Multi-dimensional cognitive state analysis method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and discloses a multi-dimensional cognitive state analysis method and device, electronic equipment and a storage medium. The method comprises the following steps: constructing a cognitive state classification system and a domain sample data set; determining target dialogue text; learning hierarchical structure features of the target dialogue text on a hyperbolic space based on the domain sample data set, and determining cognitive representations of cognitive state dimensions presented by the target dialogue text in the hyperbolic space; the cognitive state dimensions at least comprise four dimensions of emotion, thinking mode, position and intention; performing semantic training on the cognitive representations based on a preset large language model, and reasoning out a cognitive state image satisfying cognitive analysis text constraints and cognitive hierarchical structure constraints; the hierarchical overlap problem of the cognitive state is solved by using exponentially increasing space capacity, and geometric structure knowledge in the hyperbolic space is injected into the large language model, so that the joint modeling and reasoning capability of the model for the multi-dimensional cognitive state in a complex social scene is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to multi-dimensional cognitive state analysis methods, devices, electronic devices, and storage media. Background Technology

[0002] With the widespread application of artificial intelligence technology in fields such as psychological counseling, social media analysis, and dialogue systems, accurately understanding and simulating human cognitive states has become a key goal in achieving advanced artificial intelligence. Human cognitive states are a complex system composed of multiple dimensions, including emotions, thought patterns, stances, and intentions, with close interactions and hierarchical relationships among these dimensions.

[0003] Related technologies mainly focus on the independent analysis of a single psychological dimension. When constructing a comprehensive, multi-dimensional cognitive state model, they lack a systematic interaction mechanism between different cognitive dimensions. In addition, benchmarking and modeling methods lack coverage of the overall cognitive state, ignore the deep logical connections between dimensions, and lack full coverage of multiple core dimensions that connect emotions and stances. As a result, they cannot effectively distinguish different cognitive states when dealing with complex and multi-level psychological cognitive tasks. Summary of the Invention

[0004] This invention provides a multi-dimensional cognitive state analysis method, device, electronic device, and storage medium, which solves the problem that the lack of deep logical connections between emotions, ways of thinking, stances, and intentions makes it impossible to effectively distinguish different cognitive states when dealing with complex and multi-level psychological cognitive tasks.

[0005] In a first aspect, the present invention provides a multi-dimensional cognitive state analysis method, the method comprising:

[0006] Construct a cognitive state classification system and a domain sample dataset based on the cognitive state classification system; The target dialogue text is determined from the target domain sample dataset corresponding to the target domain; wherein, the target domain sample dataset stores the target dialogue text of the target domain; Based on a domain sample dataset, this study learns the hierarchical structure features of the target dialogue text in hyperbolic space, determining the cognitive state dimensions presented by the target dialogue text in hyperbolic space. The hierarchical structure features characterize the relationships of emotion, thought process, stance, and intention between the preceding and following dialogue texts in the target dialogue text. The cognitive state dimensions include at least: emotion, thought process, stance, and intention. Figure 4 One dimension; Based on a pre-defined large language model, semantic training is performed on cognitive representations to infer cognitive state profiles that satisfy the constraints of cognitive analysis text and cognitive hierarchical structure.

[0007] This invention constructs a cognitive state classification system and a domain sample dataset based on this system as the foundation for model training and evaluation. It utilizes hyperbolic space to learn the hierarchical structural features of the target dialogue text in hyperbolic space, thereby determining the emotions, thought processes, stances, and intentions presented in the target dialogue text. Figure 4 The system represents cognitive states in hyperbolic space across multiple dimensions and trains these representations semantically based on a pre-defined large language model. This allows for the inference of cognitive state profiles that satisfy both cognitive analysis text constraints and cognitive hierarchy constraints. By leveraging the exponential capacity of hyperbolic space to achieve hierarchical separation of cognitive state dimensions, the system captures the deep logic between emotions, thought processes, stances, and intentions implied in the cognitive states presented in dialogue texts. This provides a more transparent and human-cognitive analytical method for psychological research and public opinion analysis, enhancing the model's ability to jointly model and infer multidimensional cognitive states in complex social scenarios and achieving high-precision end-to-end cognitive simulation.

[0008] In one optional implementation, based on a domain sample dataset, the hierarchical structural features of the target dialogue text in hyperbolic space are learned to determine the cognitive representation of the cognitive state dimension presented by the target dialogue text in hyperbolic space, including: Extract the semantic features corresponding to the preceding and following dialogues in the target dialogue text and map them to Euclidean space to determine the Euclidean feature representation of the target dialogue text. Hyperbolic manifold projection is performed on the Euclidean feature representation based on the Euclidean tangent space to obtain the hyperbolic feature representation of the cognitive state dimension of the target dialogue text. Hyperbolic feature aggregation is performed on hyperbolic feature representation based on tangent space to obtain hierarchical relation fusion representation of cognitive state dimension of target dialogue text; By using a pre-defined loss function, the hyperbolic position of the hierarchical relationship of the cognitive state dimension is obtained in the Poincaré sphere model, thereby determining the cognitive representation.

[0009] This invention parses and maps the target dialogue text to Euclidean space to determine its Euclidean feature representation. Then, it performs hyperbolic manifold projection and hyperbolic feature aggregation to obtain a hierarchical relationship fusion representation of the cognitive state dimension of the target dialogue text. By solving the hyperbolic position of the hierarchical relationship of the cognitive state dimension in the Poincaré sphere model, the cognitive representation is determined. This solves the problem of hierarchical overlap of cognitive states with an exponentially increased spatial capacity, significantly improving the model's ability to jointly model and reason about multidimensional cognitive states in complex social scenarios.

[0010] In one optional implementation, a hyperbolic manifold projection is performed on the Euclidean feature representation based on the Euclidean tangent space to obtain a hyperbolic feature representation of the cognitive state dimension of the target dialogue text, including: Construct a hyperbolic space network and linear projection layers on hyperbolic space, wherein the linear projection layers include at least an emotion linear projection layer, a thinking style linear projection layer, a stance linear projection layer, and an intention linear projection layer; Based on hyperbolic space network, Euclidean feature representation is mapped to Euclidean tangent space of cognitive state dimension to obtain Euclidean tangent space representation; Based on a linear projection layer and a preset exponential mapping function, the Euclidean tangent space representation of the cognitive state dimension is projected onto a Poincaré spherical manifold to determine the hyperbolic features of the target dialogue text in the cognitive state dimension.

[0011] This invention constructs corresponding linear projection layers in hyperbolic space for different cognitive state dimensions. Using a hyperbolic space network, Euclidean feature representations are mapped to the Euclidean tangent space of the cognitive state dimension. Through an exponential mapping function, these representations in the tangent space are projected onto a Poincaré spherical manifold, thereby obtaining the hyperbolic features of the target dialogue text in the cognitive state dimension. By using hyperbolic space to process data with hierarchical structures or complex relationships, the invention effectively captures the complex relationships and inherent hierarchical structures of cognitive states in dialogue text, thereby improving the accuracy of cognitive state analysis.

[0012] In one optional implementation, hyperbolic feature aggregation is performed on the hyperbolic feature representation based on the tangent space to obtain a hierarchical relational fusion representation of the cognitive state dimension of the target dialogue text, including: Based on a preset logarithmic function, the hyperbolic features of the cognitive state dimension are mapped back to the origin of the tangent space to obtain the tangent space origin-aligned representation. The cognitive state space is fused by aligning the origin of the tangent space to obtain the fused cognitive anchor vector, which represents the hierarchical relationship of the cognitive state dimension of the target dialogue text.

[0013] This invention utilizes a logarithmic function to map the hyperbolic features of the cognitive state dimension back to the origin of the tangent space, completing the spatial transformation. The aligned features in the tangent space are then fused to obtain a cognitive anchor vector that can represent the hierarchical relationship of the cognitive state dimension. By aggregating the scattered hyperbolic features into a single cognitive anchor vector, the cognitive state information at different levels in the dialogue text can be effectively integrated, thereby more clearly representing the hierarchical relationship between these states and providing a structured feature representation for subsequent analysis.

[0014] In one optional implementation, a preset loss function is used to solve for the hyperbolic position of the hierarchical relationship of the cognitive state dimension in the Poincaré sphere model, thereby determining the cognitive representation, including: Determine the preset multi-task classification loss function and hyperbolic geometric regularization loss function, wherein the multi-task classification loss function is used to characterize the semantic classification accuracy in the target dialogue text; Based on the cognitive state dimension, the sum of the cross-entropy losses of the target dialogue text in the cognitive state dimension is calculated to obtain the solution of the preset multi-task classification loss function. By utilizing the Poincaré sphere distance function of the Poincaré sphere manifold, we can solve for the clustered position and discrete position of the cognitive state dimension of the target dialogue text in hyperbolic space, and obtain the solution of the hyperbolic geometric regularization loss function. Cognitive representations are determined based on the results of solving the pre-defined multi-task classification loss function and the hyperbolic geometric regularization loss function.

[0015] This invention constructs a pre-defined multi-task classification loss function and a hyperbolic geometric regularization loss function. The pre-defined multi-task classification loss function is used to measure the accuracy of the dialogue text in semantic classification, while the hyperbolic geometric regularization loss function is used to measure the positional relationship of different cognitive states in hyperbolic space, thereby determining the cognitive representation of the cognitive state. By combining the accuracy of semantic classification with the geometric structure constraints in hyperbolic space, on the one hand, the multi-task loss ensures that the model can correctly identify different cognitive states; on the other hand, the hyperbolic geometric regularization loss enables the model to simultaneously learn the hierarchical relationship between cognitive states by distributing different cognitive states in different positions in space, thus obtaining a more accurate and structured cognitive representation.

[0016] In one optional implementation, semantic training of cognitive representations is performed based on a pre-defined large language model, including: Construct a parameterized cognitive projection network, which is used to receive cognitive anchor vectors generated by a hyperbolic space network; Using cognitive anchor vectors as input to a pre-defined large language model, the embedding space dimension representation of the pre-defined large language model is determined. Based on the preset hyperbolic tangent function, the output of the preset large language model is semantically truncated to generate a dynamic soft prompt embedding representation of a preset length; A new input sequence is formed by first inputting the soft prompt embedding representation and then inputting the target dialogue text in the input order; The new input sequence is fed into a pre-defined large language model to perform semantic training on the cognitive representation.

[0017] This invention constructs a parameterized cognitive projection network to receive cognitive anchor vectors generated by a hyperbolic space network. These anchor vectors are then input into a large language model to obtain their corresponding representations in the model's embedding space. A preset hyperbolic tangent function is used to semantically truncate the output of the preset large language model, generating corresponding dynamic soft prompts. A new input sequence is formed by inputting the soft prompts first, followed by the dialogue text. This allows structured cognitive information to be input as contextual information into the model without altering its structure, thereby guiding the model to better utilize deep cognitive state features when generating or understanding text, thus improving the model's performance on complex tasks.

[0018] In one optional implementation, a cognitive state profile that satisfies the constraints of cognitive analysis text and cognitive hierarchy structure is inferred, including: When training the cognitive representation semantically using the pre-defined large language model, the hidden state corresponding to the last layer is extracted and mapped back to the tangent space dimension of the hyperbolic space network to obtain the semantic vector. Based on the cosine direction consistency between semantic vectors and cognitive anchor vectors, the cognitive hierarchy structure can be inferred. When the cosine direction consistency between the semantic vector and the cognitive anchor vector satisfies the maximum cosine similarity constraint, the cognitive state profile satisfies the cognitive hierarchy structure constraint. Determine the preset autoregressive language loss function, and infer cognitive analysis text based on the autoregressive language loss function; When it is determined that the cognitive analysis text satisfies the cognitive analysis text accuracy constraint, it is determined that the cognitive state profile satisfies the cognitive analysis text constraint.

[0019] This invention extracts the hidden states of the last layer of the model, maps them back to hyperbolic space, and determines their cosine direction consistency with the cognitive anchor vector. When the cosine direction consistency between the semantic vector and the cognitive anchor vector satisfies the maximum cosine similarity constraint, it is determined that the cognitive hierarchical structure constraint is satisfied. The generated cognitive analysis text is evaluated by calculating the autoregressive language loss function. When the cognitive analysis text satisfies the cognitive analysis text accuracy constraint, it is determined that the cognitive analysis text constraint is satisfied. Finally, a complete cognitive state profile is inferred. This ensures that the cognitive state profile conforms to the hierarchical relationship between cognitive states in structure, and that the cognitive state profile can accurately express the analysis results in content, so that the cognitive state profile has both internal structural rationality and external textual expression accuracy.

[0020] Secondly, the present invention provides a multi-dimensional cognitive state analysis device, the device comprising: The building module is used to construct a cognitive state classification system and a domain sample dataset based on the cognitive state classification system; The first determining module is used to determine the target dialogue text from the target domain sample dataset corresponding to the target domain; wherein, the target domain sample dataset stores the target dialogue text of the target domain; The second determination module is used to learn the hierarchical structure features of the target dialogue text in hyperbolic space based on the domain sample dataset, and to determine the cognitive representation of the cognitive state dimension presented by the target dialogue text in hyperbolic space. The hierarchical structure features characterize the relationship between the preceding and following dialogue texts in the target dialogue text, including emotions, thought patterns, stances, and intentions. The cognitive state dimension includes at least: emotions, thought patterns, stances, and intentions. Figure 4 One dimension; The training module is used to perform semantic training on cognitive representations based on a pre-set large language model, and to infer a cognitive state profile that satisfies the constraints of cognitive analysis text and cognitive hierarchical structure.

[0021] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect or any corresponding embodiment thereof.

[0022] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the method described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the first type of multi-dimensional cognitive state analysis method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the second process of the multi-dimensional cognitive state analysis method according to an embodiment of the present invention; Figure 4 This is a system architecture diagram of the multi-dimensional cognitive state analysis method according to an embodiment of the present invention; Figure 5 This is a structural block diagram of a multi-dimensional cognitive state analysis device according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0026] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0027] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0028] As an optional application scenario of this invention, such as Figure 1 As shown, application 101 is installed in terminal device 110, and user 130 can interact with application 101 through terminal device 110 and / or access device of terminal device 110.

[0029] For example, application 101 can be any application that provides question-and-answer related services. For instance, application 101 could be a question-and-answer interactive application, such as a text-to-text application, an image-to-text application, etc. Figure 1 In the application scenario shown, if application 101 is active, the terminal device 110 can display the interface 102 of application 101. The interface 102 may include various pages that application 101 can provide, such as interactive pages, settings pages, query pages, etc.

[0030] In some embodiments, terminal device 110 is communicatively connected to server 120 to provide services to application 101. Terminal device 110 may be a mobile terminal, fixed terminal, or portable terminal, etc., including but not limited to mobile phones, desktop computers, laptop computers, multimedia tablets, e-book devices, gaming devices, or any combination thereof, including accessories and peripherals of these devices or any combination thereof. In some embodiments, terminal device 110 may also support any type of interface, and server 120 may be various types of computing systems or servers capable of providing computing power, including but not limited to mainframes, edge computing nodes, computing devices in cloud environments, etc.

[0031] It should be noted that, Figure 1 This is merely an example of an application scenario and does not limit the scope of protection of this invention.

[0032] The embodiments of the present invention will now be described with reference to the accompanying drawings. It should be understood that the pages shown in the drawings are merely examples, and various page designs are possible in practice. The various graphic elements on the page may have different arrangements and different visual representations, one or more elements may be omitted or replaced, and one or more other elements may also be present; no limitations are imposed in the embodiments of the present invention. Furthermore, the embodiments are described primarily with respect to terminal device 110 in this context. It should be understood that the actions described relative to terminal device 110 can be performed by application 101 on terminal device 110, or can be performed by application 101 in conjunction with its server (e.g., server 120).

[0033] With the widespread application of artificial intelligence technology in fields such as psychological counseling, social media analysis, and dialogue systems, accurately understanding and simulating human cognitive states has become a key goal in achieving advanced artificial intelligence. Human cognitive states are a complex system composed of multiple dimensions, including emotions, thought patterns, stances, and intentions, with close interactions and hierarchical relationships among these dimensions.

[0034] In related technologies, the focus is mainly on the independent analysis of a single psychological dimension. For example, researchers build separate models for specific tasks such as stance detection, sentiment analysis, or intent recognition. These methods typically treat each cognitive dimension as an isolated task, using specific datasets (such as P-Stance for stance detection and SemEval-16 for sentiment analysis) to train dedicated classifiers. This fragmented perspective oversimplifies the complexity of human cognition and ignores the systematic interaction mechanisms between cognitive dimensions as pointed out in psychological research.

[0035] In recent years, with breakthroughs in Large Language Model (LLM) technology, cognitive modeling using LLM has become a mainstream trend. LLM, with its powerful semantic understanding capabilities, has demonstrated significant advantages in handling single cognitive tasks. Related technologies are attempting to utilize LLM to handle more complex scenarios. Some works are trying to combine annotation and modeling dimensions, such as combining stance and sentiment, or integrating intent information into dialogue settings. Benchmark tests (such as SemEval-16 and GunStance) are also beginning to explore combinations of certain dimensions, attempting to improve the model's understanding of human psychological states by increasing data dimensions. These LLM-based methods primarily rely on the general knowledge learned by the model during the pre-training phase, mapping text to corresponding cognitive labels through instruction fine-tuning or contextual learning, thereby achieving a certain degree of automated analysis of user psychological states.

[0036] While the aforementioned technologies perform reasonably well on single-dimensional cognitive tasks, they still have significant technical shortcomings and theoretical bottlenecks when constructing comprehensive, multi-dimensional cognitive state models.

[0037] First, the benchmarking and modeling methods in related technologies lack comprehensive coverage of cognitive states and ignore the deep logical connections between dimensions. Most current datasets only cover one or two dimensions of emotion, stance, or intention, and the deep logical connections between these dimensions are lacking. There is a lack of methods that simultaneously encompass all four key dimensions: emotion, mindset, stance, and intention. In particular, the "mindset," which plays a crucial role in connecting emotion and stance, is often overlooked in related technologies. This lack of dimension prevents models from capturing a more complete cognitive state. For example, conflicting stances may stem from deliberate analytical thinking or from the emotional outburst of anger; related technologies struggle to distinguish between these two fundamentally different cognitive paths.

[0038] Secondly, and more fundamentally, the problem lies in the severe "geometric mismatch" between the underlying geometric representation of traditional large language models and the internal structure of complex human cognitive states. (Psychological research and Gromov's work are relevant.) Hyperbolic analysis reveals that cognitive states exhibit a significant hierarchical structure, where broad concepts branch out into specific subconcepts. The representational space required for this hierarchical structure grows exponentially with the depth of the hierarchy. However, traditional large language models generally employ Euclidean space as the embedding space, whose volume only grows polynomially with dimensionality.

[0039] This difference in representational capacity leads to the phenomenon of "cognitive crowding." When attempting to forcibly embed hierarchical cognitive states with exponentially increasing demands into a Euclidean space with limited capacity, distinct cognitive states that should be semantically differentiated are forced to overlap and squeeze each other in space. This geometric mismatch directly causes a sharp decline in the performance of LLM when performing multi-dimensional joint modeling. Experimental data shows that although related models (such as GPT-4o) perform well on single-dimensional tasks, their accuracy drops significantly when predicting four dimensions simultaneously. Related technologies have failed to address this problem at the underlying mechanism of geometric representation, resulting in their inability to effectively distinguish fine-grained cognitive differences when dealing with complex, multi-layered psychological cognitive tasks, severely impacting the application potential of artificial intelligence in the field of deep psychological understanding.

[0040] For example, the spatial overlap and compression of different cognitive states refers to the following aspects: At the level of thought, different ways of thinking or perspectives intertwine in cognition. For instance, when solving a controversial social problem, it is necessary to consider both moral and legal / economic aspects. These different ways of thinking operate simultaneously in the brain, influencing and interfering with each other. Moral thinking emphasizes fairness, justice, and humanistic care; legal thinking focuses on adherence to rules and regulations; and economic thinking focuses on costs and benefits. Their overlap and compression in cognitive space can cause confusion during the thinking process, making it difficult to make quick and clear judgments. Multiple memories and emotional experiences are evoked in the same situation, causing these different experiences to interact in cognitive space. For example, when revisiting a familiar place, different emotional experiences such as joy, anger, sorrow, and happiness that occurred there in the past may simultaneously surface. Happy memories may overlap with later setbacks or sad events; these emotional experiences compress each other, making the user's current feelings complex and difficult to describe with a single emotion.

[0041] The purpose of this invention is to overcome the "cognitive crowding" defect and the lack of multidimensional cognitive data in the background technology, and to provide a multidimensional cognitive state modeling method (HyCoLLM) based on hyperbolic geometry and cognitive alignment of a large language model. This method aims to break through the geometric bottleneck of Euclidean space in representing hierarchical complex cognitive states, and achieves efficient decoupling and representation of human cognitive states by introducing hyperbolic geometric priors. Specifically, it first constructs the first model encompassing emotion, thinking style, stance, and intention. Figure 4First, a comprehensive cognitive benchmark dataset fills the data gap. Second, a hyperbolic space network is constructed using the Poincaré sphere model to solve the problem of overlapping levels of cognitive states with exponentially increasing spatial capacity. Finally, through a hyperbolic-guided alignment fine-tuning mechanism, geometric structural knowledge in hyperbolic space is injected into the large language model, thereby significantly improving the model's ability to jointly model and reason about multidimensional cognitive states in complex social scenarios, and achieving high-precision end-to-end cognitive simulation.

[0042] According to an embodiment of the present invention, a method for multi-dimensional cognitive state analysis is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0043] This embodiment provides a multi-dimensional cognitive state analysis method, which can be used in the aforementioned terminal devices, such as mobile phones and tablets. Figure 2 This is a schematic diagram of the first type of multi-dimensional cognitive state analysis method according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Construct a cognitive state classification system and a domain sample dataset based on the cognitive state classification system.

[0044] It should be noted that the cognitive state classification system includes, but is not limited to, cognitive state classification systems based on emotional dimensions, cognitive state classification systems based on thinking patterns, cognitive state classification systems based on stance dimensions, and cognitive state classification systems based on intention dimensions.

[0045] The cognitive state classification system is divided into several dimensions. The emotional dimension includes emotional categories and their corresponding specific attributes; the thinking style dimension includes thinking style categories and their corresponding specific attributes; the stance dimension includes stance categories and their corresponding specific attributes; and the intention dimension includes intention categories and their corresponding specific attributes.

[0046] The construction of this four-dimensional cognitive state classification system decomposes the complex cognitive process into emotion, thinking style, stance and intention; these four dimensions do not exist independently, but form an organic whole.

[0047] See Table 1 below for the four-dimensional cognitive state classification system.

[0048] Table 1. Classification System of Cognitive and Thinking Levels

[0049] It should be noted that, based on this four-dimensional cognitive state classification system, this embodiment collects data from social media platforms on different fields such as trade, elections, cultural diversity, and economic policies. After deduplication, cleaning, and cross-labeling by multiple experts, a domain sample dataset containing thousands of high-quality samples is constructed as the basis for subsequent model training and evaluation analysis.

[0050] Step S202: Determine the target dialogue text from the target domain sample dataset corresponding to the target domain; wherein, the target domain sample dataset stores the target dialogue text of the target domain.

[0051] It should be noted that the target domain sample dataset includes, but is not limited to, sample datasets for trade, elections, cultural diversity, and economic policy, respectively, collected from social media platforms. The target dialogue text is determined from the target domain sample dataset. The target dialogue text is a piece of natural language text that can represent an individual's cognition of relevant events or viewpoints in that domain. This target dialogue text may contain a preceding and a following paragraph, with a logical connection between them.

[0052] Step S203: Based on the domain sample dataset, learn the hierarchical structure features of the target dialogue text in hyperbolic space, and determine the cognitive representation of the cognitive state dimension presented by the target dialogue text in hyperbolic space; wherein the hierarchical structure features characterize the relationship between the preceding and following dialogue texts in the target dialogue text in terms of emotion, thinking style, stance, and intention, and the cognitive state dimension includes at least: emotion, thinking style, stance, and intention. Figure 4 Each dimension.

[0053] It should be noted that the learning target dialogue text has hierarchical structural features in hyperbolic space, that is, hierarchical structural features of the cognitive state learned on a hyperbolic manifold.

[0054] Hyperbolic space is a non-Euclidean geometric space with constant negative curvature. Hyperbolic space expands exponentially as the radius increases. The exponential capacity of hyperbolic space is used to achieve hierarchical separation of cognitive state dimensions.

[0055] Step S204: Based on the preset large language model, perform semantic training on the cognitive representation to infer a cognitive state profile that satisfies the constraints of cognitive analysis text and cognitive hierarchical structure.

[0056] It should be noted that the preset large language model can be a Large Language Model (LLM), an artificial intelligence model designed to understand and generate human language. It can be trained on large amounts of text data and can perform a wide range of tasks, including sentiment analysis. In this embodiment, semantic training of cognitive representations is performed based on this large language model.

[0057] Among them, the cognitive state profile is the analysis result of the user's emotions, thinking patterns, stance and intentions after the model is trained. In other words, the cognitive state profile represents the user's deep cognitive logic.

[0058] The multi-dimensional cognitive state analysis method provided in this embodiment constructs a cognitive state classification system and a domain sample dataset based on the cognitive state classification system as the basis for model training and evaluation analysis. It utilizes hyperbolic space to learn the hierarchical structural features of the target dialogue text in hyperbolic space, thereby determining the emotions, thought processes, stances, and intentions presented in the target dialogue text. Figure 4 The system represents cognitive states in hyperbolic space across multiple dimensions and trains these representations semantically based on a pre-defined large language model. This allows for the inference of cognitive state profiles that satisfy both cognitive analysis text constraints and cognitive hierarchy constraints. By leveraging the exponential capacity of hyperbolic space to achieve hierarchical separation of cognitive state dimensions, the system captures the deep logic between emotions, thought processes, stances, and intentions implied in the cognitive states presented in dialogue texts. This provides a more transparent and human-cognitive analytical method for psychological research and public opinion analysis, enhancing the model's ability to jointly model and infer multidimensional cognitive states in complex social scenarios and achieving high-precision end-to-end cognitive simulation.

[0059] This embodiment provides a multi-dimensional cognitive state analysis method, which can be used in the aforementioned terminal devices, such as mobile phones and tablets. Figure 3 This is a schematic diagram of the second process of the multi-dimensional cognitive state analysis method according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps: Step S301: Construct a cognitive state classification system and a domain sample dataset based on the cognitive state classification system. For details, please refer to [link to relevant documentation]. Figure 2 Step S201 of the illustrated embodiment will not be described again here.

[0060] Step S302: Determine the target dialogue text from the target domain sample dataset corresponding to the target domain; wherein, the target domain sample dataset stores the target dialogue text of the target domain. For details, please refer to... Figure 2 Step S202 of the illustrated embodiment will not be described again here.

[0061] Step S303: Based on the domain sample dataset, learn the hierarchical structure features of the target dialogue text in hyperbolic space, and determine the cognitive representation of the cognitive state dimension presented by the target dialogue text in hyperbolic space; wherein the hierarchical structure features characterize the relationship between the preceding and following dialogue texts in the target dialogue text in terms of emotion, thinking style, stance, and intention, and the cognitive state dimension includes at least: emotion, thinking style, stance, and intention. Figure 4 Each dimension.

[0062] Specifically, step S303 includes: Step S3031: Extract the semantic features corresponding to the dialogue context and dialogue background in the target dialogue text and map them to Euclidean space to determine the Euclidean feature representation of the target dialogue text.

[0063] It should be noted that this applies to the input text sequence (target dialogue text). The pre-trained Transformer encoder is used to extract contextual semantic features, and the mean or mean of the last hidden state is taken. Bit vectors as the basic Euclidean representation (Euclidean characteristic representation) , It represents a d-dimensional real vector.

[0064] Step S3032: Based on the Euclidean tangent space, perform hyperbolic manifold projection on the Euclidean feature representation to obtain the hyperbolic feature representation of the cognitive state dimension of the target dialogue text.

[0065] Specifically, step S3032 includes: Step a1: Construct a hyperbolic space network and linear projection layers on hyperbolic space, wherein the linear projection layers include at least an emotion linear projection layer, a thinking style linear projection layer, a stance linear projection layer, and an intention linear projection layer; It should be noted that, in order to geometrically decouple emotions, ways of thinking, stances, and intentions... Figure 4 In this system, four independent linear projection layers are constructed. .in, Representing the linear projection layer of emotion, Linear projection layer representing thinking patterns Indicates the position linear projection layer, This indicates a linear projection layer.

[0066] Hyperbolic space network is a neural network defined in hyperbolic space to perform data embedding and recognize state classification in hyperbolic space.

[0067] Step a2: Based on the hyperbolic space network, the Euclidean feature representation is mapped to the Euclidean tangent space of the cognitive state dimension to obtain the Euclidean tangent space representation; Specifically, the Euclidean feature representation is mapped onto the Euclidean space of the four cognitive state dimensions of emotion, thinking style, stance and intention, to obtain its Euclidean space representation.

[0068] In the context of cognitive states, Euclidean tangent space can be used to describe and analyze the local characteristics of cognitive states. Cognitive states can be represented by points in a high-dimensional space, known as the cognitive state space. Euclidean tangent space is a vector space obtained by locally linearly approximating this cognitive state at a point within the cognitive state space. For example, in artificial intelligence, Euclidean tangent space in the cognitive state dimension can be used to construct cognitive models. In reinforcement learning, for instance, an agent's cognitive state changes with its environment. Using Euclidean tangent space, the agent's cognitive state can be locally linearized, thereby simplifying the model's complexity and improving learning efficiency.

[0069] Step a3: Based on the linear projection layer and the preset exponential mapping function, the Euclidean tangent space representation of the cognitive state dimension is projected onto the Poincaré spherical manifold to determine the hyperbolic features of the target dialogue text in the cognitive state dimension.

[0070] It should be noted that the exponential mapping function is used. Projecting features onto curvature Poincaré spherical manifold The specific formula is as follows:

[0071] in, It is a vector in the tangent space; Euclidean features; Represents the exponential mapping function; This represents four independent linear projection layers; The curvature of the Poincaré spherical manifold is represented by . It represents the magnitude of the tangent space vector.

[0072] Among them, the Poincaré sphere manifold is a Poincaré sphere model in hyperbolic space. An n-dimensional manifold is a space that can be locally approximated by Euclidean tangent space. It can perform hierarchical modeling of target dialogue text and learn the hierarchical structural features of cognitive states.

[0073] The multi-dimensional cognitive state analysis method provided in this embodiment constructs corresponding linear projection layers in hyperbolic space for different cognitive state dimensions. Using a hyperbolic space network, Euclidean feature representations are mapped to the Euclidean tangent space of the cognitive state dimension. Through an exponential mapping function, these representations in the tangent space are projected onto a Poincaré spherical manifold, thereby obtaining the hyperbolic features of the target dialogue text in the cognitive state dimension. By using hyperbolic space to process data with hierarchical structures or complex relationships, the method effectively captures the complex relationships and inherent hierarchical structures of cognitive states in the dialogue text, thereby improving the accuracy of cognitive state analysis.

[0074] Step S3033: Perform hyperbolic feature aggregation on the hyperbolic feature representation based on the tangent space to obtain the hierarchical relationship fusion representation of the cognitive state dimension of the target dialogue text.

[0075] Specifically, step S3033 includes: Step b1: Based on the preset logarithmic function, the hyperbolic features of the cognitive state dimension are mapped back to the origin of the tangent space to obtain the tangent space origin-aligned representation. Specifically, to obtain a unified representation of cognitive states, this embodiment performs feature aggregation in the tangent space. The system aggregates hyperbolic features in four dimensions. Through logarithmic mapping Map back to the origin of the shear space .

[0076] Step b2 involves performing cognitive state space fusion on the tangent space origin-aligned representation to obtain the fused cognitive anchor vector, which represents the hierarchical relationship of the cognitive state dimension of the target dialogue text.

[0077] Specifically, it is fused into a unified cognitive anchor vector through attention mechanisms or splicing operations. This vector geometrically encodes a hierarchical relationship in four dimensions.

[0078] The multi-dimensional cognitive state analysis method provided in this embodiment utilizes a logarithmic function to map the hyperbolic features of the cognitive state dimension back to the origin of the tangent space, completing the spatial transformation. The aligned features in the tangent space are then fused to obtain a cognitive anchor vector that can represent the hierarchical relationship of the cognitive state dimension. Aggregating the scattered hyperbolic features into a single cognitive anchor vector can effectively integrate cognitive state information at different levels in the dialogue text, thereby more clearly representing the hierarchical relationship between these states and providing a structured feature representation for subsequent analysis.

[0079] Step S3034: Using a preset loss function, solve for the hyperbolic position of the hierarchical relationship of the cognitive state dimension in the Poincaré sphere model, and determine the cognitive representation.

[0080] Specifically, step S3034 above includes: Step c1: Determine the preset multi-task classification loss function and hyperbolic geometric regularization loss function, wherein the multi-task classification loss function is used to characterize the semantic classification accuracy in the target dialogue text; Specifically, the multi-task classification loss function Hyperbolic geometric regularization loss function .

[0081] Step c2: Based on the cognitive state dimension, solve the sum of cross-entropy losses of the target dialogue text in the cognitive state dimension to obtain the solution result of the preset multi-task classification loss function; The accuracy of semantic classification is ensured by summing the cross-entropy loss across four dimensions.

[0082] Step c3: Using the Poincaré sphere distance function of the Poincaré sphere manifold, solve for the clustering position and discrete position of the cognitive state dimension of the target dialogue text in hyperbolic space, and obtain the solution result of the hyperbolic geometric regularization loss function. Specifically, the Poincaré sphere distance function is as follows:

[0083] in, This indicates the distance of the Poincaré ball.

[0084] The contrastive loss (hyperbolic geometric regularization loss function) is constructed as follows:

[0085] in, Represents the hyperbolic geometry regularization loss function; Indicates sample The true cognitive label; Indicates sample The true cognitive label; Indicates sample Feature representation in hyperbolic space; Indicates sample Feature representation in hyperbolic space; The preset boundary threshold is used to control the minimum safe distance at which samples with different labels are pushed apart in hyperbolic space. This loss forces samples with the same cognitive label to cluster in hyperbolic space, while samples with different labels are pushed to positions with a distance greater than the edge m, thereby utilizing the exponential capacity of hyperbolic space to achieve hierarchical separation.

[0086] Step c4: Determine the cognitive representation based on the solution results of the preset multi-task classification loss function and the solution results of the hyperbolic geometric regularization loss function.

[0087] Among them, the solution of the hyperbolic geometric regularization loss function represents the clustered position and discrete position of the cognitive state dimension of the target dialogue text in the hyperbolic space, i.e., the cognitive representation.

[0088] The multi-dimensional cognitive state analysis method provided in this embodiment constructs a preset multi-task classification loss function and a hyperbolic geometric regularization loss function. The preset multi-task classification loss function is used to measure the accuracy of the dialogue text in semantic classification, and the hyperbolic geometric regularization loss function is used to measure the positional relationship of different cognitive states in hyperbolic space, thereby determining the cognitive representation of the cognitive state. By combining the accuracy of semantic classification and the geometric structure constraints in hyperbolic space, on the one hand, the multi-task loss ensures that the model can correctly identify different cognitive states; on the other hand, the hyperbolic geometric regularization loss, by distributing different cognitive states in different positions in space, enables the model to simultaneously learn the hierarchical relationship between cognitive states, thereby obtaining a more accurate and structured cognitive representation.

[0089] The multi-dimensional cognitive state analysis method provided in this embodiment parses the target dialogue text and maps it to Euclidean space to determine the Euclidean feature representation of the target dialogue text. Then, it performs hyperbolic manifold projection and hyperbolic feature aggregation to obtain the hierarchical relationship fusion representation of the cognitive state dimension of the target dialogue text. By solving the hyperbolic position of the hierarchical relationship of the cognitive state dimension in the Poincaré sphere model, the cognitive representation is determined. Thus, the problem of hierarchical overlap of cognitive states is solved with an exponentially increasing spatial capacity, which significantly improves the model's ability to jointly model and reason about multi-dimensional cognitive states in complex social scenarios.

[0090] Step S304: Based on the preset large language model, perform semantic training on the cognitive representation to infer a cognitive state profile that satisfies the constraints of cognitive analysis text and cognitive hierarchical structure.

[0091] Specifically, step S304 includes: Step S3041: Construct a parameterized cognitive projection network, which is used to receive the cognitive anchor vector generated by the hyperbolic space network; It should be noted that a parameterized cognitive projection network was designed in this embodiment. The network receives static cognitive anchors generated by the hyperbolic space network. .

[0092] Step S3042: Use the cognitive anchor vector as input to the preset large language model to determine the embedding space dimension representation of the preset large language model; Specifically, it is mapped to the input embedding space dimension of a large model through a multilayer perceptron.

[0093] Step S3043: Based on the preset hyperbolic tangent function, semantically truncate the output of the preset large language model to generate a dynamic soft cue embedding representation of a preset length; Specifically, using the hyperbolic tangent function The output is numerically truncated to prevent the prompt vector from being too large and disrupting the original semantic distribution of the LLM. The generated length is... Dynamic soft hint embedding :

[0094] in, This represents the activation function output after the cognitive anchor vector undergoes the first layer of linear mapping and layer normalization, where... This represents a non-linear activation function (such as ReLU); This represents the weight matrix of the first-layer linear projection network; This represents the weight matrix of the second-layer linear projection network; This represents the scaling factor used to control the magnitude of the soft cue vector.

[0095] Step S3044: According to the input order of first inputting the soft prompt embedding representation and then inputting the target dialogue text, a new input sequence is formed for the preset large language model; Specifically, soft prompts are placed before the original text embedding to form a new input sequence. .

[0096] Step S3045: Input the new input sequence into the preset large language model to perform semantic training on the cognitive representation.

[0097] The multi-dimensional cognitive state analysis method provided in this embodiment constructs a parameterized cognitive projection network to receive cognitive anchor vectors generated by a hyperbolic space network. These anchor vectors are then input into a large language model to obtain their corresponding representations in the model's embedding space. A preset hyperbolic tangent function is used to semantically truncate the output of the preset large language model, generating corresponding dynamic soft prompts. These soft prompts are input first, followed by the dialogue text, to form a new input sequence. This method allows structured cognitive information to be input as contextual information into the model without altering its structure, thereby guiding the model to better utilize deep cognitive state features when generating or understanding text, thus improving the model's performance on complex tasks.

[0098] Step S3046: When training the cognitive representation semantically using the preset large language model, extract the hidden state corresponding to the last layer and map it back to the tangent space dimension of the hyperbolic space network to obtain the semantic vector. Specifically, to prevent the LLM from losing geometric structure information during the generation process, this embodiment introduces topological consistency constraints. First, when the LLM generates each token, it extracts the hidden state corresponding to its last layer. Then use an alignment projector. Will Mapping back to the tangent space dimension of HCN yields semantic vectors. .

[0099] Extracting the hidden state corresponding to the last layer refers to the "feature vector" within the LLM when generating a word or other content. This vector contains the semantic information of the word in the current context. Whenever the model outputs a word, it extracts this "feature vector" generated by the last layer of the neural network during its thought process, ready for the next step of processing; for example, this hidden state represents the thought process behind the word.

[0100] Step S3047: Based on the cosine direction consistency between the semantic vector and the cognitive anchor vector, the cognitive hierarchy structure is inferred. Among them, the core of topological constraints is the calculation of semantic vectors. With cognitive anchors The consistency of the cosine direction between them allows us to deduce the hierarchical structure of cognition.

[0101] Step S3048: When the cosine direction consistency between the semantic vector and the cognitive anchor vector satisfies the maximum cosine similarity constraint, the cognitive state profile is determined to satisfy the cognitive hierarchy structure constraint. Specifically, due to the conformal nature of the Poincaré model, the angular relationships in the tangent space are equivalent to the directional relationships on the manifold. Therefore, maximizing cosine similarity is equivalent to preserving the cognitive hierarchy structure.

[0102] in, This represents a smoothing term (usually a very small constant) used to prevent the denominator from being zero. A semantic vector representing the dimension of the mapping back-cutting space. The transpose of ; Represents cognitive anchor vector The L2 norm (i.e., the magnitude of the vector); express The L2 norm; in, The semantic cognitive topological loss ensures that the generated semantic representation does not deviate from the cognitive hierarchy structure defined by the hyperbolic space, that is, ensures that the generated semantic representation (cognitive state profile) satisfies the cognitive hierarchy structure constraints.

[0103] Step S3049: Determine the preset autoregressive language loss function and infer the cognitive analysis text based on the autoregressive language loss function; in, The standard autoregressive language modeling loss (preset autoregressive language loss function) is used to ensure that LLM can generate fluent and accurate cognitive analysis text, that is, to ensure that the generated semantic representation (cognitive state profile) meets the constraints of cognitive analysis text.

[0104] Step S30410: When it is determined that the cognitive analysis text satisfies the cognitive analysis text accuracy constraint, the cognitive state profile is determined to satisfy the cognitive analysis text constraint.

[0105] The final training process of the system employs a multi-task joint optimization strategy. The hyperbolic space network parameters are kept frozen, and only the low-rank adaptation (LoRA) adapter parameters and projection network parameters of the pre-defined large language model LLM are updated. The total loss function is defined as:

[0106] in, The balancing coefficient is used to adjust the strength of the geometric constraints. During the inference phase, since the geometric prior has been internalized into the LLM through the low-rank adaptation (LoRA) parameter, the system can directly input the original text for prediction without the need for additional hyperbolic space network inference overhead, thus achieving efficient and high-precision multidimensional cognitive state modeling.

[0107] The multi-dimensional cognitive state analysis method provided in this embodiment extracts the hidden states of the last layer of the model, maps them back to hyperbolic space, and determines their cosine direction consistency with the cognitive anchor vector. When the cosine direction consistency between the semantic vector and the cognitive anchor vector satisfies the maximum cosine similarity constraint, it is determined that it satisfies the cognitive hierarchy structure constraint. The generated cognitive analysis text is evaluated by calculating the autoregressive language loss function. When the cognitive analysis text satisfies the cognitive analysis text accuracy constraint, it is determined that it satisfies the cognitive analysis text constraint, and finally, a complete cognitive state profile is inferred. This ensures that the cognitive state profile conforms to the hierarchical relationship between cognitive states in terms of structure, and also ensures that the cognitive state profile can accurately express the analysis results in terms of content, so that the cognitive state profile has both internal structural rationality and external textual expression accuracy.

[0108] The key points of the multi-dimensional cognitive state analysis method provided in this embodiment are mainly reflected in the following aspects: First, this paper, for the first time, specifically reveals the performance bottleneck of large-scale language models in handling complex cognitive tasks from a geometric topological perspective, and proposes the phenomenon of "cognitive crowding." The hierarchical cognitive states of humans have exponentially increasing spatial requirements, which creates a certain degree of geometric mismatch with the Euclidean representation space used by large language models, which only has a polynomial-level growth capacity. Based on this phenomenon, this embodiment innovatively proposes a multi-dimensional cognitive state analysis method, introducing hyperbolic geometry into the semantic space construction of large models, providing a novel approach to solving the representation problem of complex cognitive structures.

[0109] Second, this embodiment constructs a hyperbolic cognitive network (hyperbolic space network) that achieves efficient decoupling of cognitive states on a hyperbolic manifold. This network innovatively employs a Poincaré sphere model as its feature embedding space, leveraging its negative curvature and exponential spatial capacity to losslessly embed deep hierarchical structures. In this embodiment, by introducing a geometrically perceptual contrastive loss function, the hyperbolic space network can force different levels of cognitive concepts (such as concrete emotions and abstract thought processes) to maintain clear boundaries in geometric space, thereby overcoming the feature overlap problem in traditional Euclidean space.

[0110] Third, this embodiment proposes a hyperbolic guided alignment fine-tuning mechanism, constructing an efficient mapping path from geometric priors to semantic generation. This mechanism transforms static geometric anchors into dynamic soft cues through a cognitive projection network, and combines this with a unique semantic-cognitive topological loss to ensure that the internal representation of the large language model remains aligned with the topological structure of the hyperbolic space during training. This not only ensures that the semantic content generated by the model conforms to cognitive logic, but also effectively injects geometric structural knowledge into the model parameters, achieving efficient application with zero additional overhead during the inference phase.

[0111] The multi-dimensional cognitive state analysis method provided in this embodiment has the following beneficial effects: First, a significant performance improvement was achieved in the accuracy of joint modeling of multi-dimensional cognitive states. Experimental results show that it can simultaneously predict emotions, thoughts, stances, and intentions. Figure 4 In extremely challenging tasks, the parameter-based model in the multi-dimensional cognitive state analysis method provided in this embodiment achieves a joint accuracy superior to the baseline model without geometric alignment. Particularly noteworthy is its large parameter count when handling complex cognitive tasks, demonstrating the value of hyperbolic geometric priors in enhancing the model's ability to understand complex structures.

[0112] Secondly, it improves the model's parameter efficiency and inference performance. Through the hyperbolic guided alignment fine-tuning mechanism, this embodiment successfully enables smaller-scale open-source models (such as LLaMA-3.1-8B) to handle deep cognitive logic. Since geometric constraints are internalized into the model parameters during the training phase, no additional hyperbolic network structure needs to be loaded during the inference phase. Therefore, compared with the original large model, this embodiment significantly improves performance while maintaining zero inference latency, making it practically valuable for deployment.

[0113] Finally, this embodiment enhances the interpretability and structuring capabilities of cognitive state analysis. (Gromov) - Hyperbolic analysis shows that the CognitiveBench dataset (domain sample dataset) constructed in this embodiment and the model representation processed by the multi-dimensional cognitive state analysis method both exhibit extremely high hierarchical structure consistency (relative to the model). (≈1%) This means that the model is no longer simply classifying labels, but truly captures the deep logic within the cognitive state of "emotion drives thinking, and thinking determines stance," providing a more transparent and human-cognitive analytical tool for fields such as psychology research and public opinion analysis.

[0114] The following is combined with Figure 4 This describes one application embodiment of the present embodiment.

[0115] This embodiment proposes a multidimensional cognitive state modeling method (multidimensional cognitive state analysis method) based on hyperbolic geometry and cognitive alignment for large language models, named HyCoLLM (Hyperbolic Cognitive Large Language Model). The overall architecture of this system includes three core components: construction of full-dimensional cognitive benchmark data, construction of a hyperbolic cognitive network (hyperbolic space network), and hyperbolic-guided alignment fine-tuning. The core of this embodiment lies in utilizing the geometric properties of hyperbolic space to accommodate hierarchical cognitive states, and aligning this geometric prior into the semantic space of the large language model through soft cues and topological loss.

[0116] The specific implementation plan is as follows: Constructing a full-dimensional cognitive benchmark dataset: This embodiment first constructs a cognitive state classification system covering four key dimensions, decomposing the complex cognitive process into: emotion, thinking style, stance, and intention. These four dimensions do not exist independently but form an organic whole. The specific cognitive state value space is defined as shown in Table 1 of the method embodiment. For example, a cognitive state is the state an individual is in during the cognitive process, encompassing perception, thinking, memory, etc., and is characterized by dimensions such as emotion, thinking style, stance, and intention.

[0117] Based on this classification system (cognitive state classification system), this embodiment collects data from social media platforms on different fields such as trade, elections, cultural diversity, and economic policies. After deduplication, cleaning, and multi-expert cross-annotation, a dataset CognitiveBench (domain sample dataset) containing thousands of high-quality samples is constructed as the basis for subsequent model training and evaluation.

[0118] Construction and pre-training of Hyperbolic Space Network (HCN): This module, as a standalone "geometry teacher" model, is responsible for learning the hierarchical structural features of cognitive states on hyperbolic manifolds.

[0119] Step 11, Text Encoding and Feature Extraction: For the input text sequence The pre-trained Transformer encoder is used to extract contextual semantic features. The mean or average of the last hidden state is then taken. Bit vectors as the basic Euclidean representation .

[0120] Step 12, Hyperbolic manifold projection: To geometrically decouple the four dimensions, the system constructs four independent linear projection layers. First, the basic representation is mapped to a Euclidean tangent space of various dimensions, and then the exponential mapping function is used. Projecting features onto curvature Poincaré spherical manifold superior:

[0121] in is the tangent space vector.

[0122] Step 13, Hyperbolic Feature Aggregation: To obtain a unified representation of cognitive states, feature aggregation is performed in the tangent space. The system aggregates hyperbolic features across four dimensions. Through logarithmic mapping Map back to the origin of the shear space And through attention mechanisms or splicing operations, they are fused into a unified cognitive anchor vector. This vector geometrically encodes a hierarchical relationship in four dimensions.

[0123] Step 14, Geometric Perception Contrast Learning Optimization: The training objective of HCN (Hyperbolic Space Network) consists of two parts: Multi-task classification loss Based on the sum of cross-entropy losses across four dimensions, semantic classification accuracy is ensured.

[0124] Hyperbolic geometric regularization loss Using the Poincaré distance as follows:

[0125] The contrastive loss is constructed as follows:

[0126] This loss forces samples with the same cognitive label to cluster in hyperbolic space, while samples with different labels are pushed to positions at a distance greater than the edge m, thereby utilizing the exponential capacity of hyperbolic space to achieve hierarchical separation.

[0127] Hyperbolic Guided Alignment Adjustment (HGAT): This module is responsible for injecting the geometric priors learned by HCN (Hyperbolic Space Network) into large language models (LLMs, such as LLaMA-3.1-8B). To ensure the parameter efficiency of large models, this embodiment uses low-rank adaptation (LoRA) technology for fine-tuning.

[0128] Step 21, Cognitive Soft Cue Injection: In this embodiment, a parameterized cognitive projection network is designed. The network receives static cognitive anchors generated by HCN. Then, a multilayer perceptron is used to map it to the input embedding space dimension of a large model. Then, the hyperbolic tangent function is used... The output is numerically truncated to prevent the prompt vector from being too large and disrupting the original semantic distribution of the LLM. The generated length is... Dynamic soft hint embedding ;

[0129] By placing the soft cue before the original text embedding, a new input sequence is formed. .

[0130] Step 22, Semantic-Cognitive Topological Alignment: To prevent the LLM from losing geometric structure information during generation, this invention introduces topological consistency constraints. First, when the LLM generates each token, it extracts the hidden state corresponding to its last layer. Then use an alignment projector. Will Mapping back to the tangent space dimension of HCN yields semantic vectors. .

[0131] The core of topological constraints is computation. With cognitive anchors The cosine similarity between them is consistent. Due to the conformal nature of the Poincaré model, the angular relationship in the tangent space is equivalent to the directional relationship on the manifold (hyperbolic manifold). Therefore, maximizing the cosine similarity is equivalent to preserving the cognitive hierarchy structure.

[0132] Step 23, end-to-end joint optimization: The final training process of the system employs a multi-task joint optimization strategy. The HCN parameters are kept frozen, and only the LoRA adapter parameters and projection network parameters of the LLM are updated. The total loss function is defined as:

[0133] in, We use standard autoregressive language modeling loss to ensure that LLM can generate fluent and accurate cognitive analysis text. The semantic-cognitive topological loss ensures that the generated semantic representation does not deviate from the cognitive hierarchy structure defined by the hyperbolic space. The balancing coefficient is used to adjust the strength of the geometric constraints. Through this scheme, during the inference phase, since the geometric prior is internalized into the LLM via the LoRA parameter, the system can directly input the original text for prediction without additional HCN inference overhead, achieving efficient and high-precision multidimensional cognitive state modeling.

[0134] This embodiment provides an example as follows: To verify the effectiveness of the method of this invention in handling complex public opinion analysis tasks, this embodiment constructs a social media controversy topic analysis scenario. This scenario focuses on a user comment regarding statistics on social welfare distribution, specifically refuting the argument of "different ethnic groups' welfare dependence." The system needs to accurately infer the user's deep cognitive state based on logical comparisons of total and per capita data within the text: neutral sentiment, analytical thinking, ambiguous stance, and intent to correct information.

[0135] In the first phase of implementation, the system constructed a comprehensive cognitive benchmark based on the aforementioned raw corpus. The system first identified statistical concepts and rebuttal features cited in the text. Subsequently, through expert cross-annotation, it was determined that although the text contained negative vocabulary, it lacked emotional expression and the evidence was clear. Therefore, the ground truth labels were determined as neutral sentiment, evidence-based analytical thinking, ambiguous stance, and information correction intent. This multi-dimensional label set constituted the foundational ground truth for model training.

[0136] Subsequently, the system initiates a hyperbolic space network to construct geometric priors. The system extracts the basic semantic features of the text and maps them onto a hyperbolic Poincaré sphere manifold. Within this geometric space, the system automatically adjusts the distribution based on curvature characteristics: broad concepts like "neutral sentiment" are mapped near the center of the sphere, while specific concepts such as "analytical thinking" are pushed to the edges. Simultaneously, the system optimizes geometric contrast, forcibly widening the distance between the sample and intuitive thinking samples. Ultimately, the network aggregates these four decoupled features into a unified cognitive anchor vector, encoding the hierarchical logic of "neutrality due to rational analysis" in the geometric structure.

[0137] Next, the system enters the hyperbolic guided alignment fine-tuning stage. The system transforms static cognitive anchors into dynamic soft cue embedding vectors using a cognitive projector and places them before the original text input embedding. When the large language model receives mixed sequences for inference, the system extracts their semantic feature representations in real time. If the model tends to predict negative sentiment due to surface-negative words, the system will detect a significant deviation between the generated semantic feature direction and the cognitive anchor direction.

[0138] Finally, based on this deviation, the system calculates the semantic and cognitive topological consistency loss, updates the low-rank adapter parameters of the large model through backpropagation, and forcibly corrects the model's inference path, causing it to ignore surface interference and focus on underlying logic. After training, the system generates a structured cognitive profile in the final inference stage: accurately indicating that the user did not express anger, but rather analyzed through statistical logic, thus inferring that although their stance refutes the existing narrative, their attitude is ambiguous, and their motivation stems from the pursuit of facts. This demonstrates the significant effect of this invention in capturing deep cognitive logic.

[0139] For example, such as Figure 4 The diagram shows the system architecture of the multi-dimensional cognitive state analysis method.

[0140] The overall system architecture diagram of the large language model multidimensional cognitive state modeling method (multidimensional cognitive state analysis method based on hyperbolic geometry and cognitive alignment) is presented. This architecture diagram logically integrates the entire process from data construction to geometric modeling and then to semantic alignment.

[0141] The left side of the diagram illustrates the dataset construction process. The original corpus comes from social media and conversation logs, and after preprocessing, experts analyze it for emotions, thought patterns, stances, and intentions. Figure 4 The data samples are jointly labeled across multiple dimensions to form a standard data sample that includes context, target content, and structured cognitive tags.

[0142] The upper right section of the diagram details the internal mechanism of the hyperbolic space network, showing how data samples are processed through linear layers and layer normalization, and then projected onto the Poincaré sphere hyperbolic manifold through exponential mapping. The diagram vividly depicts the hierarchical distribution of different cognitive feature points on the hyperbolic surface. These features are then mapped back to the tangent space and fused through a cross-dimensional attention mechanism. Finally, the network is trained with geometric priors under the dual constraints of geometric contrast loss and classification task loss.

[0143] The lower right section of the diagram illustrates the core process of hyperbolic guided alignment fine-tuning. The cognitive anchors generated by the trained network are transformed into soft cues by a cognitive projector and then concatenated with the input of the original text before being fed into the basic large language model. The model's output is jointly constrained by the instruction fine-tuning loss used to optimize prediction accuracy and the semantic cognitive alignment loss used to maintain topological consistency. This allows for the injection of multi-dimensional cognitive capabilities without compromising the model's original semantic capabilities.

[0144] This embodiment also provides a multi-dimensional cognitive state analysis device, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0145] This embodiment provides a multi-dimensional cognitive state analysis device, such as... Figure 5 As shown, it includes: Module 501 is used to construct a cognitive state classification system and a domain sample dataset based on the cognitive state classification system. The first determining module 502 is used to determine the target dialogue text from the target domain sample dataset corresponding to the target domain; wherein, the target domain sample dataset stores the target dialogue text of the target domain; The second determining module 503 is used to learn the hierarchical structure features of the target dialogue text in hyperbolic space based on the domain sample dataset, and to determine the cognitive representation of the cognitive state dimension presented by the target dialogue text in hyperbolic space; wherein the hierarchical structure features characterize the relationship between the preceding and following dialogue text in the target dialogue text in terms of emotion, thinking style, stance, and intention, and the cognitive state dimension includes at least: emotion, thinking style, stance, and intention. Figure 4 One dimension; Training module 504 is used to perform semantic training on cognitive representations based on a preset large language model, and to infer a cognitive state profile that satisfies the constraints of cognitive analysis text and cognitive hierarchical structure.

[0146] In some alternative implementations, the second determining module 503 includes: The extraction unit is used to extract the semantic features corresponding to the preceding and following dialogue text in the target dialogue text and map them to Euclidean space to determine the Euclidean feature representation of the target dialogue text.

[0147] The projection unit is used to perform hyperbolic manifold projection on the Euclidean feature representation based on the Euclidean tangent space to obtain the hyperbolic feature representation of the cognitive state dimension of the target dialogue text.

[0148] In some alternative implementations, the projection unit includes: Construct sub-units to build hyperbolic space networks and linear projection layers on hyperbolic space, wherein the linear projection layers include at least an emotion linear projection layer, a thinking style linear projection layer, a stance linear projection layer, and an intention linear projection layer; The mapping subunit is used to map Euclidean feature representations to Euclidean tangent space representations in the cognitive state dimension based on hyperbolic space networks. The projection subunit is used to project the Euclidean tangent space representation of the cognitive state dimension onto the Poincaré spherical manifold based on the linear projection layer and the preset exponential mapping function, thereby determining the hyperbolic features of the target dialogue text in the cognitive state dimension.

[0149] The aggregation unit is used to aggregate hyperbolic features based on the tangent space to obtain a hierarchical relational fusion representation of the cognitive state dimension of the target dialogue text.

[0150] In some alternative implementations, the aggregation unit includes: The feature alignment subunit is used to map the hyperbolic features of the cognitive state dimension back to the origin of the tangent space based on a preset logarithmic function, so as to obtain the tangent space origin aligned representation. The fusion subunit is used to perform cognitive state space fusion on the tangent space origin-aligned representation to obtain the fused cognitive anchor vector. The cognitive anchor vector represents the hierarchical relationship of the cognitive state dimension of the target dialogue text.

[0151] The solver unit is used to solve for the hyperbolic position of the hierarchical relationship of the cognitive state dimension in the Poincaré sphere model using a preset loss function, and to determine the cognitive representation.

[0152] In some alternative implementations, the solving unit includes: The loss function determination subunit is used to determine the preset multi-task classification loss function and hyperbolic geometric regularization loss function, wherein the multi-task classification loss function is used to characterize the semantic classification accuracy in the target dialogue text; The first solution subunit is used to solve the sum of the cross-entropy loss of the target dialogue text in the cognitive state dimension based on the cognitive state dimension to obtain the solution result of the preset multi-task classification loss function. The second solution subunit is used to solve for the clustering position and discrete position of the cognitive state dimension of the target dialogue text in hyperbolic space using the Poincaré sphere distance function of the Poincaré sphere manifold, and obtain the solution result of the hyperbolic geometric regularization loss function. The cognitive representation determination sub-unit is used to determine the cognitive representation based on the solution results of the preset multi-task classification loss function and the solution results of the hyperbolic geometric regularization loss function.

[0153] In some alternative implementations, training module 504 includes: The cognitive projection network building unit is used to construct a parameterized cognitive projection network, which is used to receive the cognitive anchor vector generated by the hyperbolic space network. Embedding units are used to take cognitive anchor vectors as input to a pre-defined large language model and determine the dimensional representation of the embedding space of the pre-defined large language model. The semantic processing unit is used to perform semantic truncation on the output of the preset large language model based on the preset hyperbolic tangent function, and generate a dynamic soft prompt embedding representation of a preset length; The sorting unit is used to form a new input sequence for the pre-defined large language model according to the input order of first inputting soft prompt embeddings and then inputting the target dialogue text; The training unit is used to input new input sequences into a pre-defined large language model to perform semantic training on cognitive representations.

[0154] The state extraction unit is used to extract the hidden state corresponding to the last layer and map it back to the tangent space dimension of the hyperbolic space network to obtain the semantic vector when the cognitive representation is semantically trained by the preset large language model. The reasoning unit is used to infer the cognitive hierarchy based on the cosine direction consistency between the semantic vector and the cognitive anchor vector. The cognitive hierarchy structure constraint unit is used to determine that the cognitive state profile satisfies the cognitive hierarchy structure constraint when the cosine direction consistency between the determined semantic vector and the cognitive anchor vector satisfies the maximum cosine similarity constraint. The cognitive analysis text constraint unit is used to determine the preset autoregressive language loss function and infer cognitive analysis text based on the autoregressive language loss function; The cognitive profile determination unit is used to determine whether the cognitive state profile satisfies the cognitive analysis text constraints when the cognitive analysis text satisfies the cognitive analysis text accuracy constraints.

[0155] The multi-dimensional cognitive state analysis device provided in this embodiment of the invention can execute the multi-dimensional cognitive state analysis method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.

[0156] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0157] The following is a detailed reference. Figure 6 This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 6604. An input / output (I / O) interface 605 is also connected to bus 604.

[0158] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0159] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in the multi-dimensional cognitive state analysis method of the embodiments of the present invention.

[0160] Figure 6The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0161] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the multi-dimensional cognitive state analysis method shown in the above embodiments is implemented.

[0162] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0163] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A multi-dimensional cognitive state analysis method, characterized in that, The method includes: Construct a cognitive state classification system and a domain sample dataset based on the cognitive state classification system; The target dialogue text is determined from the target domain sample dataset corresponding to the target domain; wherein, the target domain sample dataset stores the target dialogue text of the target domain; Based on the domain sample dataset, the hierarchical structure features of the target dialogue text in hyperbolic space are learned, and the cognitive state dimension presented by the target dialogue text is determined as a cognitive representation in hyperbolic space; wherein the hierarchical structure features characterize the relationship between the preceding and following dialogue texts in the target dialogue text in terms of emotion, thinking style, stance and intention, and the cognitive state dimension includes at least four dimensions: emotion, thinking style, stance and intention. Based on a pre-defined large language model, the cognitive representation is semantically trained to infer a cognitive state profile that satisfies the constraints of cognitive analysis text and cognitive hierarchical structure, including: When the preset large language model performs semantic training on the cognitive representation, the hidden state corresponding to the last layer is extracted and mapped back to the tangent space dimension of the hyperbolic space network to obtain the semantic vector. Based on the cosine direction consistency between the semantic vector and the cognitive anchor vector, the cognitive hierarchy structure is inferred. When it is determined that the cosine direction consistency between the semantic vector and the cognitive anchor vector satisfies the maximum cosine similarity constraint, it is determined that the cognitive state profile satisfies the cognitive hierarchy structure constraint. A preset autoregressive language loss function is determined, and cognitive analysis text based on the autoregressive language loss function is inferred; When it is determined that the cognitive analysis text satisfies the cognitive analysis text accuracy constraint, it is determined that the cognitive state profile satisfies the cognitive analysis text constraint.

2. The method according to claim 1, characterized in that, The process of learning the hierarchical structural features of the target dialogue text in hyperbolic space based on the domain sample dataset, and determining the cognitive representation of the cognitive state dimension presented by the target dialogue text in hyperbolic space, includes: Extract the semantic features corresponding to the preceding and following dialogue texts in the target dialogue text and map them to Euclidean space to determine the Euclidean feature representation of the target dialogue text. Hyperbolic manifold projection is performed on the Euclidean feature representation based on the Euclidean tangent space to obtain the hyperbolic feature representation of the cognitive state dimension of the target dialogue text. Hyperbolic feature aggregation is performed on the hyperbolic feature representation based on the tangent space to obtain the hierarchical relationship fusion representation of the cognitive state dimension of the target dialogue text; Using a preset loss function, the hyperbolic position of the hierarchical relationship of the cognitive state dimension is obtained in the Poincaré sphere model, thereby determining the cognitive representation.

3. The method according to claim 2, characterized in that, The hyperbolic manifold projection of the Euclidean feature representation onto the Euclidean tangent space to obtain the hyperbolic feature representation of the cognitive state dimension of the target dialogue text includes: Construct a hyperbolic space network and linear projection layers on hyperbolic space, wherein the linear projection layers include at least an emotion linear projection layer, a thinking style linear projection layer, a stance linear projection layer, and an intention linear projection layer; Based on the hyperbolic space network, the Euclidean feature representation is mapped to the Euclidean tangent space of the cognitive state dimension to obtain the Euclidean tangent space representation; Based on the linear projection layer and the preset exponential mapping function, the Euclidean tangent space representation of the cognitive state dimension is projected onto the Poincaré spherical manifold to determine the hyperbolic features of the target dialogue text in the cognitive state dimension.

4. The method according to claim 3, characterized in that, The hyperbolic feature aggregation based on the tangent space of the hyperbolic feature representation to obtain the hierarchical relation fusion representation of the cognitive state dimension of the target dialogue text includes: Based on a preset logarithmic function, the hyperbolic features of the cognitive state dimension are mapped back to the origin of the tangent space to obtain the tangent space origin-aligned representation. The cognitive state space is fused by the original alignment representation of the tangent space to obtain the fused cognitive anchor vector, wherein the cognitive anchor vector represents the hierarchical relationship of the cognitive state dimension of the target dialogue text.

5. The method according to claim 2, characterized in that, The process of using a preset loss function to solve for the hyperbolic position of the hierarchical relationship of the cognitive state dimension in the Poincaré sphere model, and determining the cognitive representation, includes: A preset multi-task classification loss function and a hyperbolic geometric regularization loss function are determined, wherein the multi-task classification loss function is used to characterize the semantic classification accuracy in the target dialogue text; Based on the cognitive state dimension, the solution of the preset multi-task classification loss function is obtained by summing the cross-entropy losses of the target dialogue text in the cognitive state dimension. Using the Poincaré sphere distance function of the Poincaré sphere manifold, the clustering position and discrete position of the cognitive state dimension of the target dialogue text in hyperbolic space are solved, and the solution of the hyperbolic geometric regularization loss function is obtained. The cognitive representation is determined based on the solution results of the preset multi-task classification loss function and the solution results of the hyperbolic geometric regularization loss function.

6. The method according to claim 4, characterized in that, The semantic training of the cognitive representation based on the preset large language model includes: Construct a parameterized cognitive projection network, which is used to receive the cognitive anchor vector generated by the hyperbolic space network; Using the cognitive anchor vector as input to the preset large language model, the embedding space dimension representation of the preset large language model is determined; Based on the preset hyperbolic tangent function, the output of the preset large language model is semantically truncated to generate a dynamic soft cue embedding representation of a preset length; The new input sequence of the preset large language model is formed by first inputting the soft prompt embedding representation and then inputting the target dialogue text. The new input sequence is input into the preset large language model to perform semantic training on the cognitive representation.

7. A multi-dimensional cognitive state analysis device, characterized in that, The device includes: A construction module is used to construct a cognitive state classification system and a domain sample dataset based on the cognitive state classification system; The first determining module is used to determine the target dialogue text from the target domain sample dataset corresponding to the target domain; wherein, the target domain sample dataset stores the target dialogue text of the target domain; The second determining module is used to learn the hierarchical structure features of the target dialogue text in hyperbolic space based on the domain sample dataset, and determine the cognitive representation of the cognitive state dimension presented by the target dialogue text in hyperbolic space; wherein the hierarchical structure features characterize the relationship between the emotion, thinking mode, stance and intention between the dialogue context and the dialogue context in the target dialogue text, and the cognitive state dimension includes at least four dimensions: emotion, thinking mode, stance and intention. The training module is used to perform semantic training on the cognitive representation based on a preset large language model, and to infer a cognitive state profile that satisfies the constraints of cognitive analysis text and cognitive hierarchical structure. The training module includes: The state extraction unit is used to extract the hidden state corresponding to the last layer and map it back to the tangent space dimension of the hyperbolic space network to obtain the semantic vector when the cognitive representation is semantically trained by the preset large language model. The reasoning unit is used to infer the cognitive hierarchy based on the cosine direction consistency between the semantic vector and the cognitive anchor vector. The cognitive hierarchy structure constraint unit is used to determine that the cognitive state profile satisfies the cognitive hierarchy structure constraint when the cosine direction consistency between the determined semantic vector and the cognitive anchor vector satisfies the maximum cosine similarity constraint. The cognitive analysis text constraint unit is used to determine the preset autoregressive language loss function and infer cognitive analysis text based on the autoregressive language loss function; The cognitive profile determination unit is used to determine whether the cognitive state profile satisfies the cognitive analysis text constraints when the cognitive analysis text satisfies the cognitive analysis text accuracy constraints.

8. An electronic device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the method of any one of claims 1 to 6.

Images