Large language model perceptual alignment method and device

Abstract

The present disclosure relates to a large language model perception alignment method and device. The method comprises: obtaining output text of a large language model for a preset perception task, and generating an explicit rule set of the large language model based on the output text; obtaining human-labeled perception samples and performing semantic clustering and abstract summary on the human-labeled perception samples to generate a human perception standard rule library; constructing a perception knowledge graph based on the explicit rule set and the human perception standard rule library, the perception knowledge graph storing triplets composed of an analysis subject, a perception argument and a discrimination conclusion; for input text of the large language model, retrieving human perception standard rules related to the input text and injecting the human perception standard rules into a reasoning process of the large language model, and retrieving triplets related to the input text in the graph as explicit node / edge evidence to guide the large language model to perform graph-like thinking chain reasoning. Thus, the stability, consistency and explainability of the large language model in complex text perception and classification tasks are significantly improved.

Description

Technical Field

[0001] This disclosure relates to technical fields such as text risk identification, sentiment analysis, and content category determination; more specifically, it relates to a method and apparatus for perceptual alignment of a large language model. Background Technology

[0002] Current large language models, after instruction fine-tuning and reinforcement learning, possess robust objective information processing and text generation capabilities, demonstrating strong versatility in a wide range of practical tasks. However, for higher-order perception tasks involving text risk identification, sentiment analysis, and content category determination, existing methods still primarily rely on external reward models or general human feedback, making it difficult to directly, efficiently, and stably extract and correct implicit biases within the model in complex semantic perception. Currently, there is a lack of a systematic and structured method to directly analyze the inherent perception patterns and discrimination logic within the model and calibrate their consistency with the specific perception standards of the target scenario. Especially when users or enterprises have specific content risk control requirements (e.g., rules for identifying inappropriate content, industry classification standards, corporate compliance manuals), relying solely on general instruction fine-tuning often fails to ensure stable model performance when handling complex tasks such as implicit risks and polysemous sentiment judgment. Summary of the Invention

[0003] To address the aforementioned issues, this disclosure proposes a large language model-aware alignment method and apparatus, a computing system, and a computer-readable storage medium.

[0004] According to one aspect of this disclosure, a large language model perception alignment method is provided, comprising: acquiring the output text of a large language model for a preset perception task, and generating an explicit rule set of the large language model based on the output text through semantic vectorization, cluster analysis, and rule inversion; acquiring human-annotated perception samples, and performing semantic clustering and abstraction on the human-annotated perception samples to generate a human perception standard rule library; constructing a perception knowledge graph based on the explicit rule set and the human perception standard rule library, wherein the perception knowledge graph stores triples consisting of an analysis subject, perception arguments, and discrimination conclusions; for the input text of the large language model, retrieving human perception standard rules related to the input text from the human perception standard rule library and injecting them into the reasoning process of the large language model, and retrieving triples related to the input text from the perception knowledge graph as explicit node / edge evidence to guide the large language model to perform graph-like thought chain reasoning, so that the perception discrimination result of the large language model is aligned with the preset human standard.

[0005] Optionally, the step of generating the explicit rule set of the large language model based on the output text through semantic vectorization, cluster analysis, and rule inversion includes: converting the output text into semantic vectors to form an initial vector set; using an auxiliary semantic model to semantically enhance the vectors in the initial vector set to obtain an enhanced feature vector set; determining the optimal number of clusters for the enhanced feature vector set based on a multi-index joint decision-making strategy, and performing clustering to obtain multiple clusters; for each of the multiple clusters, selecting multiple representative samples close to the cluster center, and using a rule-summarizing large language model to summarize the multiple representative samples to generate explicit rules; and generating the explicit rule set based on the multiple explicit rules corresponding to the multiple clusters.

[0006] Optionally, the step of determining the optimal number of clusters for the enhanced feature vector set based on a multi-index joint decision-making strategy includes: determining multiple candidate cluster numbers through multiple trials, calculating the silhouette coefficient, Davies-Bouldin index, and Calinski-Harabasz index for each candidate cluster number, and determining the optimal number of clusters by combining the tree structure changes of hierarchical clustering and the performance of repeated experiments on cluster stability through a multi-index comprehensive score function.

[0007] Optionally, the step of retrieving human perception standard rules related to the input text from the human perception standard rule base and injecting them into the reasoning process of the large language model includes: vectorizing the input text and the rules in the human perception standard rule base using a multi-granularity embedding model to construct a rule vector index; retrieving multiple human perception rules most relevant to the input text from the rule vector index based on vector similarity; and inputting the retrieved multiple human perception rules as explicit prompts, along with the input text, into the large language model to drive the large language model to perform reasoning and judgment.

[0008] Optionally, the steps of constructing a perceptual knowledge graph based on the explicit rule set and the human perception standard rule base include: extracting triples comprising an analysis subject, perceptual arguments, and a judgment conclusion from the explicit rule set, the human perception standard rule base, and the thought chain text of the large language model; using the analysis subject whose frequency of occurrence is greater than a preset threshold as the core node, aggregating the related perceptual arguments and judgment conclusions to form triples, and clustering the formed triple set according to the analysis subject to form a preliminary graph edge set; converting the preliminary graph edge set into a graph structure grammar and storing it in a graph database to construct the perceptual knowledge graph.

[0009] Optionally, the step of retrieving relevant triples from the perceptual knowledge graph as explicit node / side evidence to guide the large language model in graph-based thought chain reasoning includes: converting the triples in the perceptual knowledge graph into natural language description text and vectorizing them to construct a triple vector index; retrieving triple text related to the current context from the triple vector index and injecting it as explicit node / side evidence into the reasoning process of the large language model to guide the large language model to generate a graph-based thought chain containing the explicit node / side evidence.

[0010] Optionally, the large language model perceptual alignment method further includes: when multiple related triple texts are retrieved, the large language model reverse maps the multiple related triple texts back to the original graph elements, and performs logical completion and / or conflict detection based on the association path of the original graph elements in the perceptual knowledge graph, so as to further constrain the graph-like thought chain reasoning of the large language model.

[0011] According to another aspect of this disclosure, a large language model perception alignment device is provided, comprising: an explicit rule set generation unit configured to acquire output text of a large language model for a preset perception task, and generate an explicit rule set of the large language model based on the output text through semantic vectorization, cluster analysis, and rule inversion; a human perception standard rule base generation unit configured to acquire human-annotated perception samples, and perform semantic clustering and abstraction summarization on the human-annotated perception samples to generate a human perception standard rule base; and a perception knowledge graph construction unit configured to construct a knowledge graph based on the explicit rule set. A perceptual knowledge graph is constructed using the aforementioned human perception standard rule base. The perceptual knowledge graph stores triples consisting of an analysis subject, perceptual arguments, and a judgment conclusion. The reasoning unit is configured to, for the input text of the large language model, retrieve human perception standard rules related to the input text from the human perception standard rule base and inject them into the reasoning process of the large language model. It also retrieves triples related to the input text from the perceptual knowledge graph as explicit node / edge evidence to guide the large language model to perform graph-like thought chain reasoning, so that the perceptual judgment result of the large language model is aligned with the preset human standard.

[0012] Optionally, the explicit rule set generation unit is further configured to: convert the output text into semantic vectors to form an initial vector set; use an auxiliary semantic model to semantically enhance the vectors in the initial vector set to obtain an enhanced feature vector set; determine the optimal number of clusters for the enhanced feature vector set based on a multi-index joint decision-making strategy, and perform clustering to obtain multiple clusters; for each of the multiple clusters, select multiple representative samples close to the cluster center, and use a rule summarization large language model to summarize the multiple representative samples to generate explicit rules; and generate the explicit rule set based on the multiple explicit rules corresponding to the multiple clusters.

[0013] Optionally, the explicit rule set generation unit is further configured to: determine multiple candidate cluster numbers through multiple trials, calculate the silhouette coefficient, Davies-Bouldin index, and Calinski-Harabasz index of each candidate cluster number, and determine the optimal cluster number by combining the tree structure changes of hierarchical clustering and the performance of repeated cluster stability experiments through a multi-index comprehensive score function.

[0014] Optionally, the reasoning unit is further configured to: vectorize the input text and the rules in the human perception standard rule base using a multi-granularity embedding model to construct a rule vector index; retrieve multiple human perception rules most relevant to the input text from the rule vector index based on vector similarity; and input the retrieved multiple human perception rules as explicit prompts, along with the input text, into the large language model to drive the large language model to perform reasoning and judgment.

[0015] Optionally, the perceptual knowledge graph construction unit is further configured to: extract triples including the analysis subject, perceptual arguments, and judgment conclusions from the explicit rule set, the human perception standard rule base, and the thought chain text of the large language model; take the analysis subject whose frequency is greater than a preset threshold as the core node, aggregate the related perceptual arguments and judgment conclusions to form triples, and cluster the formed triple set according to the analysis subject to form a preliminary graph edge set; convert the preliminary graph edge set into graph structure grammar and store it in the graph database to construct the perceptual knowledge graph.

[0016] Optionally, the reasoning unit is further configured to: convert the triples in the perceptual knowledge graph into natural language descriptive text and vectorize them to construct a triple vector index; retrieve the triple text related to the current context from the triple vector index and inject it as explicit node / edge evidence into the reasoning process of the large language model to guide the large language model to generate a graph-like thought chain containing the explicit node / edge evidence.

[0017] Optionally, the reasoning unit is further configured to: when multiple related triplet texts are retrieved, the large language model reverse-maps the multiple related triplet texts back to the original graph elements, and performs logical completion and / or conflict detection based on the association path of the original graph elements in the perceptual knowledge graph, so as to further constrain the graph-like thought chain reasoning of the large language model.

[0018] According to another aspect of this disclosure, a computing system is provided that includes at least one computing device and at least one storage device for storing instructions, wherein the instructions, when executed by the at least one computing device, cause the at least one computing device to perform the large language model-aware alignment method as described above.

[0019] According to another aspect of this disclosure, a computer-readable storage medium is provided for storing instructions, wherein when the instructions are executed by at least one computing device, the at least one computing device causes the at least one computing device to perform the large language model-aware alignment method as described above.

[0020] By adopting this disclosure, the stability, consistency and interpretability of large language models in complex text perception and classification tasks (e.g., malicious content identification and sentiment analysis) are significantly improved, overcoming the technical bottlenecks of traditional fine-tuning methods such as weak ability to align perception standards, difficulty in rule transfer, and opaque inference chains; and achieving controllable, interpretable and stable perception and risk control alignment effects without sacrificing generation capabilities. Attached Figure Description

[0021] The above and / or other objects and advantages of this disclosure will become clearer from the following description of embodiments in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a large language model-aware alignment method according to an exemplary embodiment of the present disclosure; Figure 2 This is a schematic diagram illustrating a large language model-aware alignment method according to an embodiment of the present disclosure; Figure 3 A block diagram of a large language model-aware alignment apparatus according to an exemplary embodiment of the present disclosure is shown; Figure 4This is a block diagram illustrating a computing system including at least one computing device and at least one storage device of storage instructions according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0022] The following description, in conjunction with the accompanying drawings, provides specific embodiments to aid the reader in gaining a comprehensive understanding of the methods, apparatus, and / or systems described herein. However, upon understanding this disclosure, various changes, modifications, and equivalents of the methods, apparatus, and / or systems described herein will become apparent. For example, the order of operations described herein is merely illustrative and is not limited to those orders set forth herein, but may be altered as will become clear upon understanding this disclosure, except for operations that must occur in a specific order. Furthermore, for clarity and conciseness, descriptions of features known in the art may be omitted.

[0023] For high-order perception tasks involving text risk identification, sentiment analysis, and content category determination, existing methods still mainly rely on external reward models or general human feedback, making it difficult to directly, efficiently, and stably extract and correct implicit biases in complex semantic perception within the model. Therefore, a new technical framework is needed to: efficiently extract the model's own perception logic of text features; directly enhance the model's decision-making ability using a human-defined perception criterion library (e.g., review rules, classification definitions); and optimize the knowledge base to achieve a more robust, stable, and controllable text perception and correction capability. Compared with traditional large language model alignment techniques, the large language model perception alignment method disclosed in this paper combines the three dimensions of "perceptual logic extraction—discrimination criterion enhancement—reasoning structuring" for the first time. It mines the model's inherent content recognition rules through clustering, performs explicit knowledge enhancement using a human perception criterion library, and uses Graph-CoT (Graphic Thinking Chain) to transform the classification decision process into a verifiable graph-structured reasoning chain, achieving a shift from "implicit semantic alignment" to "explicit, controllable, and adjustable perception modeling." The large language model-based perceptual alignment method disclosed herein not only reveals the model's internal patterns of perceiving text risk and sentiment, but also provides an engineering framework that allows for the customization of perceptual standards for specific users and industry scenarios, filling the technical gaps in existing methods for perceiving and controlling high-transparency, high-consistency content. Specifically, the large language model-based perceptual alignment method disclosed herein can be widely applied to the following fields and solve corresponding technical problems.

[0024] When processing online comments, bullet screens, or user-generated content, traditional risk control models often rely on keyword matching or general black-box classifiers, making it difficult to identify high-order semantic risks such as "implicit attacks," "sarcasm," or "homophonic variations." For example, some metaphorical malicious verbal attacks or slang from specific circles do not contain obvious prohibited words, leading to missed detections by traditional models. The perceptual alignment method of this disclosure utilizes a constructed human perception standard rule base (e.g., rules containing the latest online slang variations and attack pattern definitions) and a perceptual knowledge graph to guide the large language model in graph-based reasoning. The model no longer simply looks at surface-level words but links "implicit expressions" to the core nodes of "aggressive intent" through paths in the graph, thereby accurately identifying deep semantic risks, significantly reducing the missed detection rate, and providing logical justification for the judgment (e.g., "The text detected the use of metaphor X; according to rule base R, this metaphor is judged as hate speech in context Y").

[0025] In financial report review, contract risk assessment, and legal document drafting assistance, large language models often suffer from "illusion" problems or a lack of domain rigor, leading to biased or inconsistent interpretations of clauses. Fine-tuning general instructions is insufficient to ensure that the model consistently adheres strictly to specific industry compliance manuals or legal provisions. By transforming industry-specific compliance manuals into explicit rule sets and injecting them into the reasoning process, the large language model perception alignment method disclosed herein forces large language models to have a legal basis when generating judgments. Graph-CoT ensures that when determining the compliance of a clause, the model must find the corresponding legal basis node in the knowledge graph as explicit evidence. This not only solves the "illusion" problem of models arbitrarily interpreting clauses but also provides reviewers with a traceable compliance proof path, enhancing the professionalism and interpretability of the review.

[0026] In psychological counseling dialogue analysis or consumer sentiment surveys, general models often only recognize coarse-grained emotions like "positive / negative," struggling to distinguish between "extreme despair," "anxiety," and "ordinary sadness." They are also susceptible to biases in training data, leading to inappropriate feedback. By extracting psychological professional standards or refined emotion classification systems as a rule base for human perception standards, the perceptual alignment method of this disclosure can calibrate the model's perception of emotional intensity. During inference, the model retrieves triples related to "emotional trigger—expression mode—emotional category" from the graph, accurately mapping the user's implicit expressions to specific psychological state categories. This ensures the model possesses empathic ability and judgment accuracy aligned with human expert standards when handling sensitive topics.

[0027] Figure 1 This is a flowchart illustrating a large language model-aware alignment method according to an exemplary embodiment of the present disclosure. Figure 2This is a schematic diagram illustrating a large language model-aware alignment method according to an embodiment of the present disclosure.

[0028] like Figure 1 As shown, in step S101, the output text of the large language model for a preset perception task is obtained, and an explicit rule set of the large language model is generated based on the output text through semantic vectorization, cluster analysis, and rule inversion. This disclosure proposes a cluster analysis method for systematically extracting the implicit perception logic and discrimination rules of a large language model. Its core idea is to obtain a set of explicit, interpretable, and criterion-alignable perception feature mapping rules by performing semantic modeling, distribution aggregation, and rule inversion on the generated content of the large language model in generalized perception tasks (e.g., text classification, sentiment analysis, risk identification).

[0029] In the example, the output text is converted into semantic vectors to form an initial vector set. An auxiliary semantic model is then used to semantically augment the vectors in the initial vector set, resulting in an augmented feature vector set. (The auxiliary semantic model not only performs synonym expansion and semantic disambiguation but also supplements implicit risk dimensions based on context. After this stage, the separability of the feature space is significantly enhanced, laying a more stable mathematical foundation for subsequent clustering operations.) Based on a multi-index joint decision-making strategy, the optimal number of clusters for the augmented feature vector set is determined (for example, multiple candidate cluster numbers can be determined through multiple trials, and the silhouette coefficient, Davies-Bouldin index, and Calinski-Harab index can be calculated for each candidate cluster number). The ASZ metric, combined with the tree-like structure changes of hierarchical clustering and the performance of repeated clustering stability experiments, determines the optimal number of clusters through a multi-metric comprehensive scoring function. This avoids the biases caused by traditional single-metric methods, making the semantic and perceptual category differences between samples more clearly expressed in the vector space. Clustering is then performed to obtain multiple clusters. For each cluster, several representative samples close to the cluster center are selected (compared to random sampling, the representative sample extraction method based on center distance significantly improves the accuracy and stability of the model's value inversion). A rule-summarizing large language model is then used to summarize these representative samples to generate explicit rules. Based on the multiple explicit rules corresponding to the multiple clusters, an explicit rule set is generated. This enables the transformation from "implicit value space" to "explicit rule system," which not only allows for in-depth analysis of the perception and discrimination mechanisms of large language models on complex texts, but also provides a directly usable structured rule foundation for subsequent perception standard enhancement, Graph-CoT inference structuring, and perception correction models. This makes the entire perception logic alignment process interpretable, verifiable, and engineerable.

[0030] For example, in such Figure 2 The clustering-based rule extraction method for large language models disclosed herein is based on the core idea of ​​a clustering analysis method for systematically extracting the implicit perceptual logic and discrimination rules of large language models. This method involves semantic modeling, distribution aggregation, and rule inversion of the generated content of large language models in generalized perceptual tasks (e.g., text classification, sentiment analysis, risk identification, etc.) to obtain a set of explicit, interpretable, and criterion-alignable perceptual feature mapping rules.

[0031] Specifically, the semantic vector space is first constructed from the model's extensive reasoning processes and final responses. For each piece of text output by the large language model, its vectorized representation is denoted as... All samples constitute a vector set In the initial stage, the text can be preprocessed, including word segmentation, cleaning, and stop word removal, and the term frequency-inverse text frequency (TF-IDF) or sentence vector embedding can be calculated to map the text semantics to a high-dimensional vector space. Based on this, by statistically analyzing keyword frequencies, mining co-occurrence relationships, and analyzing the topic structure of the corpus, a perceptual boundary for the model's understanding of specific text types (e.g., malicious verbal attacks, indecent information, or sadness) can be constructed. This distribution serves as a primary clustering space, providing a coarse-grained structural framework for subsequent semantic aggregation.

[0032] Since relying solely on statistical methods is insufficient to fully characterize the deep semantic features of large language models in complex content determination, a larger-scale auxiliary model with enhanced semantic understanding capabilities is further introduced. This involves enhancing the perceptual accuracy of the initially constructed feature space. The basic approach is to process each vector... Input this auxiliary model to obtain a high-quality vector representation with semantic expansion. This constitutes a new set of features. The auxiliary model can not only perform synonym expansion and semantic disambiguation, but also supplement implicit risk dimensions based on context. After this stage of processing, the separability of the feature space is significantly enhanced, laying a more stable mathematical foundation for subsequent clustering operations.

[0033] To determine the truly reasonable number of clusters in the value rule space A multi-indicator joint decision-making strategy was adopted for cluster number selection. In multiple experiments, different... The values ​​are clustered, and the silhouette coefficient of each cluster result is calculated. Davies-Bouldin index Calinski-Harabasz index Evaluation indicators, etc.

[0034] Simultaneously, combining the tree-like structure changes of hierarchical clustering and the performance of repeated clustering stability experiments (specifically, this refers to double-checking the number of clusters K through structural intuition and statistical robustness; the tree-like structure changes of hierarchical clustering refer to observing the "distance jumps" in the clustering tree diagram. If the required semantic distance suddenly increases significantly when merging two clusters, it indicates that the two clusters are fundamentally very different (e.g., "generally negative" and "extreme hatred"), and this is the natural semantic boundary point, which should not be forcibly merged. The performance of repeated clustering stability experiments refers to conducting multiple random initialization or resampling experiments after selecting a candidate K value. If the sample classification results under the same K value (e.g., K=5) remain highly consistent in multiple experiments (i.e., samples are always assigned to the same group), it indicates that the classification structure is stable; conversely, if the results fluctuate drastically, it indicates that the K value is unreliable), the optimal number of clusters is finally selected through the following optimization objective: .

[0035] in, This is a multi-indicator comprehensive scoring function. Therefore, it avoids the biases introduced by traditional single-indicator methods, making the semantic and perceptual category differences between samples more clearly expressed in the vector space. After this stage of processing, the separability of the feature space is significantly enhanced, laying a more stable mathematical foundation for subsequent clustering operations.

[0036] After obtaining the final clustering structure, calculate the center vector within each cluster. And calculate the distance from the center for all points within the cluster. Based on the distance ranking, the most representative samples closest to the center are selected to form a representative set. .

[0037] These samples are considered typical cases that best reflect the value semantics of the cluster. Compared with random sampling, the representative sample extraction method based on center distance can significantly improve the accuracy and stability of the model's value inversion.

[0038] To further transform these representative samples into "value judgment rules" that are interpretable, summable, and usable for criterion alignment, another large-scale model with stronger reasoning capabilities is introduced. As a rule summarizer, it summarizes each representative sample set Input the model to obtain the corresponding value rule entries as follows: These rules typically include the preconditions for value judgments, preference tendencies, rationale, conflict resolution methods, and typical inference chain patterns (CoT Patterns). By performing rule inversion on representative samples from all clusters, the value decision rule set of the tested model is finally obtained. .

[0039] This set of rules systematically reveals the inherent judgment methods of the model across different value dimensions, realizing the transformation from an "implicit value space" to an "explicit rule system." Overall, this method not only provides in-depth analysis of the perception and discrimination mechanisms of large language models on complex texts, but also provides a directly usable structured rule foundation for subsequent perception standard enhancement, Graph-CoT inference structuring, and perception correction models, making the entire perception logic alignment process interpretable, verifiable, and engineerable.

[0040] In step S102, human-annotated perceptual samples are obtained, and semantic clustering and abstraction are performed on these samples to generate a human-perceptual standard rule base. In step S103, a perceptual knowledge graph is constructed based on the explicit rule set and the human-perceptual standard rule base. The perceptual knowledge graph stores triples consisting of the analysis subject, perceptual arguments, and judgment conclusions. In the example, triples including the analysis subject, perceptual arguments, and judgment conclusions are extracted from the explicit rule set, the human-perceptual standard rule base, and the thought chain text of the large language model. Analysis subjects with a frequency greater than a preset threshold are used as core nodes, and their related perceptual arguments and judgment conclusions are aggregated to form triples. The resulting triple set is then clustered according to the analysis subject to form a preliminary graph edge set. The preliminary graph edge set is converted into graph structure grammar and stored in a graph database to construct the perceptual knowledge graph.

[0041] In step S104, for the input text of the large language model, human perception standard rules related to the input text are retrieved from the human perception standard rule base and injected into the reasoning process of the large language model. Furthermore, triples related to the input text are retrieved from the perception knowledge graph as explicit node / edge evidence to guide the large language model in graph-based thought chain reasoning, so that the perception judgment result of the large language model aligns with the preset human standards. In the example, the input text and the rules in the human perception standard rule base can be vectorized using a multi-granularity embedding model to construct a rule vector index. Based on vector similarity, the most relevant human perception rules to the input text are retrieved from the rule vector index. These retrieved human perception rules are used as explicit prompts and input into the large language model along with the input text to drive the large language model in reasoning and judgment. Furthermore, in the example, triples in the perceptual knowledge graph can be converted into natural language descriptive text and vectorized to construct a triple vector index. Triple text relevant to the current context is retrieved from this index and injected as explicit node / edge evidence into the reasoning process of the large language model, guiding it to generate a graph-like thought chain containing explicit node / edge evidence. Additionally, when multiple relevant triple texts are retrieved, the large language model maps these texts back to the original graph elements and performs logical completion and / or conflict detection based on the association paths of the original graph elements in the perceptual knowledge graph, further constraining the graph-like thought chain reasoning of the large language model.

[0042] For example, after completing the clustering inversion of the internal perception logic of the large language model, a knowledge enhancement system based on a human perception standard library was further constructed. This system explicitly and controllably injects real human classification standards and risk control norms into the model's reasoning process, addressing deep-seated problems in complex semantic perception scenarios, such as "unstable discrimination standards," "fluctuations in perception results due to contextual changes," and "poor classification consistency across different scenarios." For instance, 5000 real human-annotated samples were first collected, covering multi-dimensional perception tasks such as malicious content identification, text type classification, sentiment analysis, risk level assessment, and compliance review. Each data point consists of an event description and a corresponding human perception category label, forming a set. To construct a representative distribution of human perception experience, semantic clustering methods were used to divide these 5000 data points into seven clusters with clear semantic consistency. Each cluster reflects a stable perception dimension, such as physical conflict, malicious verbal attacks, indecent information, negative emotions, credit risk, sensitive topics, and compliant content.

[0043] To transform the raw labeled data into "HumanPerception Rules" that can be directly invoked by the model, several representative samples can be selected within each cluster, and a larger, more inductively powerful language model can be used to perform a few-shot rule summarization. For clustering... Construct a representative subset It also establishes a unified rule format: "an abstract description of a specific event + the type of human value judgment about that event," allowing the teacher model to output generalizable rules. By repeating this process for all clusters, a criterion library containing, for example, 52 high-quality rules was eventually constructed. These rules condense the core patterns of value judgments from 5,000 human annotations, providing an organized and logical explicit knowledge base for subsequent reasoning enhancement.

[0044] After constructing the human rule base, a complete RAG (Retrieval-Augmented Generation) enhancement module can be designed to introduce human value norms into the reasoning process, enabling the generation process to be based on "explicit rules" rather than entirely relying on implicit preferences in the model parameters. BGE-M3 (BAAI GeneralEmbedding–Multi-Function Multi-Granularity) is used as the vectorization model to encode input events and rule entries into multi-granularity, multi-semantic level embedding vectors. Specifically, given user input... First, the vector representation is obtained through BGE-M3. Simultaneously, the same vectorization process can be applied to the 52 value rules to construct a rule vector library. To support efficient retrieval, all rule vectors are stored in a vector index structure based on FAISS / Elastic Search, enabling subsequent retrieval to be completed in sublinear time.

[0045] During the reasoning phase, it can be based on the input The semantic features are used to perform vector retrieval, selecting the top vectors that best match the input semantics from the rule base. According to the rules, we can obtain , in, Cosine similarity or BGE-M3-adapted multi-vector matching strategies are typically employed. Because BGE-M3 supports multi-granularity semantic representations across short texts, long texts, and keywords, its retrieval results are more robust, accurately matching the most suitable human value rules even when faced with different contextual structures, different event expressions, and even implicit user intent. This characteristic greatly enhances the stability of the RAG enhancement module, enabling the system to provide consistent criteria in value judgment tasks.

[0046] After retrieving the rule set Then, the system injects these entries as explicit prompts into the reasoning process of the large language model, so that model generation no longer depends solely on the parameter space, but is driven by both human rules and the model itself. The reasoning process is formalized as follows:

[0047] in, For the final output, This is an augmented large language model. In fact, this augmentation method significantly changes the model's thought chain structure: the model not only uses rules as the basis for value reasoning, but also performs chain reasoning of "rule interpretation - context mapping - value judgment" in generation, thereby improving decision transparency.

[0048] Furthermore, constraint mechanisms for rule referencing can be designed. When the reasoning chain generated by the model deviates from human criteria, structured prompts can force the model to reconsider the most relevant rules, thereby achieving a weak form of forward checking and ensuring that the final output follows a value system consistent with humans. By combining the high-precision retrieval of BGE-M3 with rule-enhanced reasoning, a truly "controllable value reasoning system" is constructed, capable of maintaining stable, consistent, and transparent value orientations across various subjective tasks. Overall, this RAG enhancement method based on BGE-M3 vector retrieval and a human value rule base not only provides explicit value constraints for large language models but also endows the generation process with contextual relevance, criterion consistency, and reasoning interpretability, providing a solid knowledge foundation and reasoning framework for subsequent Graph-CoT optimization.

[0049] In addition, such as Figure 2The structured perceptual reasoning and graph construction steps of the large language model shown aim to combine human perceptual standards (e.g., review rules, risk control specifications) with structured feature evidence extracted from the model and corpus into a queryable and reasonable perceptual knowledge graph, thereby supporting accurate classification reasoning in the form of Graph-CoT (graph-cotton). To achieve this goal, the regularized risk control text and the analytical thought chain generated by the model are first subjected to "Subject-Argument-Conclusion" (SAC) triple extraction. Specifically, this is achieved by directionally transferring basic semantic analysis techniques from natural language processing to the risk control reasoning scenario, thereby mapping unstructured text to a specific SAC logical paradigm. For each reasoning chain or rule sample... The set of triples is obtained through preprocessing, syntactic analysis, and semantic role labeling: , in, Indicates the subject of the discussion (subject entity or event). The arguments (evidence, reasons, facts, or assumptions) that support the conclusion. This represents the conclusion or value judgment drawn based on the arguments. To ensure the consistency of the extraction results, techniques such as named entity recognition, coreference resolution, and syntactic dependency parsing were applied in the extraction process. The extracted text was also normalized (normalized entity names, unified time and metric expressions, and removed redundant modifiers) to ensure that the nodes in the subsequent graph construction are highly comparable.

[0050] After obtaining a large number of SAC triples, the frequency of occurrence of the "analysis object topic" (i.e., subject or core entity) in all triples is statistically analyzed, and based on this, the top 2000 analysis objects with the highest frequency are selected as the candidate set of core nodes for the knowledge graph. Formally, let... Define a frequency function for all subject sets. Then select set The selection of the top 2000 analysis objects was not a simple arbitrary threshold setting, but rather an engineering trade-off derived from balancing coverage and retrieval efficiency through multiple experiments. This scale covers common value judgment scenarios while maintaining controllable latency for graph queries and vector retrieval. For each core node, its related arguments and conclusions are aggregated, and the resulting triplet sets are clustered by subject to form a preliminary graph edge set. .

[0051] Next, the SAC triples are converted into a graph-structured grammar and implemented using a graph database as the storage engine. The knowledge graph is formally constructed as a directed graph. , where the set of nodes Includes core subject nodes, argument nodes, and conclusion nodes (conclusion nodes can also be considered as special attribute nodes), and edge set. This represents inference relationships such as "subject → argument," "argument → conclusion," or "subject → conclusion." To support complex logical relationships, metadata (e.g., confidence level, source sample ID, timestamp, whether the rule source is a human criterion or model induction) is appended to the edges, allowing graph queries to be weighted by confidence level or filtered by source. This metadata becomes particularly important for subsequent path constraints and conflict detection as the amount of graph data and the complexity of relationships increase.

[0052] The Neo4j interface provided by LangChain can be used to write text and structured triples into the Neo4j graph database, leveraging its mature graph query language (Cypher) to achieve complex path finding and subgraph matching. The process of transforming the extracted results into Node / Relationship form and writing them into Neo4j may encounter several challenges, such as transaction rollback during large-scale concurrent writes, schema drift during complex text-to-graph schema mapping, and entity alignment challenges during multi-source heterogeneous evidence fusion. To address these engineering challenges, batch write queues, idempotent write strategies, rule-based entity standardization pipelines, and incremental index updates are employed. Before writing, a "schema validator" performs syntax verification and conflict detection on the subgraph structure to be written, reducing the failure rate and data inconsistency probability of the graph database in a production environment. To ensure the graph knowledge base can be efficiently utilized by downstream RAG / Graph-CoT inference modules, the edges of the graph (each...) are... The triples are transformed into plain text descriptions of natural language and a textual corpus is constructed. The transformation is then formally converted into mapping functions. : For example, the generated text format is "Subject: Argument: Conclusion: ", or the more compact natural sentence structure "targeting ,based on The conclusion is The purpose of this textualization is to make structured graph knowledge compatible with existing text-based retrieval-enhancement pipelines, enabling RAG retrieval, vectorization, and LLM inference to interact within the same semantic space.

[0053] After further standardization (e.g., denoising, deduplication, and templated processing) of the textualized triplet corpus, the text is mapped to a vector representation using the aforementioned embedding model, and a FAISS vector index is constructed: for each text... Calculate its vector representation The vector is then stored in a vector library along with its corresponding primitive information (e.g., node ID, edge type, confidence level). During a query, for the input... The schematic hints or reasoning segments that need to be verified are first textualized and embedded as... Then, in FAISS, a nearest neighbor search is performed to obtain the most relevant set of triplet texts. The retrieved triplet text can be back-mapped back to the original graph elements, and the generation of the model can be further constrained by graph path completion or conflict detection; at the same time, it can also be directly used as explicit node / edge evidence in Graph-CoT to guide the model to reason on the graph structure, thereby constructing a graph-like thought chain rather than a linear CoT.

[0054] To ensure the long-term availability and timeliness of the perceptual knowledge base, a closed-loop update mechanism can be designed: when the model discovers new effective perceptual feature triples (e.g., novel variant sensitive words, implicit risk expression patterns) during online inference or manual review, pattern verification is first performed in the pre-write buffer, followed by aggregation updates (e.g., merging synonymous feature nodes, weighting risk correlation edges), and triggering incremental reconstruction of the vector library or insertion of new vectors to maintain consistency between the FAISS index and the graph database. In terms of engineering implementation, to address potential synchronization delays between Neo4j and FAISS, an event-driven synchronizer is implemented. Any update to the graph structure triggers the corresponding feature textualization and vector insertion tasks, and idempotent keys are used to prevent duplicate writes. Overall, by structurally extracting the "analysis subject—perceived evidence—judgment conclusion" chain, constructing core feature nodes based on frequency filtering, implementing the graph database using LangChain / Neo4j, achieving reversible conversion from triples to plain text, and tightly integrating FAISS vectorized retrieval, a Graph-CoT perceptual knowledge base has been formed that supports both graph logic verification and text retrieval. This system provides the model with a structured chain of classification and judgment evidence, and also ensures efficient retrieval and incremental updates in engineering, thereby significantly improving the accuracy, interpretability, and maintainability of complex semantic perception and risk control reasoning based on the perceptual standard library.

[0055] By adopting the large language model perception alignment method according to exemplary embodiments of the present disclosure, the stability, consistency and interpretability of the large language model in complex text perception and classification tasks (e.g., malicious content identification, sentiment analysis) are significantly improved. It overcomes the technical bottlenecks of traditional fine-tuning methods, such as weak perception standard alignment ability, difficulty in rule transfer, and opaque inference chain. It can achieve controllable, interpretable and stable perception and risk control alignment effect without sacrificing generation ability.

[0056] In practical testing of the large language model-aware alignment method according to exemplary embodiments of this disclosure, a baseline model was first constructed that does not rely on external rule bases and knowledge graphs. The model was then directly trained on labeled data for complex text classification and risk identification using only traditional fine-tuning methods. While this baseline model can provide reasonable answers in some standardized scenarios with obvious features, it is relatively unstable in risk scenarios with obscure semantics, polysemous conflicts, and ambiguous boundaries (e.g., sarcastic remarks, metaphorical attacks). Its classification consistency is only about 78%–82%, and the reasoning process often lacks feature-based citations or explanatory chains.

[0057] To enhance the model's stability and controllability, a RAG retrieval mechanism from the human perception standard library was gradually introduced, enabling the model to explicitly reference 52 perception discrimination rules summarized from 5000 annotations during the inference phase. After enabling RAG, the model's perception consistency improved to approximately 89%, with most errors concentrated in deep semantic dependencies or long text context scenarios. Further integration with a Neo4j-based Graph-CoT perception knowledge graph allowed the model to utilize structured feature triples for inference completion, significantly improving semantic ambiguity resolution and risk evidence interpretation, raising overall consistency to approximately 94%. Furthermore, the interpretability and auditability of the inference chain were greatly enhanced, demonstrating the substantial performance advantages of the proposed perception standard enhancement and graph-structured inference method.

[0058] The three-stage system of "perceptual logic extraction—RAG enhancement—Graph-CoT graph structure reasoning" proposed in this publication significantly improves the stability, consistency, and interpretability of large language models in complex text perception and classification tasks (e.g., malicious content identification, sentiment analysis), overcoming the technical bottlenecks of traditional fine-tuning methods such as weak alignment of perception criteria, difficulty in rule transfer, and opaque inference chains. In the example, firstly, by clustering, semantically abstracting, and summarizing 5000 human-annotated data points, 52 high-quality explicit rules covering common content risk control and sentiment perception dimensions are constructed. This allows the model's reasoning process to move away from relying entirely on parameter memorization and possess a discriminative basis that can be reviewed and invoked across scenarios. Secondly, by introducing a BGE-M3-based RAG retrieval mechanism, the model can automatically match the most relevant perception criteria items in each inference, making the judgment results more stable in different contexts and significantly reducing the inconsistency in classification output caused by semantic drift and differences in expression. In actual testing, the model's perception discrimination consistency improved from approximately 80% at the baseline to 89%. More importantly, by constructing a Neo4j-based Graph-CoT perceptual knowledge graph, a large number of "analysis subject—perceptual evidence—judgment conclusion" triples are embedded into the structured reasoning network, enabling the model to achieve chain-based reasoning completion in scenarios with complex logical relationships, polysemous conflicts, and deep semantic dependencies. Graph enhancement allows the model to explicitly reference perceptual evidence paths, improving the transparency of judgment criteria and effectively reducing misjudgment bias caused by broken reasoning chains. After integrating RAG and Graph-CoT, the overall consistency of the model is improved to approximately 94%, the inference interpretability is significantly enhanced, and it exhibits stronger robustness in cross-task and cross-scenario transfer. Thus, a controllable, interpretable, and stable alignment effect between perception and risk control can be achieved without sacrificing generation capabilities.

[0059] Figure 3 A block diagram of a large language model-aware alignment device according to an exemplary embodiment of the present disclosure is shown.

[0060] like Figure 3As shown, the large language model perception alignment apparatus 300 according to an exemplary embodiment of the present disclosure includes: an explicit rule set generation unit 301, configured to acquire the output text of the large language model for a preset perception task, and generate an explicit rule set of the large language model based on the output text through semantic vectorization, cluster analysis, and rule inversion; a human perception standard rule base generation unit 302, configured to acquire human-annotated perception samples, and perform semantic clustering and abstraction summarization on the human-annotated perception samples to generate a human perception standard rule base; and a perception knowledge graph construction unit 303, configured to... The system constructs a perceptual knowledge graph by combining a set of rules and a human perception standard rule base. The perceptual knowledge graph stores triples consisting of the analysis subject, perceptual arguments, and judgment conclusions. The reasoning unit 304 is configured to retrieve human perception standard rules related to the input text from the human perception standard rule base and inject them into the reasoning process of the large language model. It also retrieves triples related to the input text from the perceptual knowledge graph as explicit node / edge evidence to guide the large language model to perform graph-like thought chain reasoning, so that the perceptual judgment result of the large language model is aligned with the preset human standards.

[0061] In the example, the explicit rule set generation unit 301 is further configured to: convert the output text into semantic vectors to form an initial vector set; use an auxiliary semantic model to semantically enhance the vectors in the initial vector set to obtain an enhanced feature vector set; determine the optimal number of clusters for the enhanced feature vector set based on a multi-index joint decision-making strategy, and perform clustering to obtain multiple clusters; for each cluster, select multiple representative samples close to the cluster center, and use a rule summarization large language model to summarize the multiple representative samples to generate explicit rules; and generate an explicit rule set based on the multiple explicit rules corresponding to the multiple clusters.

[0062] In the example, the explicit rule set generation unit 301 is further configured to: determine multiple candidate cluster numbers through multiple trials, calculate the silhouette coefficient, Davies-Bouldin index, and Calinski-Harabasz index for each candidate cluster number, and determine the optimal cluster number by combining the tree structure changes of hierarchical clustering and the performance of repeated experiments on cluster stability through a multi-index comprehensive score function.

[0063] In the example, the reasoning unit 304 is further configured to: vectorize the input text and the rules in the human perception standard rule base respectively through a multi-granularity embedding model to construct a rule vector index; retrieve the most relevant human perception rules to the input text from the rule vector index based on vector similarity; and input the retrieved human perception rules as explicit prompts, together with the input text, into the large language model to drive the large language model to perform reasoning and judgment.

[0064] In the example, the perceptual knowledge graph construction unit 303 is further configured to: extract triplets including the analysis subject, perceptual arguments, and judgment conclusions from the explicit rule set, the human perception standard rule base, and the thought chain text of the large language model; take the analysis subject whose frequency is greater than a preset threshold as the core node, aggregate the related perceptual arguments and judgment conclusions to form triplets, and cluster the formed triplet set according to the analysis subject to form a preliminary graph edge set; convert the preliminary graph edge set into graph structure grammar and store it in the graph database to construct the perceptual knowledge graph.

[0065] In the example, the reasoning unit 304 is further configured to: convert triples in the perceptual knowledge graph into natural language descriptive text and vectorize them to construct a triple vector index; retrieve triple text related to the current context from the triple vector index and inject it as explicit node / side evidence into the reasoning process of the large language model to guide the large language model to generate a graph-like thought chain containing explicit node / side evidence.

[0066] In the example, the reasoning unit 304 is further configured to: when multiple related triplet texts are retrieved, the large language model reverse-maps the multiple related triplet texts back to the original graph elements, and performs logical completion and / or conflict detection based on the association path of the original graph elements in the perceptual knowledge graph, so as to further constrain the graph-like thought chain reasoning of the large language model.

[0067] The above combination Figure 1 and Figure 2 The specific operations shown are respectively by Figure 3 The corresponding unit in the large language model perception alignment device 300 shown is used to perform the operation. Here, the specific operational details will not be elaborated.

[0068] Figure 4 This is a block diagram illustrating a computing system including at least one computing device and at least one storage device of storage instructions according to an exemplary embodiment of the present disclosure.

[0069] like Figure 4As shown, the computing system 400 provided according to an exemplary embodiment of the present invention includes a computing device 401 and a storage device 402. The storage device 402 stores computer-executable instructions. When the computer-executable instructions are executed by the computing device 401, the large language model-aware alignment method described in any of the foregoing embodiments is executed.

[0070] The computing device 401 can be deployed in a server or client, or on a node device in a distributed network environment. Furthermore, the computing device 401 can be a PC, tablet, personal digital assistant, smartphone, web application, or other device capable of executing the aforementioned set of instructions. Here, the computing device is not necessarily a single computing device; it can be any collection of devices or circuits capable of executing the aforementioned instructions (or instruction sets) individually or in combination. The computing device can also be part of an integrated control system or system manager, or can be configured to interconnect with a portable electronic device locally or remotely (e.g., via wireless transmission). In the computing device, the processor includes a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, the processor also includes analog processors, digital processors, microprocessors, multi-core processors, processor arrays, network processors, etc.

[0071] According to another aspect of this disclosure, a computer-readable storage medium is provided that stores instructions, which, when executed by at least one computing device, cause the at least one computing device to perform the large language model-aware alignment method described in any of the foregoing embodiments. The computer-readable storage medium includes magnetic media such as floppy disks and magnetic tapes, optical media (including optical disc (CD) ROMs and DVD-ROMs), magneto-optical media such as floppy discs, hardware devices such as ROMs and RAMs designed for storing and executing program commands, and flash memory. The instructions may include language code executable by a computer using an interpreter and machine language code generated by a compiler.

[0072] By adopting this disclosure, the stability, consistency and interpretability of large language models in complex text perception and classification tasks (e.g., malicious content identification and sentiment analysis) are significantly improved, overcoming the technical bottlenecks of traditional fine-tuning methods such as weak ability to align perception standards, difficulty in rule transfer, and opaque inference chains; and achieving controllable, interpretable and stable perception and risk control alignment effects without sacrificing generation capabilities.

[0073] The processes, methods, or algorithms disclosed herein can be transmitted to, or implemented by, a processing device, controller, or computer, which may include any existing programmable electronic control unit or a dedicated electronic control unit. Similarly, the processes, methods, or algorithms can be stored in various forms as data and instructions executable by a controller or computer, including but not limited to information permanently stored on non-writable storage media (such as ROM devices) and information variablely stored on writable storage media (such as floppy disks, magnetic tapes, CDs, RAM devices, and other magnetic and optical media). The processes, methods, or algorithms can also be implemented in a software executable object. Optionally, the processes, methods, or algorithms can be implemented wholly or partially using suitable hardware components (such as ASICs, FPGAs, state machines, controllers, or other hardware components or devices) or a combination of hardware components, software components, and firmware components.

[0074] Although this disclosure includes specific examples, it will be apparent to those skilled in the art that various changes in form and detail may be made to these examples without departing from the spirit and scope of the claims and their equivalents. The examples described herein are to be considered merely for descriptive purposes and not for limiting purposes. The description of features or aspects in each example is to be considered applicable to similar features or aspects in other examples. Suitable results may be obtained if the described techniques are performed in a different order, and / or if components in the described system, architecture, apparatus, or circuit are combined in a different manner and / or if components in the described system, architecture, apparatus, or circuit are replaced or supplemented with other components or their equivalents. Therefore, the scope of this disclosure is not limited by the specific embodiments but by the claims and their equivalents, and all variations within the scope of the claims and their equivalents shall be construed as included in this disclosure.

Claims

1. A perceptual alignment method for large language models, characterized in that, The large language model-aware alignment method includes: Obtain the output text of the large language model for a preset perception task, and generate the explicit rule set of the large language model based on the output text through semantic vectorization, cluster analysis and rule inversion; Obtain human-annotated perceptual samples, and perform semantic clustering and abstraction summarization on the human-annotated perceptual samples to generate a human perception standard rule base. Based on the explicit rule set and the human perception standard rule base, a perceptual knowledge graph is constructed, which stores triplets consisting of the analysis subject, perceptual arguments, and judgment conclusions. For the input text of the large language model, human perception standard rules related to the input text are retrieved from the human perception standard rule base and injected into the reasoning process of the large language model. Furthermore, triples related to the input text are retrieved from the perception knowledge graph as explicit node / edge evidence to guide the large language model in graph-based thought chain reasoning, so that the perception discrimination result of the large language model aligns with the preset human standards. The steps for generating the explicit rule set of the large language model based on the output text through semantic vectorization, cluster analysis, and rule inversion include: The output text is converted into semantic vectors to form an initial vector set; The vectors in the initial vector set are semantically enhanced using an auxiliary semantic model to obtain an enhanced feature vector set. Based on a multi-index joint decision-making strategy, the optimal number of clusters for the enhanced feature vector set is determined, and clustering is performed to obtain multiple clusters. For each of the multiple clusters, select multiple representative samples close to the cluster center, and use a rule summarization big language model to summarize the multiple representative samples to generate explicit rules. The explicit rule set is generated based on the explicit rules corresponding to the multiple clusters.

2. The large language model perceptual alignment method according to claim 1, characterized in that, The steps for determining the optimal number of clusters for the enhanced feature vector set based on a multi-index joint decision-making strategy include: Multiple candidate cluster numbers were determined through repeated experiments. The silhouette coefficient, Davies-Bouldin index, and Calinski-Harabasz index of each candidate cluster number were calculated. The optimal cluster number was determined by combining the tree structure changes of hierarchical clustering and the performance of repeated cluster stability experiments through a multi-index comprehensive scoring function.

3. The large language model perceptual alignment method according to claim 1, characterized in that, The steps of retrieving human perception standard rules related to the input text from the human perception standard rule base and injecting them into the reasoning process of the large language model, for the input text of the large language model, include: The input text and the rules in the human perception standard rule base are vectorized using a multi-granularity embedding model to construct a rule vector index. Based on vector similarity, the most relevant human-perceptual rules to the input text are retrieved from the rule vector index. The retrieved multiple human perception rules are used as explicit prompts and input together with the input text into the large language model to drive the large language model to perform reasoning and judgment.

4. The large language model perceptual alignment method according to claim 1, characterized in that, The steps for constructing a perceptual knowledge graph based on the explicit rule set and the human perception standard rule base include: Extract triplets comprising the analytical subject, perceptual arguments, and discriminative conclusions from the explicit rule set, the human perception standard rule base, and the thought chain text of the large language model; Using the analysis subject whose frequency of occurrence exceeds a preset threshold as the core node, the related perceptual arguments and judgment conclusions are aggregated to form a triplet, and the formed triplet set is clustered according to the analysis subject to form a preliminary graph edge set; The initial graph edge set is converted into graph structure grammar and stored in a graph database to construct the perceptual knowledge graph.

5. The large language model perceptual alignment method according to claim 1, characterized in that, The steps of retrieving triples related to the input text from the perceptual knowledge graph as explicit node / edge evidence to guide the large language model in graph-based reasoning include: The triples in the perceptual knowledge graph are converted into natural language description text and vectorized to construct a triple vector index. The text of triples relevant to the current context is retrieved from the triple vector index and injected as explicit node / edge evidence into the reasoning process of the large language model to guide the large language model to generate a graph-like thought chain containing the explicit node / edge evidence.

6. The large language model perceptual alignment method according to claim 5, characterized in that, The large language model-aware alignment method also includes: When multiple related triplet texts are retrieved, the large language model reverse maps the multiple related triplet texts back to the original graph elements, and performs logical completion and / or conflict detection based on the association path of the original graph elements in the perceptual knowledge graph, so as to further constrain the graph-like thought chain reasoning of the large language model.

7. A large language model perceptual alignment device, characterized in that, The large language model-aware alignment device includes: The explicit rule set generation unit is configured to acquire the output text of the large language model for a preset perception task, and generate the explicit rule set of the large language model based on the output text through semantic vectorization, cluster analysis and rule inversion. The human perception standard rule base generation unit is configured to acquire human-annotated perception samples and perform semantic clustering and abstraction summarization on the human-annotated perception samples to generate a human perception standard rule base. The perceptual knowledge graph construction unit is configured to construct a perceptual knowledge graph based on the explicit rule set and the human perception standard rule base. The perceptual knowledge graph stores triplets consisting of the analysis subject, perceptual arguments, and judgment conclusions. The reasoning unit is configured to, for the input text of the large language model, retrieve human perception standard rules related to the input text from the human perception standard rule base and inject them into the reasoning process of the large language model; and retrieve triples related to the input text from the perceptual knowledge graph as explicit node / edge evidence to guide the large language model to perform graph-based thought chain reasoning, so that the perceptual discrimination result of the large language model is aligned with the preset human standard. The explicit rule set generation unit is further configured as follows: The output text is converted into semantic vectors to form an initial vector set; The vectors in the initial vector set are semantically enhanced using an auxiliary semantic model to obtain an enhanced feature vector set. Based on a multi-index joint decision-making strategy, the optimal number of clusters for the enhanced feature vector set is determined, and clustering is performed to obtain multiple clusters. For each of the multiple clusters, select multiple representative samples close to the cluster center, and use a rule summarization big language model to summarize the multiple representative samples to generate explicit rules. The explicit rule set is generated based on the explicit rules corresponding to the multiple clusters.

8. A computing system comprising at least one computing device and at least one storage device for storing instructions, characterized in that, When the instruction is executed by the at least one computing device, it causes the at least one computing device to perform the large language model-aware alignment method according to any one of claims 1 to 6.

9. A computer-readable storage medium for storing instructions, characterized in that, When the instruction is executed by at least one computing device, it causes the at least one computing device to perform the large language model-aware alignment method according to any one of claims 1 to 6.

Images