controlled leadership style.

[0082] S3. Neural Network Modeling of Decision-Making Patterns. This step aims to address the "black box" problem of existing neural network models in decision analysis, which lack theoretical support and have poor interpretability. This method integrates the structured thinking of the Analytic Hierarchy Process (AHP) with bounded rationality decision theory to construct a neural network for decision-making patterns with a hierarchical architecture of "perception-selection-reasoning." It simulates the nonlinear information selection and reasoning process carried out by an individual under the constraints of their own cognitive abilities, value preferences, and other intrinsic traits, in response to external environmental stimuli, thus achieving interpretable and reproducible modeling of the decision-making process.

[0083] Network Structure Design: The decision-making neural network in this embodiment adopts a hierarchical "perception-selection-reasoning" architecture, which differs from the flattened concatenation of all features in traditional deep learning models. It restores the inherent logic of decision-making through layered modeling. For example... Figure 3 As shown, the neural network consists of an environmental factor input layer, a personal factor screening layer, and a deep decision reasoning layer. Its core design logic is as follows: environmental factors serve as the external input driving decision-making, quantified personal factors serve as the internal "gating threshold" constraining decision-making, the personal factor screening layer achieves cognitive filtering of environmental information based on personal characteristics, and the reasoning layer completes the nonlinear fusion of effective information, ultimately outputting the probability of a decision occurring.

[0084] The specific implementation methods for each level are as follows:

[0085] Environmental factor input layer: As the perceptual front end of the neural network, it receives and encodes the external objective conditions that trigger decision-making. Specifically, it acquires a set of environmental factors, which is extracted from the situation-inference diagram and filtered according to preset rules. An environmental factor, denoted as . ,in The number of environmental factors, typically the number of environmental factors. The range of values ​​is The selection is performed in descending order based on frequency of occurrence. In practice, this can be dynamically adjusted using a preset feature contribution threshold to ensure that the selected factors cover the four core dimensions of stakeholder groups, attitudes of others, cultural background, and related events, while avoiding overfitting due to excessively high feature dimensions. Each discrete symbolic factor is then... It is mapped to a high-dimensional dense embedding vector.

[0086] Personal Factor Screening Layer: This layer simulates a person's cognitive mechanism by gating and screening the embedding vectors output from the environmental factor input layer based on personal traits. Specifically:

[0087] The number of neurons in the personal factor screening layer corresponds one-to-one with the input personal influencing factors. Each neuron is responsible for processing the filtering effect of a single dimension of personal traits on environmental stimuli. The values ​​of each personal factor obtained by quantification through the analytic hierarchy process in the preceding step S2 are directly assigned to the bias terms of the neurons. Bias term The expression is:

[0088] ;

[0089] in This represents the quantified individual ability score (as stated in the previous text). (Mapped from) This represents a score indicating the quantified closeness of team relationships. This represents the quantified leadership relationship score. This indicates how a person treats specific environmental factors. Value orientation.

[0090] This layer uses the ReLU activation function to construct a stimulus-response threshold mechanism. The activation function expression is as follows:

[0091] ;

[0092] in This represents the learnable weight matrix of the filtering layer. Then it represents the input number of the first... Embedding vectors of environmental factors.

[0093] When the weighted intensity of external environmental stimuli and the sum of internal personal trait bias are greater than zero, the embedding vector of the corresponding environmental information is activated and transmitted to the deep decision reasoning layer; otherwise, the output is zero. This simulates the process by which a person automatically ignores / shiels specific environmental information due to personal bias or ability limitations.

[0094] Deep decision reasoning layer: used for non-linear fusion and decision reasoning of the filtered effective information, ultimately outputting the probability of the decision occurring. Specifically:

[0095] Summation and pooling are performed on all valid feature vectors output from the individual factor selection layer to generate a global state vector reflecting the current decision-making situation. ;in, This is the output feature vector (i.e., effective information feature) of the personal factor screening layer. This term will only have a non-zero output when the environmental stimulus exceeds the personal trait threshold.

[0096] Nonlinear reasoning, using the global state vector The input is fed into a multilayer perceptron (MLP) containing several fully connected layers, utilizing the learnable weight matrix in the MLP structure. and nonlinear activation functions , global state vector This is mapped to a latent feature space composed of the embedded vectors. Within this space, high-order features of fragmented cues are extracted by simulating logical association analysis within a person's brain, as shown in the following equation:

[0097] ;

[0098] in, This is the bias vector of the multilayer perceptron.

[0099] The probability output is mapped to the [0,1] interval using the Sigmoid activation function to obtain the final predicted probability value. As shown below:

[0100] ;

[0101] in, Transpose the weight vector of the output layer; This is the output vector of the previous layer; For the bias term of the output layer;

[0102] Model Training Phase: In this embodiment, when training the decision-making neural network described above, the objective is to minimize the binary cross-entropy loss function between the predicted value and the historical true decision label, driving the network parameter update. The loss function expression is as follows:

[0103] ;

[0104] in, This refers to the set of model parameters to be optimized in a neural network. This refers to the number of training samples, i.e., the number of decision events. The label value is the actual historical decision (usually 0 or 1). This is the probability of the decision occurring as predicted by the model.

[0105] During training, the bias term bᵢ of neurons in the personal factor screening layer is strictly frozen and does not participate in parameter updates. This ensures that the model learns only the correlation between environmental factors and decision results while respecting the constraints of the individual's established personal traits. This guarantees that the model conforms to the theory of bounded rationality in decision-making and improves the interpretability of decision results.

[0106] In summary, the character decision-making pattern analysis method proposed in this embodiment, based on the analysis results of decision-influencing factors, predefines constrained entity types such as characters, organizations, locations, and events, and relationship types such as sentiment polarity relationships; it constructs an extraction template based on reading comprehension to extract triples that can represent character decision-making related information from the text; and it designs an entity-time consistency alignment algorithm based on quadruples to accurately associate and align semantic triples with time information in the text, thereby extracting the time dimension information of the triples and constructing an integrated decision-making pattern situation-reasoning graph.

[0107] Subsequently, by combining node screening and semantic classification techniques of large language models, four types of environmental factors—stakeholder groups, attitudes of others, cultural background, and related events—are accurately extracted and mapped into vectorized embedded representations. The topological structure information and node semantic information of the graph are deeply integrated to construct an objective quantitative calculation model of personal influencing factors, thereby achieving a quantitative representation of personal influencing factors in the decision-making process. This method provides protection for this stage.

[0108] Finally, by integrating the structural ideas of the analytic hierarchy process (AHP), a neural network for decision-making patterns is constructed. Quantified personal factors are mapped to fixed bias terms in the network layers, and nonlinear activation functions are used to simulate cognitive gating and psychological threshold mechanisms in the decision-making process. Ultimately, this achieves the modeling and representation of individual decision-making patterns and processes.

Claims

1. A method for analyzing character decision-making patterns based on a situation-reasoning diagram, comprising the following steps: S1. Constructing a Decision-Making Pattern-Reasoning Diagram: From open-source text on the internet, entities and relations are extracted using a general information extraction model and prompt words; a time alignment algorithm is used to attach time information in the text to corresponding semantic triples, generating quadruples containing subject entity, object entity, relation, and timestamp. This allows for the construction of a dynamic knowledge graph; the extracted relationship types include affective relationships used to characterize attitude tendencies. S2. Extraction and Quantification of Decision-Making Influencing Factors: Based on the aforementioned dynamic knowledge graph, environmental influencing factors and personal influencing factors are extracted and quantified respectively; wherein: The extraction of environmental influencing factors includes: taking the decision event entity as the anchor point, backtracking a preset time window, extracting the nodes and edges connected to the anchor point entity within the window to form an event background subgraph; filtering based on the frequency of the entity's occurrence in historical events to obtain a candidate entity set; and using a large language model to classify the candidate entities into stakeholder groups, other people's attitudes, cultural backgrounds, or related event categories according to predefined prompts. The personal influencing factors include at least personal ability, value orientation, and team relationships, and their quantification methods are as follows: Personal ability: The event weight is determined based on the structural complexity of the event background subgraph. The weight Attitude changes among key stakeholders that characterize the social effects of decisions A weighted summation is performed to obtain the individual's overall ability score. ,in The number of events that trigger character decisions; Value Orientation: By constructing and analyzing the individual value orientation tensor The calculations show that the tensor reflects the distribution of a person's emotional inclination towards the target entity; Team relationships: The net semantic density of the network is obtained by calculating and normalizing the network subgraph of team member interactions with assigned emotional weights; S3. Character Decision-Making Pattern Modeling and Analysis: Construct a character decision-making pattern neural network containing a screening layer and a fusion reasoning layer; use the quantified personal influencing factors as the gate threshold of the screening layer to perform cognitive filtering on the input environmental influencing factors; the fusion reasoning layer performs nonlinear fusion on the filtered effective information and finally outputs the probability of the decision occurring.

2. The method according to claim 1, characterized in that, In S1, the emotional relationship includes at least one of support, opposition, condemnation, and appreciation.

3. The method according to claim 1, characterized in that, In S1, the time alignment algorithm specifically includes: Time information is extracted as a special relation type to obtain time triples; based on the principles of entity consistency and text proximity, the time values ​​in the time triples are attached to the corresponding semantic triples to synthesize the quadruples.

4. The method according to claim 1, characterized in that, In S2, predefined categories of environmental factors include: stakeholders, attitudes of others, cultural background, and related events.

5. The method according to claim 1, characterized in that, In S2, when calculating individual capabilities, the event weights... The calculation formula is: ； in, and Let N be the number of nodes and edges in the event background subgraph corresponding to the i-th decision time, respectively, and N be the total number of events.

6. The method according to claim 1, characterized in that, In S2, the personal value orientation tensor Calculated using the following formula: ； in, For the physical form of a person, For the target entity, For emotional relationships.

7. The method according to claim 1, characterized in that, In S2, team relationship scores are calculated. The formula is: ； in, For the Sigmoid function, For the number of team members, Let represent the net interaction strength between members i and j. Positive relationships are denoted as positive weights, and negative relationships as negative weights.

8. The method according to claim 1, characterized in that, The personal influencing factors also include leadership relationships, which are quantified as follows: Construct an information propagation probability matrix P based on the team member interaction subgraph, where , For nodes The degree; Simulate information through random walk within a finite number of steps. The internal dissemination process is characterized by the proportion of nodes covered by information and the speed of dissemination, which reflects the leadership style.

9. The method according to claim 1, characterized in that, In S3, the filtering layer includes a fixed bias set according to the individual influencing factor vector to simulate the threshold psychological mechanism.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 9.

Images