Dynamic rule base construction method, device and medium based on semantic connection computing

Abstract

This application discloses a method, device, and medium for constructing a dynamic rule base based on semantic connection calculation, relating to the field of artificial intelligence technology. The method includes: constructing a manifold subspace of an initial concept manifold to obtain a concept manifold; determining a continuous thought chain based on the concept manifold through differentiability of concept neighborhoods; dividing the language field into a token vector field and a semantic connection field based on the continuous thought chain; optimizing semantic relationships based on the semantic connection field to obtain a semantic field arrangement sentence; and performing a predicted probability analysis of the next token in the semantic field arrangement sentence based on the token vector field to determine the probability of the next token being generated. This application addresses the technical problem of poor universality of large-scale model inference and judgment using the above method.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, and medium for constructing a dynamic rule base based on semantic connection computing. Background Art

[0002] The "reasoning" capabilities of large language models are based on statistical pattern recognition, possessing implicit reasoning and emergence capabilities. However, they lack true causal reasoning and are subject to hallucinations and difficulties with multi-step reasoning. Strictly speaking, large language models themselves lack "reasoning" capabilities. For the time being, they can be considered a form of "inference," still relying on cognitive approximations based on correlations. Therefore, building true reasoning capabilities in large models is a top priority.

[0003] Currently, integrating symbolic logic into neural network training presents significant technical challenges, and the complex and diverse constraints corresponding to rule-based judgments in various scenarios present challenges that must be addressed. The industry is still exploring the implementation of weak / strong coupling for neural symbolic logic. Technical solutions such as neural symbolic programming, neural rule engines, and differentiable transformers are all areas of ongoing research. The difficulties in widespread adoption of large-scale model inference and judgment in existing technologies can be attributed to the following: large models do not understand the abstract process of a concept, probabilistic models cannot effectively adapt to all conceptual reasoning, and precision is difficult to embed. These issues are caused by the fact that each concept has its own sub-attributes, each concept has its own context and scope of application, and precision in thought chains is difficult to embed. Summary of the Invention

[0004] The embodiments of the present application provide a method, device and medium for constructing a dynamic rule base based on semantic connection calculation, which solves the technical problem of poor universality of large model inference judgment.

[0005] In a first aspect, an embodiment of the present application provides a method for constructing a dynamic rule base based on semantic connection calculation, characterized in that the method includes: constructing a manifold subspace of an initial concept manifold to obtain a concept manifold; based on the concept manifold, determining a continuous thinking chain through concept neighborhood differentiability; according to the continuous thinking chain, dividing the language field into a token vector field and a semantic connection field; according to the semantic connection field, obtaining a semantic field arrangement sentence through semantic relationship optimization; based on the token vector field, performing a predicted probability analysis of the next token on the semantic field arrangement sentence to determine the probability of generating the next token.

[0006] In one implementation of the present application, a manifold subspace is constructed on the initial concept manifold to obtain a concept manifold, specifically including: obtaining a local active coordinate frame of the initial concept manifold, and selecting a sub-attribute of the concept as a unit vector of the local active coordinate frame to determine the local concept space; semantically mapping the local concept space into an N-dimensional space to obtain a concept manifold to be defined; wherein, the N-dimensional space includes: N-dimensional Euclidean space and geometric space; according to a preset application scenario, a manifold definition is performed on the concept manifold to be defined to obtain a concept manifold; wherein, the types of concept manifolds include: parameterized manifolds, data manifolds, and concept manifolds.

[0007] In one implementation of the present application, based on the concept manifold, a continuous thought chain is determined through concept neighborhood differentiability, specifically including: determining a token chain based on the concept manifold, and screening the token chain for text coherence to obtain an incoherent token chain; determining the disconnected area of ​​the incoherent token chain, and calling the latent space based on the disconnected area; based on the latent space, performing differentiable supplementation of the disconnected area across the neighborhood latent space to determine a continuous thought chain.

[0008] In one implementation of the present application, the language field is divided into a token vector field and a semantic connection field according to a continuous thought chain, specifically including: setting a word-unit function, and based on the continuous thought chain and the word-unit function, constructing a semantic connection field through word-unit semantic organization; according to the semantic connection field, constructing a token vector field through token state vector analysis.

[0009] In one implementation of the present application, a semantic field arrangement sentence is obtained based on the semantic connection field through semantic relationship optimization, which specifically includes: obtaining word units, and based on the word units, defining the geometric structure of the semantic manifold; constructing the connection coefficient of the local coordinate system of the semantic manifold after the geometric structure definition to determine the semantic connection features; and arranging and organizing the semantic connection features of the word units to obtain the semantic field arrangement sentence.

[0010] In one implementation of the present application, based on the token vector field, a predicted probability analysis of the next token is performed on the semantic field arrangement sentence to determine the probability of the next token being generated, specifically including: based on the token vector field, a predicted probability analysis of the next token is performed on the semantic field arrangement sentence to obtain the generation status of the next token; obtaining a word-unit function and a word-unit complex function corresponding to the word-unit function, and determining the probability of the next token being generated through token collapse analysis based on the word-unit function and the word-unit complex function.

[0011] In one implementation of the present application, the probability of generating the next token is determined through token collapse analysis based on the word-gram function and the word-gram complex function, specifically including: when the token prediction state is that the next token is not generated, the token prediction state is uncertain; when the token prediction state is that the next token is generated, the token prediction state is collapsed; when the token prediction state is that the collapse occurs, the field function corresponding to the semantic field arrangement sentence is determined based on the word-gram function and the word-gram complex function, and the square of the modulus of the field function of the next token and the field function corresponding to the semantic field arrangement sentence is set as the probability of generating the next token.

[0012] In one implementation of the present application, based on the token vector field, a predicted probability analysis of the next token is performed on the semantic field arrangement statement to determine the probability of generating the next token. The method also includes: constructing a semantic dynamic rule library based on the token vector field and the semantic connection field.

[0013] In the second aspect, an embodiment of the present application also provides a dynamic rule base construction device based on semantic connection calculation, characterized in that the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: construct a manifold subspace of the initial concept manifold to obtain a concept manifold; based on the concept manifold, determine a continuous thinking chain through concept neighborhood differentiability; according to the continuous thinking chain, divide the language field into a token vector field and a semantic connection field; according to the semantic connection field, obtain a semantic field arrangement sentence through semantic relationship optimization; based on the token vector field, perform a predicted probability analysis of the next token on the semantic field arrangement sentence to determine the probability of generating the next token.

[0014] In a third aspect, an embodiment of the present application also provides a non-volatile computer storage medium for a dynamic rule base construction method based on semantic connection calculation, which stores computer executable instructions, and is characterized in that the computer executable instructions are set to: construct a manifold subspace of the initial concept manifold to obtain a concept manifold; based on the concept manifold, determine a continuous thinking chain through concept neighborhood differentiability; according to the continuous thinking chain, divide the language field into a token vector field and a semantic connection field; according to the semantic connection field, obtain a semantic field arrangement sentence through semantic relationship optimization; based on the token vector field, perform a predicted probability analysis of the next token on the semantic field arrangement sentence to determine the probability of generating the next token.

[0015] The embodiments of the present application provide a method, device and medium for constructing a dynamic rule base based on semantic connection calculation, which improves the data accuracy of the spatial framework through a unified neural symbolic modeling method based on intrinsic geometry, dynamically fills in missing tokens through a continuous reasoning method in a cross-neighborhood latent space, and performs token prediction and word arrangement and organization analysis through a unified language field construction method based on semantic connection. It solves the technical problem of poor universality of large model inference judgment, realizes the construction of concept tangent space, the differentiability of neighborhood and the dynamic formation of continuous reasoning chains across knowledge fields, and can be applied to scenarios such as compliance audit and source control. In addition, computing semantic connections can realize the self-assembly and self-construction of knowledge structures, mine strong correlations of information, and be applied to scenarios such as compliance analysis and early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0017] Figure 1 A flow chart of a method for constructing a dynamic rule base based on semantic connection calculation provided in an embodiment of the present application;

[0018] Figure 2 A schematic diagram of constructing a conceptual manifold provided in an embodiment of the present application;

[0019] Figure 3 A schematic diagram of differentiability of a conceptual neighborhood provided in an embodiment of the present application;

[0020] Figure 4 A schematic diagram of differentiating states of a language field provided in an embodiment of the present application;

[0021] Figure 5 A schematic diagram of the internal structure of a dynamic rule base construction device based on semantic connection computing provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0023] The embodiments of the present application provide a method, device and medium for constructing a dynamic rule base based on semantic connection calculation, which improves the data accuracy of the spatial framework through a unified neural symbolic modeling method based on intrinsic geometry, dynamically fills in missing tokens through a continuous reasoning method in a cross-neighborhood latent space, and performs token prediction and word arrangement and organization analysis through a unified language field construction method based on semantic connection. It solves the technical problem of poor universality of large model inference judgment, realizes the construction of concept tangent space, the differentiability of neighborhood and the dynamic formation of continuous reasoning chains across knowledge fields, and can be applied to scenarios such as compliance audit and source control. In addition, computing semantic connections can realize the self-assembly and self-construction of knowledge structures, mine strong correlations of information, and be applied to scenarios such as compliance analysis and early warning.

[0024] The technical solutions proposed in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0025] Figure 1 A flow chart of a method for constructing a dynamic rule base based on semantic connection calculation provided in an embodiment of the present application. Figure 1 As shown, the embodiment of the present application provides a method for constructing a dynamic rule base based on semantic connection calculation, which specifically includes the following steps:

[0026] Step 101: construct a manifold subspace on the initial concept manifold to obtain a concept manifold.

[0027] For example, in order to no longer need to construct an external coordinate system to study the properties of the conceptual manifold, only a local active coordinate frame is established on the surface of the manifold. This application constructs a manifold subspace of the initial conceptual manifold to obtain the conceptual manifold, thereby getting rid of the dependence on the coordinate system and directly describing the local geometric properties of the conceptual manifold based on the geometric structure of the model space. It also achieves data dimensionality reduction through intrinsic geometry, so that the data can be better analyzed, visualized and understood while maintaining its geometric characteristics.

[0028] Specifically, a manifold subspace is constructed on the initial concept manifold to obtain a concept manifold, including: obtaining a local active coordinate frame of the initial concept manifold, and selecting a sub-attribute of the concept as a unit vector of the local active coordinate frame to determine the local concept space; semantically mapping the local concept space into an N-dimensional space to obtain a concept manifold to be defined; wherein, the N-dimensional space includes: an N-dimensional Euclidean space and a geometric space; according to a preset application scenario, a manifold definition is performed on the concept manifold to be defined to obtain a concept manifold; wherein, the types of concept manifolds include: parameterized manifolds, data manifolds, and concept manifolds.

[0029] In one embodiment, to incorporate continuous reasoning capabilities, a concept manifold is established. This concept manifold lacks external coordinates and instead has a local active coordinate frame. The sub-attributes of a concept are used as unit vectors of the local active coordinate frame. All concept points within a concept region can be represented in the local active coordinate frame. The local space of an n-dimensional manifold is homeomorphic to an n-dimensional Euclidean space. Figure 2 A schematic diagram of constructing a conceptual manifold provided in an embodiment of the present application is shown as follows: Figure 2 As shown, in the subspace of the concept manifold, "penguin, swallow" are both "bird", and the concept of "bird" includes "beak, wings, feathers"; similarly, the subspace relationship of insects is similar and will not be repeated here.

[0030] After the construction of the original concept manifold is completed, many of the concepts are distributed in different subspaces. At this time, the construction of the subspace is not perfect enough, so it is necessary to select the sub-attributes of the concept as the unit vectors of the local active coordinate frame to determine the local concept space, which can not only get rid of the coordinate restrictions, but also meet the construction of the concept level in the local manifold (such as primary, secondary, etc.).

[0031] It should be noted that the construction of conceptual manifolds is a manifestation of intrinsic geometry. High-dimensional data distribution usually has a certain geometric structure. In the specific process of analyzing the structure, algorithms such as manifold learning can be used to reduce its dimensionality and explain its intrinsic geometric structure.

[0032] Further, As a representation of the concept manifold, is the dimension of the manifold, The essence represents real numbers (determined by the mathematical concept of manifold); at this point, the conceptual manifold is constructed, and for the concepts of "insect" and "bird", there is also a common feature (intersection) that they can fly.

[0033] pass 、 、 Mapping the concept homeomorphism to Euclidean geometric space, we get three ; three They can also be mapped to each other, that is, and . and yes 、 The inverse of the homeomorphic mapping can be offset in the mutual mapping process. 、 ; Finally, the middle circle After the mutual mapping is completed, , representing the intersection of the concepts “insect” and “bird” as “fly”.

[0034] It should be noted that in a smooth manifold, if the coordinate mapping in a very small space is performed, since the curvature tends to 0, it can be mapped through mapping methods such as three-dimensional geometry.

[0035] Furthermore, the local concept space can be studied in n-dimensional Euclidean space, and the definition of manifold can be diversified according to the application scenario.

[0036] Parameterized manifolds: For example, a family of probability distributions in statistical models (e.g., the parameter space of a Gaussian distribution), whose tangent space describes the effects of small perturbations of the parameters.

[0037] Data manifold: After a high-dimensional dataset is embedded in a low-dimensional space through dimensionality reduction (such as t-SNE and UMAP in manifold learning), its local linear structure is characterized by the tangent space.

[0038] Concept manifold: Abstract semantics (such as lexical semantics and cognitive structures) can be modeled as manifolds through geometrization.

[0039] This application adopts a neural symbolic unified modeling method based on intrinsic geometry. It no longer studies the properties of the concept manifold by constructing an external coordinate system, but only establishes a local active coordinate frame on the surface of the manifold. This method gets rid of the dependence on the coordinate system and directly describes the local geometric properties of the concept manifold based on the geometric structure of the model space. It focuses on processing high-dimensional data and revealing its potential low-dimensional manifold structure. Its core idea is to assume that high-dimensional data samples are distributed on a potential low-dimensional manifold. Even if the data is in a high-dimensional space, it can be described by a lower-dimensional structure. This method drives the manifold learning algorithm, which aims to discover the intrinsic low-dimensional representation of the data, so that the data can be better analyzed, visualized and understood while maintaining its geometric characteristics. This method is conducive to studying the self-driving mechanism of changes in language space structure and studying the source of changes in semantic fields.

[0040] Step 102: Based on the concept manifold, determine the continuous thinking chain through concept neighborhood differentiability.

[0041] Exemplarily, this application is based on the concept manifold, and through the differentiability of the concept neighborhood, it determines the continuous thinking chain, realizes the supplement of the discontinuous space in the concept manifold, and the model performs reasoning in the latent space without generating intermediate language expressions, skipping the discontinuous areas in the token space, thereby improving the overall stability of the concept manifold.

[0042] Specifically, based on the concept manifold, the continuous thought chain is determined through the differentiability of the concept neighborhood, including: determining the token chain based on the concept manifold, and screening the token chain for text coherence to obtain an incoherent token chain; determining the disconnected area of ​​the incoherent token chain, and calling the latent space based on the disconnected area; based on the latent space, performing differentiable supplementation of the disconnected area across the neighborhood latent space to determine the continuous thought chain.

[0043] In one embodiment, a significant problem arises when large language models (LLMs) reason with language: the amount of reasoning required for each specific inference token varies significantly, yet current LLM architectures allocate nearly identical computational budgets to predicting each token. These issues have been addressed by forcing the LLM to generate concise reasoning chains or performing additional reasoning before generating a few key tokens. However, these solutions remain constrained to the language space and do not address the fundamental problem. Ideally, the LLM would be able to reason freely, without any linguistic constraints, and then translate its findings into language only when necessary. LLMs are limited to reasoning in "language space," typically using chains of thought (CoTs) to express reasoning processes and solve complex reasoning problems. However, the language space may not always be the optimal space for reasoning. Most tokens in the reasoning chain are generated simply for fluency and contribute little to the actual reasoning process.

[0044] Figure 3 A schematic diagram of a conceptual neighborhood differentiability provided in an embodiment of the present application.

[0045] Most word tokens are primarily used for text coherence and are not essential for reasoning. However, some key tokens require complex planning, posing a significant challenge for LLMs. To explore the potential of LLMs for reasoning in an unrestricted latent space, rather than using natural language, a new paradigm, Coconut (continuous chain of thought), is introduced. The "hidden state" generated by a Transformer layer is converted into a token. This conversion process compresses information from an infinite mathematical space consisting of continuous values ​​to a finite set of only tens of thousands of discrete words. This "dimensionality reduction" not only results in efficiency losses but can also compress or distort the model's internal representation. To address this, the model architecture is modified: bypassing the token generation step, the hidden state is looped directly back into the input embedding and processed again through the Transformer layers. This allows the model to always process information within a continuous mathematical space, without being constrained by the discrete boundaries imposed by human language.

[0046] The construction of the concept manifold realizes the construction without the need for an external coordinate frame, the differentiability of the concept neighborhood realizes the supplementary processing of problems such as discontinuous language data in the framework, and the latent space is used to realize the supplement of the discontinuous part of the token space.

[0047] This application uses a continuous reasoning method in a cross-neighborhood latent space to construct a language model that primarily reasones in the latent space. The trained model not only learns to reason in the latent space but also enables it to autonomously determine when to end reasoning and return to language generation. The outputs of the model's intermediate modules are not immediately converted into tokens but are instead fed back into the module as embedded vectors for further processing. The model autonomously learns how many times to perform cyclical reasoning for different tasks, and reasoning ends when the changes in the embedded vectors of the cyclic outputs stabilize. In this process, the model gradually establishes connections between tokens. The final layer synthesizes all information to generate a new sequence of embedding vectors. The model outputs multiple tokens, gradually demonstrating its solution path. The model's last hidden state is considered a representation of "continuous thought" and is directly used as the embedding for the next input, rather than being decoded into a token. This allows the model to reason in the latent space without generating intermediate language expressions, skipping discontinuous regions in the token space. This not only helps reveal the model's reasoning mechanism but also significantly improves the accuracy of the answers.

[0048] Step 103: Based on the continuous thought chain, the language field is divided into the token vector field and the semantic connection field.

[0049] For example, this application proposes the concept of a language field, dividing it into a token vector field and a semantic connection field. Tokens serve as the foundational building blocks of a language, while the semantic connection field organizes words into sentences with fixed semantics. The large model predicts the probability of the next token in the token vector field.

[0050] Specifically, according to the continuous thought chain, the language field is divided into the token vector field and the semantic connection field, including: setting the word function, and based on the continuous thought chain and the word function, constructing the semantic connection field through word semantic organization; according to the semantic connection field, constructing the token vector field through the token state vector analysis.

[0051] Step 104: According to the semantic connection field, the semantic relationship is optimized to obtain the semantic field arrangement sentence.

[0052] Specifically, according to the semantic connection field, a semantic field arrangement sentence is obtained through semantic relationship optimization, including: obtaining word units, and based on the word units, defining the geometric structure of the semantic manifold; constructing the connection coefficient of the local coordinate system of the semantic manifold after the geometric structure definition to determine the semantic connection features; arranging and organizing the semantic connection features of the word units to obtain the semantic field arrangement sentence.

[0053] In one embodiment, semantic connections are the specific implementation of affine connections on manifolds in semantic space, used to describe the covariant derivatives and parallel movement rules of vector fields in semantic space. The core of this approach lies in defining the geometric structure of semantic manifolds (such as word embedding spaces and concept spaces) and constructing connection coefficients in a local coordinate system. The semantic space is modeled as a smooth manifold, whose local coordinates correspond to semantic features (such as the dimensions of word vectors and the abstract properties of concepts). For example, in the Word2Vec model, the word embedding space can be viewed as a submanifold of the semantic space, reflecting the semantic relationships between words. Computing semantic connections is a core task in geometric semantics, requiring the integration of differential geometry, statistical learning, and domain knowledge. Within a given semantic field, word units interact with each other through the medium of the semantic field, being organized and arranged by it to form sentences with specific semantic meanings.

[0054] The semantic field arrangement statement indicates the state in which the words should be arranged and the output result under the configuration of the semantic field.

[0055] Step 105: Based on the token vector field, perform a prediction probability analysis on the next token of the semantic field arrangement sentence to determine the generation probability of the next token.

[0056] For example, a word function is a vector in complex Hilbert space. The probability of the model generating the next token based on the previous sentence is the square of the modulus of the inner product of the next token vector and the previous sentence vector. Before the next token is generated, its state is uncertain. When a token is generated, the state collapses. The word function rotates under the influence of the semantic field, occupying different semantic states. The modulus of the word function is its fundamental meaning.

[0057] Specifically, based on the token vector field, the predicted probability analysis of the next token is performed on the semantic field arrangement sentence to determine the probability of the next token being generated, specifically including: based on the token vector field, the predicted probability analysis of the next token is performed on the semantic field arrangement sentence to obtain the generation status of the next token; obtaining the word-unit function and the word-unit complex function corresponding to the word-unit function, and determining the probability of the next token being generated through token collapse analysis based on the word-unit function and the word-unit complex function.

[0058] Furthermore, based on the word-gram function and the word-gram complex function, the probability of the next token being generated is determined through token collapse analysis, specifically including: when the token prediction state is that the next token has not been generated, the token prediction state is uncertain; when the token prediction state is that the next token has been generated, the token prediction state is collapsed; when the token prediction state is that the collapse has occurred, based on the word-gram function and the word-gram complex function, the field function corresponding to the semantic field arrangement sentence is determined, and the square of the modulus of the field function of the next token and the field function corresponding to the semantic field arrangement sentence is set as the probability of the next token being generated.

[0059] Furthermore, based on the token vector field, a prediction probability analysis of the next token is performed on the semantic field arrangement sentence to determine the probability of the next token being generated. The method also includes: constructing a semantic dynamic rule library based on the token vector field and the semantic connection field.

[0060] Figure 4 A schematic diagram of the differentiating states of a language field provided in an embodiment of the present application.

[0061] In one embodiment, the word function is a vector in complex Hilbert space. The probability of the model generating the next token based on the previous sentence is the square of the modulus of the inner product of the next token vector and the previous sentence vector. Before the next token is generated, its state is uncertain. When a token is generated, its state collapses. The word function rotates under the influence of the semantic field, occupying different semantic states.

[0062] Therefore, the large model predicts that the probability of the next token being generated is the square of the modulus of the inner product of the field function of the next token and the field function of the previous statement (which is consistent with the next token in the quantum superposition state, where observation will collapse and its state cannot be determined without observation).

[0063] The above is an embodiment of the method proposed in this application. Based on the same inventive concept, this application embodiment also provides a dynamic rule base construction device based on semantic connection calculation, whose structure is as follows: Figure 5 shown.

[0064] Figure 5 The internal structure diagram of a dynamic rule base construction device based on semantic connection calculation provided in the embodiment of the present application. Figure 5 As shown, the equipment includes:

[0065] at least one processor 501;

[0066] and, a memory 502 in communication with the at least one processor;

[0067] The memory 502 stores instructions that can be executed by at least one processor, and the instructions are executed by the at least one processor 501 to enable the at least one processor 501 to:

[0068] The manifold subspace of the initial concept manifold is constructed to obtain the concept manifold; based on the concept manifold, the continuous thought chain is determined through the differentiability of the concept neighborhood; according to the continuous thought chain, the language field is divided into the token vector field and the semantic connection field; based on the semantic connection field, the semantic field arrangement sentence is obtained through semantic relationship optimization; based on the token vector field, the predicted probability of the next token in the semantic field arrangement sentence is analyzed to determine the probability of the next token being generated.

[0069] Some embodiments of the present application provide corresponding Figure 1 A non-volatile computer storage medium constructed based on a dynamic rule base of semantic connection computing stores computer executable instructions, wherein the computer executable instructions are set as follows:

[0070] The manifold subspace of the initial concept manifold is constructed to obtain the concept manifold; based on the concept manifold, the continuous thought chain is determined through the differentiability of the concept neighborhood; according to the continuous thought chain, the language field is divided into the token vector field and the semantic connection field; based on the semantic connection field, the semantic field arrangement sentence is obtained through semantic relationship optimization; based on the token vector field, the predicted probability of the next token in the semantic field arrangement sentence is analyzed to determine the probability of the next token being generated.

[0071] The various embodiments in this application are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences from the other embodiments. In particular, the IoT device and media embodiments are generally similar to the method embodiments, so their description is relatively simple. For relevant portions, refer to the description of the method embodiments.

[0072] The system and medium provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the system and medium also have similar beneficial technical effects to their corresponding methods. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the system and medium will not be repeated here.

[0073] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0075] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0077] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0078] Memory may include non-permanent storage in a computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0079] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can be implemented using any method or technology to store information. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change RAM (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0080] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0081] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A method for constructing a dynamic rule base based on semantic connection computing, characterized in that: The method comprises: Constructing manifold subspaces on the initial concept manifold to obtain the concept manifold; Based on the concept manifold, the continuous thinking chain is determined through the differentiability of the concept neighborhood; According to the continuous thought chain, the language field is divided into the token vector field and the semantic connection field; According to the semantic connection field, a semantic field arrangement sentence is obtained by optimizing the semantic relationship; Based on the token vector field, performing a prediction probability analysis of the next token on the semantic field arrangement sentence to determine the generation probability of the next token; The initial concept manifold is constructed into a manifold subspace to obtain the concept manifold, which includes: Obtaining a local active coordinate frame of the initial concept manifold and selecting a sub-attribute of a concept as a unit vector of the local active coordinate frame to determine a local concept space; Mapping the local concept space semantics into N-dimensional space to obtain a concept manifold to be defined; According to a preset application scenario, manifold definition is performed on the concept manifold to be defined to obtain the concept manifold; Based on the concept manifold, the continuous thinking chain is determined through the differentiability of the concept neighborhood, which specifically includes: Determining a token chain based on the concept manifold, and screening the token chain for text coherence to obtain an incoherent token chain; Determining a disconnected region of the incoherent token chain, and calling a latent space based on the disconnected region; According to the latent space, the disconnected region is supplemented with differentiable cross-neighborhood latent space to determine the continuous thinking chain.

2. A method for constructing a dynamic rule base based on semantic connection computing according to claim 1, characterized in that: The N-dimensional space includes: N-dimensional Euclidean space and geometric space; The types of the concept manifold include: parameterized manifold, data manifold, and concept manifold.

3. The method for constructing a dynamic rule base based on semantic connection computing according to claim 1, characterized in that: According to the continuous thought chain, the language field is divided into the token vector field and the semantic connection field, which specifically include: Setting a word-element function, and constructing the semantic connection field through word-element semantic organization based on the continuous thought chain and the word-element function; According to the semantic connection field, the token vector field is constructed by analyzing the state vector of the token.

4. The method for constructing a dynamic rule base based on semantic connection computing according to claim 1, characterized in that: According to the semantic connection field, a semantic field arrangement sentence is obtained through semantic relationship optimization, specifically including: Obtaining word-units, and defining a geometric structure of a semantic manifold based on the word-units; Constructing a connection coefficient of a local coordinate system for the semantic manifold after the geometric structure is defined to determine a semantic connection feature; The semantic connection features are arranged and organized for the word units to obtain the semantic field arrangement sentences.

5. The method for constructing a dynamic rule base based on semantic connection computing according to claim 1, characterized in that: Based on the token vector field, a prediction probability analysis of the next token is performed on the semantic field arrangement sentence to determine the probability of the next token being generated, specifically including: Based on the token vector field, performing a prediction probability analysis of the next token on the semantic field arrangement sentence to obtain a generation state of the next token; A word-unit function and a word-unit complex function corresponding to the word-unit function are obtained, and a probability of generating the next token is determined through token collapse analysis based on the word-unit function and the word-unit complex function.

6. The method for constructing a dynamic rule base based on semantic connection computing according to claim 5, characterized in that: According to the word-unit function and the word-unit complex function, the probability of generating the next token is determined through token collapse analysis, specifically including: When the token prediction state is that the next token has not been generated, the token prediction state is uncertain; When the token prediction state is to generate the next token, the token prediction state is to collapse; When the token prediction state is collapsed, the field function corresponding to the semantic field arrangement sentence is determined based on the word function and the word complex function, and the square of the modulus of the field function of the next token and the field function corresponding to the semantic field arrangement sentence is set as the probability of generating the next token.

7. The method for constructing a dynamic rule base based on semantic connection computing according to claim 1, characterized in that: Based on the token vector field, performing a prediction probability analysis of the next token on the semantic field arrangement sentence to determine a generation probability of the next token, the method further comprising: Based on the token vector field and the semantic connection field, a semantic dynamic rule library is constructed.

8. A dynamic rule base construction device based on semantic connection computing, characterized in that: The device comprises: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: Constructing manifold subspaces on the initial concept manifold to obtain the concept manifold; Based on the concept manifold, the continuous thinking chain is determined through the differentiability of the concept neighborhood; According to the continuous thought chain, the language field is divided into the token vector field and the semantic connection field; According to the semantic connection field, a semantic field arrangement sentence is obtained by optimizing the semantic relationship; Based on the token vector field, performing a prediction probability analysis of the next token on the semantic field arrangement sentence to determine the generation probability of the next token; The initial concept manifold is constructed into a manifold subspace to obtain the concept manifold, which includes: Obtaining a local active coordinate frame of the initial concept manifold and selecting a sub-attribute of a concept as a unit vector of the local active coordinate frame to determine a local concept space; Mapping the local concept space semantics into N-dimensional space to obtain a concept manifold to be defined; According to a preset application scenario, manifold definition is performed on the concept manifold to be defined to obtain the concept manifold; Based on the concept manifold, the continuous thinking chain is determined through the differentiability of the concept neighborhood, which specifically includes: Determining a token chain based on the concept manifold, and screening the token chain for text coherence to obtain an incoherent token chain; Determining a disconnected region of the incoherent token chain, and calling a latent space based on the disconnected region; According to the latent space, the disconnected region is supplemented with differentiable cross-neighborhood latent space to determine the continuous thinking chain.

9. A non-volatile computer storage medium for a method for constructing a dynamic rule base based on semantic connection computing, storing computer-executable instructions, characterized in that: The computer executable instructions are configured to: Constructing manifold subspaces on the initial concept manifold to obtain the concept manifold; Based on the concept manifold, the continuous thinking chain is determined through the differentiability of the concept neighborhood; According to the continuous thought chain, the language field is divided into the token vector field and the semantic connection field; According to the semantic connection field, a semantic field arrangement sentence is obtained by optimizing the semantic relationship; Based on the token vector field, performing a prediction probability analysis of the next token on the semantic field arrangement sentence to determine the generation probability of the next token; The initial concept manifold is constructed into a manifold subspace to obtain the concept manifold, which includes: Obtaining a local active coordinate frame of the initial concept manifold and selecting a sub-attribute of a concept as a unit vector of the local active coordinate frame to determine a local concept space; Mapping the local concept space semantics into N-dimensional space to obtain a concept manifold to be defined; According to a preset application scenario, manifold definition is performed on the concept manifold to be defined to obtain the concept manifold; Based on the concept manifold, the continuous thinking chain is determined through the differentiability of the concept neighborhood, which specifically includes: Determining a token chain based on the concept manifold, and screening the token chain for text coherence to obtain an incoherent token chain; Determining a disconnected region of the incoherent token chain, and calling a latent space based on the disconnected region; According to the latent space, the disconnected region is supplemented with differentiable cross-neighborhood latent space to determine the continuous thinking chain.

Images