Hybrid architecture semantic restoration method and device based on DFA and large language model

Abstract

The invention discloses a hybrid architecture semantic restoration method and device based on a DFA and a large language model, and relates to the technical field of semantic restoration, the method comprises the following steps: receiving an original power text sequence, loading a terminology benchmark library, and carrying out full scanning based on the DFA to output a candidate anchor point set; locking the candidate anchor point set, extracting a context feature window of each anchor point, and calculating a semantic coupling probability of each anchor point and a context in the context feature window; determining a repair path between each anchor point and a context in the context feature window based on the semantic coupling probability, and performing semantic repair on each anchor point and the context based on the repair path to obtain a corresponding standardized term fragment; and integrating the standardized term fragments to form a complete standardized text sequence corresponding to the original power text sequence. According to the invention, the problem of low accuracy and efficiency during conversion from a spoken text to a standardized term in the prior art is solved.

Description

A Hybrid Architecture-Based Semantic Repair Method and Apparatus Based on DFA and Large Language Model Technical Field

[0001] This invention relates to the field of semantic repair technology, and in particular to a semantic repair method and apparatus based on a hybrid architecture of DFA and large language model. Background Technology

[0002] In industrial-grade RAG systems, intelligent search, automated defect record processing, and automated report generation in vertical fields such as power design and operation and maintenance, the text content of front-line staff's operation instructions, defect descriptions, and search queries often contains a large number of colloquial expressions, industry non-standard abbreviations, and long-tail informal expressions. However, the power professional knowledge base and business systems use a standardized professional terminology system. The difference in expression between the two can easily lead to a mismatch between the text semantics and the knowledge base content, resulting in problems such as knowledge retrieval failure and low accuracy of data structuring processing, which seriously affects the operating efficiency and decision-making accuracy of power business systems.

[0003] To address the alignment issue between colloquial text and standardized terminology, existing technologies primarily employ three approaches: rule / dictionary-based hard replacement, optimized replacement based on NLP segmentation and NER techniques, and lexical mapping using a pure Large Language Model (LLM). Additionally, some improved solutions utilize offline thesaurus combined with dual-link search to rewrite non-standard queries. However, these approaches all suffer from significant technical limitations, making them ill-suited for the highly specialized, complexly nested terminology, and demanding requirements for processing efficiency and accuracy within the power industry. Specific problems include: 1. Rule / dictionary-based hard replacement... The replacement scheme uses a pre-defined vocabulary mapping table to perform simple string matching and direct replacement of text. This scheme lacks the ability to perceive the textual context and cannot recognize the numerous nested entities and complex professional terms present in the power industry. It is prone to destroying entity boundaries and incorrectly disassembling professional entities that should be treated as a whole. For example, replacing "ground screw" in "ground nut" with "ground anchor bolt" results in the incorrect expression "ground anchor bolt mother". At the same time, mechanically replacing local words of independent professional entities will destroy the logical structure of the original entities and fail to meet the structural integrity requirements of power professional terms.

[0004] 2. The optimized replacement scheme based on NLP word segmentation and NER technology, although adding an entity recognition step on the basis of hard replacement, does not fully cover the colloquial and non-standard expressions in the power field due to the lack of a general word segmenter's lexicon. It is prone to word segmentation ambiguity and over-segmentation problems at non-standard terms, resulting in low long-tail entity recognition rate. Moreover, in highly specialized long-tail distribution scenarios such as power engineering, the NER model has limited coverage of on-site colloquialisms and informal abbreviations, which is prone to recognition omissions, ultimately causing terminology conversion failure. It still cannot solve the problem of accurate alignment between colloquial text and standardized terms.

[0005] 3. The vocabulary mapping scheme of pure large language model (LLM) inputs all the correspondence between spoken language and professional vocabulary into the model for semantic transformation. This scheme has significant instability and illusion problems. Without strict constraints, the model is prone to generating random vocabulary mappings, distorting certain power professional terms into fictitious words, and even mapping polysemous terms to meanings outside the power field. At the same time, when processing massive amounts of deterministic power terms, the model consumes a lot of tokens, and its performance is limited by the length of the context window. There are serious problems of resource redundancy and response latency, which cannot meet the rigid requirements of real-time and high concurrency in scenarios such as power dispatching and compliance review.

[0006] 4. Existing solutions for rewriting non-standard queries segment the input text and correct non-standard entities through offline thesaurus mining or entity recognition models, and then combine the results of dual-link search fusion. However, these solutions still fail to overcome the technical bottlenecks of traditional word segmentation and entity recognition, lack the ability to determine the semantic boundaries of nested composite entities, and do not solve the efficiency problem of processing massive amounts of terms. They cannot fundamentally avoid core problems such as semantic fragmentation, polysemy, and low recall accuracy.

[0007] In summary, existing technologies cannot simultaneously meet the three core requirements of the power industry for high-concurrency processing efficiency of terminology conversion, semantic accuracy in complex contexts, and structural integrity of professional entities. There is an urgent need for a new semantic repair method to achieve accurate and efficient conversion of colloquial power texts into standardized terms, thereby solving the industry pain point of semantic alignment in vertical domain texts. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a semantic repair method and apparatus based on a hybrid architecture of DFA and large language model, which aims to solve the problems of low accuracy and efficiency in the conversion of colloquial text to standardized terminology in the prior art.

[0009] This invention proposes a semantic repair method based on a hybrid architecture of DFA and a large language model. The method includes: receiving the original power text sequence to be analyzed; loading a professional terminology benchmark library and performing a millisecond-level full scan based on DFA to output a candidate anchor set; locking the candidate anchor set and extracting the context feature window of each anchor in the candidate anchor set; calculating the semantic coupling probability between each anchor and the context in the context feature window based on a large model with thought chain reasoning ability; determining the repair path between each anchor and the context in the context feature window based on the semantic coupling probability; performing semantic repair on each anchor and the context based on the repair path to obtain the corresponding standardized term fragments; and integrating the standardized term fragments to form a complete standardized text sequence corresponding to the original power text sequence.

[0010] Furthermore, in the aforementioned hybrid architecture semantic repair method based on DFA and a large language model, the steps of receiving the original power text sequence to be analyzed, loading a technical terminology benchmark library, and performing a millisecond-level full scan based on DFA to output a candidate anchor set include: receiving the original power text sequence. ,in, For a single character in the text, Set the total text length; load the preset electrical industry slang-professional vocabulary ontology base library. Benchmark library It includes industry-proven 1:1 deterministic terminology mappings; and utilizes a DFA-based deterministic perception function. A millisecond-level full scan is performed on the initialized raw power text sequence to quickly locate potential non-standard terminology candidate segments; based on these candidate segments, segments matching the benchmark library are selected. Matched candidate terms form a set of candidate anchor points. .

[0011] Furthermore, in the aforementioned hybrid architecture semantic repair method based on DFA and a large language model, the formula for calculating the semantic coupling probability is as follows: In the formula, This represents the semantic coupling probability, with a value range of [0,1]. A large model with the ability to reason through thought processes. anchor point The semantic feature vector of the power industry, For context window The semantic feature vector obtained through its model processing, The inner product of eigenvectors. For logical constraint coefficients within a professional field, The normalization function maps the core calculation results to probability values ​​in the interval [0,1].

[0012] Furthermore, in the aforementioned hybrid architecture semantic repair method based on DFA and large language model, the step of determining the repair path between each anchor point and the context in the context feature window based on the semantic coupling probability includes: when the semantic coupling probability is greater than a preset coupling probability threshold, the repair path between each anchor point and the context in the context feature window adopts a composite repair path; when the semantic coupling probability is less than or equal to the preset coupling probability threshold, the repair path between each anchor point and the context in the context feature window adopts an atomic mapping path.

[0013] Furthermore, in the aforementioned hybrid architecture semantic repair method based on DFA and a large language model, the step of semantically repairing each anchor point and context based on the repair path to obtain the corresponding standardized term fragment includes: when a composite repair path is adopted, initiating the semantic repair mechanism. ;in, As anchor point, For context, Standard terminology corresponding to composite entities of anchor points and context; when using atomic state mapping paths, a regular alignment mechanism is performed. ;in, As anchor point, The standard terminology corresponding to anchor points.

[0014] Furthermore, in the above-mentioned hybrid architecture semantic repair method based on DFA and large language model, the normalization function of the standardized text sequence is: In the formula, The final output is a standardized text result. This is the original electrical text sequence. A benchmark database for electrical industry slang and professional vocabulary. For context-constrained parameters, For the set of candidate anchor points and preliminary matching of standard terminology mapping, For probabilistic decision functions based on large models with reasoning capabilities, The computation time for the fusion function; loading the terminology benchmark library and performing a millisecond-level full scan based on DFA to output the candidate anchor point set. It satisfies the linear time complexity model, and the specific formula is as follows: In the formula, For input text sequence Total length, For the benchmark library The number of terms that successfully match the input text, i.e., the set of candidate anchors. The number of elements, For the first The character length of each successfully matched term. The total character length of all matching terms.

[0015] Furthermore, in the aforementioned hybrid architecture semantic repair method based on DFA and a large language model, after the step of integrating standardized term fragments to form a complete standardized text sequence corresponding to the original power text sequence, the method further includes: [further details about the standardized text sequence are needed for a complete translation]. Feedback is sent to subsequent business systems to provide accurate semantic input for actual business scenarios; a retrieval gain model is defined to evaluate the effect of normalization processing on knowledge activation, with the specific formula as follows: In the formula, To retrieve the gain coefficient, Original power text sequence Number of search hits in the RAG system For standardized text The number of search hits, For standardized text semantic feature vectors, This is the feature vector of the index entries in the technical terminology benchmark database. Let be the cosine distance between the two, with values ​​ranging from [0,1]. It is a positive gain index; if If the value is below the threshold, supplement the professional terminology benchmark library. The terms used, the size of the context window, or the size of the preset probability threshold.

[0016] Another objective of this invention is to provide a semantic repair device based on a hybrid architecture of DFA and a large language model. The device includes: a receiving module for receiving the original power text sequence to be analyzed, loading a technical terminology benchmark library, and performing a millisecond-level full scan based on DFA to output a candidate anchor set; a locking module for locking the candidate anchor set and extracting the context feature window of each anchor in the candidate anchor set, and calculating the semantic coupling probability between each anchor and the context in the context feature window based on a large model with thought chain reasoning capabilities; a repair module for determining the repair path between each anchor and the context in the context feature window based on the semantic coupling probability, and performing semantic repair on each anchor and its context based on the repair path to obtain the corresponding standardized term fragments; and an integration module for integrating the standardized term fragments to form a complete standardized text sequence corresponding to the original power text sequence.

[0017] Another object of the present invention is to provide a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0018] Another object of the present invention is to provide an electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0019] This invention receives the original power industry text sequence and loads a professional terminology benchmark library. Based on Descriptive Fractional Analysis (DFA), it performs a millisecond-level full scan to output a candidate anchor point set. It then locks the candidate anchor point set and extracts the context feature window for each anchor point. Relying on a large model with thought chain reasoning capabilities, it calculates the semantic coupling probability between the anchor point and its context. Based on the semantic coupling probability, it determines a repair path—either composite state repair or atomic state mapping—and performs semantic repair on the anchor point and its context according to the path to obtain standardized terminology fragments. Finally, it integrates these standardized terminology fragments to form a complete standardized text sequence. This invention utilizes the strong determinism and high-concurrency scanning capabilities of DFA to achieve efficient localization of non-standard terms. Combined with the deep semantic reasoning capabilities of the large model's thought chain, it accurately determines entity boundaries and eliminates ambiguity. Simultaneously, the differentiated repair paths of composite and atomic states ensure the semantic atomicity of professional entities, balancing the processing efficiency and semantic accuracy of power industry terminology conversion. This fundamentally solves the technical problems of semantic fragmentation, entity boundary destruction, ambiguity illusion, low processing efficiency, and poor recall accuracy in existing technologies for processing colloquial power industry text. It addresses the issues of low accuracy and efficiency in existing technologies for converting colloquial text to standardized terminology.

[0020] In addition, the embodiments of the present invention have at least the following beneficial effects: 1. The first "dual-modal collaborative" architecture: the basic processing framework is built through DFA automata to ensure strong determinism and high concurrency in text segmentation and entity fragment capture, which can efficiently cope with the parallel processing needs of massive power operation and maintenance and scheduling instructions; at the same time, it links LLM nodes to carry out deep semantic analysis, and with the contextual understanding ability of the large model, it can achieve high fault tolerance in handling colloquial expressions and ambiguous words, taking into account both processing efficiency and intelligent parsing level, and solving the contradiction of traditional technology that is "efficient but not accurate, intelligent but weak in performance".

[0021] 2. Overcoming the "Semantic Fragmentation" Dilemma: Addressing the characteristics of numerous nested engineering terms and complex expressions in the power sector, a "semantic atomicity" protection mechanism is established. This mechanism can accurately identify the semantic boundaries of composite and independent entities, avoiding mechanical disassembly and one-sided replacement of nested entities. For example, for composite entities composed of "ground screw + mother," it avoids the problem of generating incorrect terms after traditional disassembly. For independent entities such as "substation," it prevents semantic conflicts caused by hard replacement of local terms, ensuring the structural integrity and semantic accuracy of engineering terminology.

[0022] 3. Enhance the recall purity of the RAG system: Reconstruct the knowledge activation logic, upgrading the traditional "fuzzy probability matching" to "deterministic entity-level indexing." Based on accurate entity recognition and semantic repair results, a standardized entity knowledge index is constructed. This effectively eliminates information noise caused by non-standard colloquialisms and terminological biases, enabling the RAG system to accurately recall task-related professional knowledge in power design, operation and maintenance, and scheduling scenarios. This significantly reduces decision-making biases caused by information distortion, improving the reliability and security of power scenario decisions. Attached Figure Description

[0023] Figure 1 is a flowchart of the hybrid architecture semantic repair method based on DFA and large language model in the first embodiment of the present invention; Figure 2 is a structural block diagram of the hybrid architecture semantic repair device based on DFA and large language model in the third embodiment of the present invention.

[0024] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0025] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0026] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] Please refer to Figure 1 for Example 1, which shows the semantic repair method based on a hybrid architecture of DFA and a large language model in the first embodiment of the present invention. The method includes steps S10 to S13.

[0029] Step S10: Receive the original power text sequence to be analyzed, load the technical terminology benchmark library, and perform a millisecond-level full scan based on DFA to output a set of candidate anchor points.

[0030] The system first receives the original power text sequence. ,in For a single character in the text, Given the total length of the text, initialize the input sequence.

[0031] Subsequently, the technical terminology benchmark library is loaded; the pre-set electrical slang-technical vocabulary ontology benchmark library is also loaded. The library contains industry-verified 1:1 deterministic terminology mappings (such as "CT" → "current transformer", "ground screw" → "anchor bolt", etc.), providing a basis for subsequent terminology matching.

[0032] Then, a full scan based on DFA is performed; a deterministic sensing function based on DFA (Deterministic Finite State Automaton) is then used. (“ "Represents a perceptual mapping relationship," "Deterministic" is an abbreviation for "Deterministic", which represents the strong deterministic matching characteristic of the function. It performs a full scan of the initialized original power text in milliseconds to quickly locate potential non-standard term candidate segments. The core logic is: first, add the benchmark library. All terms are constructed into a finite state transition graph, and then the input text is traversed once to complete the full term matching without repeated scanning, ensuring efficient location of potential non-standard term candidate segments.

[0033] Finally, output the set of candidate anchor points; filter them by scanning to find those that match the benchmark library. Matched candidate terms form a set of candidate anchor points. This provides the core analysis object for subsequent steps.

[0034] In addition, the computation time of this stage It satisfies the linear time complexity model, and the specific formula is as follows: In the formula For input text sequence Total length, For the benchmark library The number of terms that successfully match the input text (i.e., the set of candidate anchors) (number of elements) For the first The character length of each successfully matched term. The total character length of all matching terms. The core value of this efficiency model lies in quantifying and ensuring the high-concurrency processing capability of the perception layer, directly addressing the resource redundancy and response latency pain points of pure LLM solutions in massive text processing. By explicitly defining the linear time complexity characteristics, it ensures that when processing large batches of power operation and maintenance logs, design specifications, and other texts, the system's processing time only increases linearly with the text length and the length of the matching terms, without any additional exponential overhead. This better and more stably adapts to the rigid requirements of real-time performance and high concurrency in scenarios such as power dispatching and compliance review.

[0035] Step S11: Lock the candidate anchor set and extract the context feature window of each anchor in the candidate anchor set. Calculate the semantic coupling probability between each anchor and the context in the context feature window based on a large model with thought chain reasoning ability.

[0036] The candidate anchor point locking and context feature window extraction mainly involve the following steps: locking the candidate anchor point set, extracting the context feature window, and defining the overall normalization function. Specifically: locking the candidate anchor point set; based on the candidate anchor point set A output by S1, clarifying the core objects of subsequent analysis to ensure the effective transmission of anchor point information.

[0037] Extract the contextual feature window; for each anchor point Automatically extract its immediate context feature window ( As The core input parameters of the function capture contextual information that directly participates in the semantic coupling probability calculation, such as... middle The corresponding context feature vector, i.e. based on generate. This is the key basis for determining whether "ground snail + mother" is a composite entity and "main transformer + CT" has no composite association, ensuring that semantic decisions are not divorced from the specific context. The default context feature window range is 3-5 characters before and after the anchor point, which can be flexibly configured according to the characteristics of power industry terminology. It captures the contextual association information of the anchor point (such as the word "mother" after "ground snail") and passes it into the overall normalization formula; it defines an overall normalization function; this function is named "normalization function" because its essence is to unify and "merge" the colloquial, non-standard, and multi-form terminology expressions in the power industry into the standardized terminology system stipulated by the national / industry standards of the power industry, eliminate semantic differences caused by differences in expression, and achieve "consistency and standardization" of terminology semantics and format.

[0038] To achieve the organic integration of "perception layer results" and "decision layer analysis", a core normalization function is defined for the entire process. The specific formula is as follows:

[0039] In the formula The final output is a standardized text result. This is the original input text. A benchmark database for electrical industry slang and professional vocabulary. Contextual constraint parameters (including classification rules for professional categories in the power field, rules for taboos in terminology collocation, and the coupling probability threshold required later) wait), For the output result (candidate anchor set) and preliminary matching of standard terminology mapping). The core significance of this probabilistic decision function based on LLM's CoT (with "anchor point + context + constraint parameters" as input and semantic judgment result as output) is to make up for the lack of contextual understanding in the deterministic perception layer. By using the logical reasoning ability of LLM, it can accurately determine whether the anchor point is a composite entity and whether there are complex semantic issues such as polysemy, providing core decision basis for nested entity repair and ambiguity elimination. As a fusion function, it integrates two types of results according to the principle of "deterministic mapping first, probabilistic correction and completion". Its core significance is to achieve the organic synergy between "high concurrency efficiency of the perception layer" and "precise semantics of the decision layer". It retains the efficiency of simple term conversion and ensures the accuracy of conversion in complex scenarios, fundamentally solving the contradiction of "high efficiency but insufficient accuracy, intelligence but weak concurrency" in traditional single-modal processing.

[0040] Furthermore, the formula for calculating the semantic coupling probability is as follows: In the formula, This represents the semantic coupling probability, with a value range of [0,1]. A large model with the ability to reason through thought processes. anchor point The semantic feature vector of the power industry, For context window The semantic feature vector obtained through its model processing, The inner product of eigenvectors. For logical constraint coefficients within a professional field, The normalization function maps the core calculation results to probability values ​​in the interval [0,1].

[0041] Specifically, LLM decision nodes are employed to provide computational support for deep semantic analysis. The Chain of Reasoning (CoT) technology is introduced to enhance the logical reasoning capabilities of LLM, enabling it to make accurate semantic judgments based on the power industry context and avoid unconstrained probabilistic predictions. For example, the essence of CoT is to allow large models to mimic the human logical reasoning process: instead of directly outputting conclusions, it first breaks down the problem step by step, derives intermediate logic, and finally arrives at the result. For semantic repair scenarios in the power sector, the approach is to have a large model analyze first and then output semantic repair decisions, rather than directly providing a "replace / not replace" conclusion. This avoids errors caused by unconstrained probabilistic predictions. Specifically, a general-purpose large model can be superimposed with CoT (Coding on Technology) prompts and power sector-specific fine-tuning. Common, general-purpose large models with strong logical reasoning capabilities can be used as a foundation. CoT prompts guide the model's step-by-step reasoning, and power sector corpora are used for lightweight fine-tuning. Power sector knowledge graph association rules are embedded to improve the recognition accuracy of industry slang and nested entities. Alternatively, to address the rigid requirements of high concurrency and low latency in power systems, lightweight models such as Qwen3 14B and Qwen3 30B-A3B can be selected. Predefined power sector CoT reasoning templates solidify the logic of coupled probability calculation and entity boundary determination into the model's reasoning steps, avoiding meaningless probability exploration and improving decision-making efficiency (e.g., the template includes "anchor point + context" power entity combination rules, directly limiting the reasoning scope). Therefore, for those skilled in the art, obtaining and setting up a large model with thought chain reasoning capabilities is understandable and will not be elaborated upon here.

[0042] Step S12: Determine the repair path between each anchor point and the context in the context feature window based on the semantic coupling probability, and perform semantic repair on each anchor point and the context based on the repair path to obtain the corresponding standardized term fragment.

[0043] Among them, the coupling probability With preset threshold (Default 0.75) Compare and determine whether the anchor point and the context constitute a composite entity, and output the decision instruction of "composite state repair" or "atomic state mapping", that is, whether to use the composite state repair path or the atomic state mapping path.

[0044] Complex repair branch ( ); Initiate semantic repair mechanism ,in For anchor points (from the candidate anchor point set) ), For context (from the set of candidate anchors) ), Standard terminology for composite entities (taken from the benchmark library) (Middle Composite Entity Mapping Table). Through the overall mapping of "anchor point + context", the semantic atomicity of composite entities is ensured not to be violated. For example, "ground screw + mother" is completely mapped to "anchor bolt nut".

[0045] Atomic state mapping branch ( ); Perform regular alignment mechanism ,in For anchor points (from the candidate anchor point set) ), Standard terminology for anchor points (taken from the benchmark library) (1:1 deterministic mapping relationship). Simultaneously through... and The coefficients eliminate ambiguity, for example, aligning "CT" to "current transformer" and excluding its medical meaning.

[0046] Output standardized terminology fragments; integrate the standardized terminology fragments processed by the two branches to generate complete standardized text later. Provides the foundation.

[0047] Step S13: Integrate the standardized terminology fragments to form a complete standardized text sequence corresponding to the original power text sequence.

[0048] In summary, the hybrid architecture semantic repair method based on DFA and a large language model in the above embodiments of the present invention receives the original power text sequence and loads a professional terminology benchmark library. It then performs a millisecond-level full scan based on DFA to output a candidate anchor point set; locks the candidate anchor point set and extracts the context feature window for each anchor point; calculates the semantic coupling probability between the anchor point and the context using a large model with thought chain reasoning capabilities; determines the repair path (complex state repair or atomic state mapping) based on the semantic coupling probability; performs semantic repair on the anchor point and context according to the path to obtain standardized terminology fragments; and integrates the standardized terminology fragments to form a complete standardized text sequence. This method utilizes the strong determinism and high-concurrency scanning capability of DFA to achieve efficient localization of non-standard terms, combines the deep semantic reasoning capability of the large model's thought chain to accurately determine entity boundaries and eliminate ambiguity, and ensures the semantic atomicity of professional entities through differentiated repair paths of complex and atomic states. It balances the processing efficiency and semantic accuracy of power terminology conversion, fundamentally solving the technical problems of semantic fragmentation, entity boundary destruction, ambiguity illusion, low processing efficiency, and poor recall accuracy in existing power colloquial text processing technologies. It solves the problem of low accuracy and efficiency in the conversion of colloquial text to standardized terminology in existing technologies.

[0049] Example 2 This example also proposes a hybrid architecture semantic repair method based on DFA and a large language model. The difference between the hybrid architecture semantic repair method based on DFA and a large language model in this example and the hybrid architecture semantic repair method based on DFA and a large language model in Example 1 is that: after the step of integrating standardized term fragments to form a complete standardized text sequence corresponding to the original power text sequence, it further includes: [further details about the standardized text sequence are needed for a complete translation.] Feedback is sent to subsequent business systems to provide accurate semantic input for actual business scenarios; a retrieval gain model is defined to evaluate the effect of normalization processing on knowledge activation, with the specific formula as follows: In the formula, To retrieve the gain coefficient, Original power text sequence Number of search hits in the RAG system For standardized text The number of search hits, For standardized text semantic feature vectors, This is the feature vector of the index entries in the technical terminology benchmark database. Let be the cosine distance between the two, with values ​​ranging from [0,1]. It is a positive gain index; if If the value is below the threshold, supplement the professional terminology benchmark library. The terms used, the size of the context window, or the size of the preset probability threshold.

[0050] The process involves integrating standardized terminology fragments to form a complete standardized text sequence corresponding to the original power text sequence. Subsequent steps include standardized text sequence feedback, retrieval gain evaluation, and benchmark library and parameter optimization. First, the semantically repaired standardized text is transformed into usable input for actual power business operations, enabling the seamless integration of technical solutions with business scenarios. After generating the complete standardized text sequence, the system pushes it to various subsequent business systems in the power sector via a pre-defined interface protocol. This protocol can employ commonly used industrial-grade data interaction methods such as RESTful APIs and RPCs to ensure data transmission stability and compatibility.

[0051] Subsequent business systems include power industry-grade RAG systems, power intelligent search platforms, automated power operation and maintenance defect record processing systems, and power automated report generation systems. Standardized text sequences serve as the core semantic input for these systems, replacing the original colloquial and non-standardized raw text, thus ensuring the accuracy of business system processing from the data source.

[0052] For example, in the automated processing system for power operation and maintenance defect records, after the original non-standard text is repaired into standardized text, the system can directly complete the structured entry of defect ledgers, automatic determination of defect levels, and automatic generation of operation and maintenance work orders based on the text; in the power RAG system, standardized text is used as a search question and can be directly matched with standardized professional content in the knowledge base, avoiding knowledge retrieval failures due to differences in expression.

[0053] After receiving the standardized text sequence, the business system performs data format verification. If the verification passes, the data is directly incorporated into the business processing flow. If the verification fails, the system sends an error message to the semantic repair system, which then re-executes the repair process to ensure that the text sent back to the business system is valid data that conforms to power industry standards.

[0054] Secondly, the actual effect of semantic repair is evaluated through a quantitative model. The quantitative dimension focuses on the degree of improvement in knowledge activation, that is, the effect of standardized text in professional system retrieval and matching. The core is to calculate the retrieval gain coefficient through the retrieval gain model to realize the quantification and judgment of the repair effect.

[0055] The retrieval gain model uses the retrieval gain coefficient as the core evaluation indicator. This coefficient is calculated using a fixed formula. Each parameter in the formula is a quantitative indicator that can be directly collected or calculated in the power business system. The specific acquisition and calculation methods of each indicator are as follows: 1. The number of retrieval hits of the original power text sequence in the RAG system. This is the number of relevant knowledge base contents matched by the system according to the original retrieval logic after the original colloquial text without any semantic repair is input into the power RAG system. This value is directly obtained from the retrieval log statistics of the RAG system. For example, if only 1 relevant content is hit after the original text "main transformer CT abnormality" is input, the value is 1.

[0056] 2. The number of retrieval hits for standardized text is the number of relevant knowledge base contents matched when the semantically repaired standardized text is input into the same power RAG system and the same retrieval logic is applied. This number is also obtained from the retrieval log statistics of the RAG system. For example, if the standardized text "main transformer current transformer operation abnormal" is input and 8 relevant contents are matched, then the value is 8.

[0057] 3. The semantic feature vector of standardized text is a numerical vector obtained by vectorizing the standardized text through a word vector model pre-trained in the power field. The word vector model can be Word2Vec, BERT, or other models adapted to professional texts. The model is fine-tuned with a power professional corpus to ensure that the vectorization result can accurately represent the semantic features of power text. For example, the standardized text "the secondary equipment box of the current transformer is damaged" is vectorized to generate a one-dimensional vector composed of several numerical values.

[0058] 4. The feature vectors of the index entries in the professional terminology benchmark library are numerical vectors obtained by pre-vectorizing all standardized terms and professional terms in the professional terminology benchmark library using the same power field word vector model. All index entries generate corresponding feature vectors and store them in the vector index library of the benchmark library to achieve fast vector matching and calculation.

[0059] 5. The cosine distance between the two is calculated using the cosine similarity algorithm, which compares the semantic feature vector of the standardized text with the feature vector of the indexed entries in the professional terminology benchmark library. The value ranges from 0 to 1. The closer the value is to 0, the higher the semantic similarity between the two. The closer the value is to 1, the lower the semantic similarity between the two. This value is automatically calculated by the system calling the vector calculation engine.

[0060] 6. The positive gain metric is the reciprocal of the cosine distance mentioned above. This metric transforms the inverse metric of semantic similarity into a positive metric. The smaller the cosine distance, the larger the positive gain metric, which represents the higher the semantic matching degree between the standardized text and the benchmark library index entries.

[0061] The retrieval gain coefficient is calculated from the above indicators using the formula. This coefficient is a unitless quantitative value. The larger the value, the better the normalization process of semantic repair improves knowledge activation. The smaller the value, the worse the improvement. This quantitative indicator can intuitively determine the effectiveness of the semantic repair scheme in practical applications and provide data basis for subsequent optimization.

[0062] Finally, based on the evaluation results of the retrieval gain model, the core basic library and key parameters of semantic repair are optimized in a closed loop to achieve self-iteration and self-optimization of the semantic repair scheme, ensuring that it continuously adapts to the ever-changing terminology and business needs in the power field.

[0063] Specifically: 1. Setting the retrieval gain coefficient threshold: The system will preset a reasonable threshold for the retrieval gain coefficient based on the accuracy requirements of different power business scenarios. This threshold is determined by power industry technical personnel in conjunction with actual business results. For example, the power RAG system has high requirements for retrieval accuracy, so the threshold can be set to 5; the power automation report generation system has relatively moderate requirements for retrieval results, so the threshold can be set to 3. The threshold is a fixed value and serves as the core basis for determining whether the repair effect meets the standards. When the actual value of the calculated retrieval gain coefficient is lower than this preset threshold, it is determined that the current semantic repair effect does not meet the business requirements, triggering subsequent optimization operations.

[0064] 2. Supplementing the Terminology Baseline: The terminology baseline is the foundation of semantic repair. When the retrieval gain coefficient is low, the primary optimization direction is to supplement the terminology in the baseline, focusing on non-standard electrical slang, long-tail terms, and newly emerging abbreviations not covered in the baseline. The system will first locate the reason for the low retrieval gain coefficient. If it is determined that some non-standard terms in the original text have not been recognized by the baseline, resulting in incomplete repair, then the terminology will be supplemented through a combination of manual annotation and automatic mining. For example, "relay protection panel" is a non-standard abbreviation commonly used by field maintenance personnel, and there is no corresponding entry in the baseline. This results in the repaired text still not being able to accurately match the knowledge base. In this case, "relay protection panel" and its corresponding standard term "relay protection panel" will be added to the baseline, and a 1:1 deterministic mapping relationship will be established. At the same time, the relevant content in the composite entity mapping table will be supplemented to ensure that the term can be accurately identified in the future.

[0065] 3. Adjusting the Context Window Size: The context window is the core range for extracting contextual information surrounding the anchor point. Its size directly affects the accuracy of calculating the semantic coupling probability between the anchor point and its context. If the window is too large, irrelevant contextual information will be introduced; if the window is too small, key semantic association information will be missing. Both will lead to errors in semantic repair judgment, thereby reducing the retrieval gain effect. When the retrieval gain coefficient is determined to be low due to an unreasonable context window size, the system will adjust the window range according to the actual error case. The default range of the context window is 3-5 characters before and after the anchor point, which can be increased or decreased according to the number of characters. For example, if the context of the anchor point "oil pillow" is "oil pillow position gauge display abnormality", and the original window is 3 characters before and after the anchor point, only the "oil pillow position gauge" part of the content is extracted, resulting in failure to determine it as a composite entity and poor repair effect, then the window is adjusted to 5 characters before and after the anchor point to fully extract the context information and ensure accurate determination of the coupling relationship between the anchor point and its context.

[0066] 4. Adjusting the Preset Probability Threshold: The preset probability threshold is the core basis for determining whether an anchor point and its context constitute a composite entity. The default value is 0.75. When the semantic coupling probability between the anchor point and its context is higher than this threshold, it is determined to be a composite entity and composite state repair is performed. When it is lower than or equal to this threshold, it is determined to be an independent entity and atomic state mapping is performed. Setting the threshold too high or too low will lead to entity identification errors and affect the repair effect. When the low retrieval gain coefficient is caused by an unreasonable preset probability threshold, the system will fine-tune the threshold based on actual identification error cases, gradually optimizing it in units of 0.05. For example, if the original threshold is 0.75, and some anchor points that are actually composite entities have a coupling probability of 0.73 with their context, they are identified as independent entities because they are below the threshold, resulting in semantic fragmentation after repair. In this case, the threshold is adjusted to 0.70 to ensure that such composite entities can be accurately identified, improving the repair effect.

[0067] The above three optimization operations can be executed separately or in combination. The system will select the corresponding optimization method according to the specific reasons for the low retrieval gain coefficient. After optimization, the original raw power text sequence will be re-input into the semantic repair system to perform the full-process semantic repair operation, and the retrieval gain coefficient will be calculated again until the coefficient is higher than the preset threshold to achieve the compliance of the semantic repair effect, forming a complete closed-loop system of "repair - application - evaluation - optimization - re-repair".

[0068] In summary, the semantic repair method based on the hybrid architecture of DFA and large language model in the above embodiments of the present invention receives the original power text sequence and loads the professional term benchmark library, and outputs the candidate anchor point set through millisecond-level full-scale scanning based on DFA; locks the candidate anchor point set and extracts the context feature window of each anchor point, and calculates the semantic coupling probability between the anchor point and the context relying on the large model with the ability of chain-of-thought reasoning; determines the repair path of composite state repair or atomic state mapping according to the semantic coupling probability, and performs semantic repair on the anchor point and the context according to the path to obtain the standardized term fragment; integrates the standardized term fragments to form a complete standardized text sequence. Utilize the strong determinacy and high-concurrency scanning ability of DFA to achieve efficient positioning of non-standard terms, combine the deep semantic reasoning ability of the large model's chain of thought to accurately determine entity boundaries and eliminate polysemy ambiguity, and at the same time ensure the semantic atomicity of professional entities through the differential repair paths of composite state and atomic state, taking into account the processing efficiency and semantic accuracy of power term conversion, and fundamentally solving the technical problems of semantic fragmentation, entity boundary destruction, polysemy ambiguity hallucination, low processing efficiency and poor recall accuracy existing in the processing of power colloquial texts in the prior art. Solve the problem of low accuracy and efficiency in the conversion of colloquial texts to standardized terms in the prior art.

[0069] In addition, in order to further and more detailedly describe the progress of the embodiments of the present invention, exemplarily, several specific application scenarios are listed, including: semantic alignment and intelligent search of the power engineering knowledge base. Suppose a first-line operation and maintenance personnel searches in the system: "The operation of the main transformer CT is abnormal, resulting in the basic stability range height difference exceeding the threshold". The performance comparison of each technical solution is as follows: 1. Code hard replacement: Since "basic stability range" may appear in "basic stability range height difference" or "basic stability range distance", hard replacement is very likely to cause entity boundary cutting errors, resulting in the inability to match "basic stability range height difference" in the standard library.

[0070] 2. NLP word segmentation replacement: The general word segmentation tool will mis-segment "basic stability range height difference" into "basic / stability / range / height / difference", resulting in the inability to recognize the industry-specific colloquial term "basic stability range" because it is chopped up.

[0071] 3. Pure LLM replacement: Facing the polysemous word "CT" (also called CT in the medical field), the LLM may produce hallucinations when lacking strict constraints and map it to a meaning in a non-power field.

[0072] This invention's embodiments accurately locate three core coordinate categories—"main transformer," "CT," and "maintaining elevation difference"—through the DFA perception layer. The LLM decision layer, relying on the CoT (Cooperation of Thought) mechanism, determines that the current context belongs to the professional category of "transmission structure," and then performs a strongly deterministic mapping. Its core advantage lies in effectively solving the problem of poor retrieval results in vertical domains, significantly improving the recall performance of non-standard long-tail keywords compared to traditional solutions.

[0073] The system automatically processes power operation and maintenance defect records in a structured manner. It automatically processes on-site defect descriptions entered by maintenance personnel on mobile devices, such as: "During inspection, it was found that the door lock of the PT secondary box on the main transformer side is damaged, and the oil level gauge is displaying abnormally." A comparison of the performance of various technical solutions reveals the following: 1. Hard code replacement: The system recognizes that "PT" is mapped to the standard term "voltage transformer." However, due to a lack of context awareness, the system mechanically rewrites "PT secondary box" as "voltage transformer secondary box." Although the terminology is correct, in the power standard ledger, the atomic name of this component should be "transformer secondary equipment box." This hard replacement disrupts the hierarchical structure of the equipment components.

[0074] 2. Hard code replacement: The system recognizes that "PT" is mapped to the standard term "voltage transformer". However, due to a lack of context awareness, it mechanically rewrites "PT secondary box" as "voltage transformer secondary box". Although the terminology is correct, in the power standard ledger, the atomic name of this component should be "transformer secondary equipment box". This hard replacement disrupts the hierarchical structure of the equipment components.

[0075] 3. Pure LLM replacement: For directional descriptions such as "main transformer side", LLM is susceptible to probabilistic prediction and may overgeneralize it to "main transformer end", causing the converted text to lose the determinism of power dispatch instructions. At the same time, the huge mapping library may directly fill most of the context, greatly increasing the computational overhead and format instability.

[0076] In this embodiment of the invention, the DFA sensing layer locks in core anchor points such as "PT" and "oil pillow", and then the LLM decision layer identifies that "PT" belongs to the composite component "PT secondary box".

[0077] 1. Semantic Repair: The algorithm does not simply replace the colloquial word "PT", but instead activates a semantic repair mechanism to reconstruct and map "PT secondary box" as a standard entity "mutual inductor secondary equipment box".

[0078] Structure protection: It ensures the semantic atomicity of nested entities in complex phrases, overcomes the dilemma of "semantic fragmentation", and improves the accuracy of structured archiving of operation and maintenance data.

[0079] Please refer to Figure 2 for Embodiment 3, which shows the hybrid architecture semantic repair device based on DFA and a large language model proposed in the third embodiment of the present invention. The device includes: a receiving module 100, used to receive the original power text sequence to be analyzed, load a professional terminology benchmark library, and perform a millisecond-level full scan based on DFA to output a candidate anchor set; a locking module 200, used to lock the candidate anchor set and extract the context feature window of each anchor in the candidate anchor set, and calculate the semantic coupling probability between each anchor and the context in the context feature window based on a large model with thinking chain reasoning ability; a repair module 300, used to determine the repair path between each anchor and the context in the context feature window based on the semantic coupling probability, and perform semantic repair on each anchor and the context based on the repair path to obtain the corresponding standardized term fragments; and an integration module 400, used to integrate the standardized term fragments to form a complete standardized text sequence corresponding to the original power text sequence.

[0080] The functions or operation steps implemented by the above modules are largely the same as those in the above method embodiments, and will not be repeated here.

[0081] Example 4: Another aspect of the present invention provides a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method described in any one of Examples 1 to 2 above.

[0082] Example 5: Another aspect of the present invention provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, it implements the steps of the method described in any one of Examples 1 to 2 above.

[0083] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0084] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable storage medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0085] More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable storage media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0086] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0087] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0088] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A semantic repair method based on a hybrid architecture of DFA and a large language model, characterized in that, The method includes: receiving the original power text sequence to be analyzed, loading a professional terminology benchmark library and performing a millisecond-level full scan based on DFA to output a candidate anchor set; locking the candidate anchor set and extracting the context feature window of each anchor in the candidate anchor set, calculating the semantic coupling probability between each anchor and the context in the context feature window based on a large model with thinking chain reasoning ability; determining the repair path between each anchor and the context in the context feature window based on the semantic coupling probability, and performing semantic repair on each anchor and the context based on the repair path to obtain the corresponding standardized term fragments; integrating the standardized term fragments to form a complete standardized text sequence corresponding to the original power text sequence.

2. The semantic repair method based on a hybrid architecture of DFA and a large language model according to claim 1, characterized in that, The steps of receiving the original power text sequence to be analyzed, loading a technical terminology benchmark library, and performing a millisecond-level full scan based on DFA to output a candidate anchor set include: receiving the original power text sequence. ,in, For a single character in the text, Set the total text length; load the preset electrical industry slang-professional vocabulary ontology base library. Benchmark library It includes industry-proven 1:1 deterministic terminology mappings; and utilizes a DFA-based deterministic perception function. A millisecond-level full scan is performed on the initialized raw power text sequence to quickly locate potential non-standard terminology candidate segments; based on these candidate segments, segments matching the benchmark library are selected. Matched candidate terms form a set of candidate anchor points. 。 3. The semantic repair method based on a hybrid architecture of DFA and a large language model according to claim 2, characterized in that, The formula for calculating the semantic coupling probability is: In the formula, This represents the semantic coupling probability, with a value range of [0,1]. A large model with the ability to reason through thought processes. anchor point The semantic feature vector of the power industry, For context window The semantic feature vector obtained through its model processing, The inner product of eigenvectors. For logical constraint coefficients within a professional field, The normalization function maps the core calculation results to probability values ​​in the interval [0,1].

4. The semantic repair method based on a hybrid architecture of DFA and a large language model according to claim 3, characterized in that, The step of determining the repair path between each anchor point and the context in the context feature window based on the semantic coupling probability includes: when the semantic coupling probability is greater than a preset coupling probability threshold, the repair path between each anchor point and the context in the context feature window adopts a composite state repair path; when the semantic coupling probability is less than or equal to the preset coupling probability threshold, the repair path between each anchor point and the context in the context feature window adopts an atomic state mapping path.

5. The semantic repair method based on a hybrid architecture of DFA and a large language model according to claim 4, characterized in that, The step of semantically repairing each anchor point and its context based on the repair path to obtain the corresponding standardized term fragment includes: when a composite repair path is used, the semantic repair mechanism is activated. ;in, As anchor point, For context, Standard terminology corresponding to composite entities of anchor points and context; when using atomic state mapping paths, a regular alignment mechanism is performed. ;in, As anchor point, The standard terminology corresponding to anchor points.

6. The semantic repair method based on a hybrid architecture of DFA and a large language model according to claim 5, characterized in that, The normalization function for the standardized text sequence is: In the formula, The final output is a standardized text result. This is the original electrical text sequence. A benchmark database for electrical industry slang and professional vocabulary. For context-constrained parameters, For the set of candidate anchor points and preliminary matching of standard terminology mapping, For probabilistic decision functions based on large models with reasoning capabilities, The computation time for the fusion function; loading the terminology benchmark library and performing a millisecond-level full scan based on DFA to output the candidate anchor point set. It satisfies the linear time complexity model, and the specific formula is as follows: In the formula, For input text sequence Total length, For the benchmark library The number of terms that successfully match the input text, i.e., the set of candidate anchors. The number of elements, For the first The character length of each successfully matched term. The total character length of all matching terms.

7. The semantic repair method based on a hybrid architecture of DFA and a large language model according to claim 1, characterized in that, The step of integrating standardized terminology fragments to form a complete standardized text sequence corresponding to the original power text sequence further includes: [further details about the standardized text sequence]. Feedback is sent to subsequent business systems to provide accurate semantic input for actual business scenarios; a retrieval gain model is defined to evaluate the effect of normalization processing on knowledge activation, with the specific formula as follows: In the formula, To retrieve the gain coefficient, Original power text sequence Number of search hits in the RAG system For standardized text The number of search hits, For standardized text semantic feature vectors, This is the feature vector of the index entries in the technical terminology benchmark database. Let be the cosine distance between the two, with values ​​ranging from [0,1]. It is a positive gain index; if If the value is below the threshold, supplement the professional terminology benchmark library. The terms used, the size of the context window, or the size of the preset probability threshold.

8. A semantic repair device based on a hybrid architecture of DFA and a large language model, the device comprising: The receiving module is used to receive the raw power text sequence to be analyzed, load the technical terminology benchmark library, and perform a millisecond-level full scan based on DFA to output a set of candidate anchor points. The locking module is used to lock the candidate anchor set and extract the context feature window of each anchor in the candidate anchor set. Based on the large model with the ability of thought chain reasoning, the semantic coupling probability between each anchor and the context in the context feature window is calculated. The repair module is used to determine the repair path between each anchor point and the context in the context feature window based on the semantic coupling probability, and to perform semantic repair on each anchor point and the context based on the repair path to obtain the corresponding standardized term fragments; the integration module is used to integrate the standardized term fragments to form a complete standardized text sequence corresponding to the original power text sequence.

9. A readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1 to 7.

Images