Text polishing method for power transmission line crossing frame construction auxiliary generation scheme based on T5 pre-training model

Abstract

The invention discloses a T5 pre-training model-based power transmission line crossing frame construction auxiliary generation scheme text retouching method, which comprises the following steps of: constructing a high-quality parallel corpus, and performing multi-task fine tuning on a T5 pre-training model; adopting a term constraint decoding mechanism to carry out retouching processing on the output of the T5 pre-training model; performing standardization compliance inspection through a rule post-processing engine; and generating and outputting a standardized construction scheme. Compared with pure artificial retouching, the method has the advantages that the processing efficiency is remarkably improved, the labor cost is reduced, and compared with a single rule or single statistical method, the method gives consideration to both language fluency and format compliance while ensuring the accuracy, and is easy for engineering deployment and industry popularization; the method is superior to a traditional manual or general AI retouching scheme in the aspects of fidelity, specialty, normalization and auditing performance, unification of construction technology document language quality improvement and production efficiency optimization is achieved, and the method has high practicability and popularization value.

Description

TECHNICAL FIELD

[0001] The application relates to a text polishing method, in particular to a power transmission line crossing tower construction auxiliary generation scheme text polishing method based on a T5 pre-training model, and belongs to the technical field of natural language processing and intelligent processing of engineering texts. BACKGROUND

[0002] In recent years, with the rapid development of natural language processing technology, automatic text polishing has gradually become a key technology in intelligent writing assistance, automatic document processing and other scenarios. In the Chinese field, the text polishing task mainly includes grammar correction, semantic optimization, style unification, logic enhancement and terminology standardization and other sub-goals, especially in the construction scheme of power transmission line crossing tower, which puts forward higher requirements on language accuracy, terminology consistency and logical rigor. The development of generative artificial intelligence makes it possible to generate power crossing tower construction schemes with the help of large models, and it is necessary to polish the automatically generated construction scheme text content to make the generated scheme more feasible.

[0003] In the field of power construction, the related technology of crossing tower is also developing. For example, the system and method for erecting or dismantling a crossing tower disclosed in CN120582010A uses a modular telescopic design through telescopic wire carrying modules, power driving modules and closed-loop tension control systems, aiming to solve the problems of long construction time and complex process of traditional crossing tower construction, shortening the construction time from 3-5 days to several hours and improving the work efficiency. This scheme focuses on the physical improvement of mechanical structure and control system; for example, the crossing tower docking device and crossing tower disclosed in CN120728426A provide a crossing tower docking device, which tightens the net through the cooperation of the insertion mechanism and the access mechanism, aiming to reduce the working intensity and high-altitude operation risk of workers in the docking process of crossing tower construction. This invention also focuses on the optimization of hardware devices to improve the safety and convenience of the construction process; although the above-mentioned patents in the field of electric power have proposed improvement schemes in the physical structure, construction efficiency or safety of the crossing tower, they have not involved the quality control and optimization of the construction scheme text content itself. Looking back at the automatic text polishing technology itself, the mainstream technical route still has the following bottlenecks: 1. The method based on rules and templates, which relies on manually written grammar rules, regular expressions and terminology dictionaries. This method is stable in specific formats, but has poor generalization ability and is difficult to handle complex semantics or sentence restructuring; 2. The method based on statistics and traditional machine learning, which uses n-gram language models, maximum entropy models or CRF for local error detection and correction. This method relies on feature engineering and has limited ability to model the context, making it difficult to support paragraph-level logic polishing; 3. End-to-end methods based on pre-trained language models: With the rise of pre-trained models such as BERT, T5, and BART, researchers have begun to model the polishing task as a sequence-to-sequence (Seq2Seq) generation problem. For example, T5 is used to unify the task into a "text-to-text" format, achieving state-of-the-art results in the Chinese grammar correction (CGEC) task. In addition, large language models (LLMs) such as ChatGLM and Qwen have also been tried for polishing. Their advantages lie in strong generalization ability and contextual understanding, but they have problems such as high deployment cost, uncontrollable output, and the generation of factual illusions.

[0004] Despite the progress made in the aforementioned technologies, the following key bottlenecks still exist in the specific scenario of Chinese technical text polishing: 1. The problem of over-generation of large models and the illusion of fact: When polishing general large language models, they tend to be rewritten rather than refined, which can easily introduce technical details or incorrect parameters not mentioned in the original text. This may have serious consequences in engineering documents, resulting in output that lacks interpretability and makes it difficult to trace the source of errors. This does not meet the principle of minimum modification and maximum fidelity of industrial-grade documents. 2. General polishing models lack domain-specific expertise. Existing open-source polishing models are mostly trained on general corpora, which lacks sufficient understanding of professional terms in power, civil engineering, and machinery, such as crossing frames, wind deflection angles, and tension stringing, which are commonly used in power construction schemes. This can easily lead to misuse of terminology or replacement with colloquial expressions, which undermines the standardization of technical documents and lacks the ability to enforce industry standard formats. 3. It is difficult to balance the performance and controllability of lightweight models. Traditional Seq2Seq models, such as LSTM, perform poorly in long-distance dependencies and semantic coherence. Although some studies have attempted to fine-tune T5 or BART for Chinese text editing, most of them have not optimized the data construction and decoding for the characteristics of technical texts, resulting in grammatically correct but unprofessional editing results. 4. Purely data-driven methods ignore rule constraints, and the end-to-end model relies entirely on the distribution of training data. If the training set lacks specific terminology or format samples, it cannot generalize. Furthermore, it lacks integration with external knowledge, such as explicit interaction mechanisms with terminology databases and standard documents, making it difficult to guarantee that the output conforms to industry standards. Summary of the Invention

[0005] The purpose of this invention is to provide a text polishing method for auxiliary generation schemes of power transmission line crossing frame construction based on the T5 pre-trained model in order to solve at least one of the above-mentioned technical problems.

[0006] This invention achieves the above objective through the following technical solution: a text polishing method for auxiliary generation schemes of power transmission line crossing frame construction based on a T5 pre-trained model, which includes the following steps: S1. Construct a high-quality parallel corpus and fine-tune the T5 pre-trained model through multiple tasks; S2. Use a terminology constraint decoding mechanism to refine the model output; S3. Perform standardized compliance checks through a rule post-processing engine; S4. Generate and output standardized construction plans.

[0007] As a further aspect of the present invention: S1 specifically includes: S11. Construct a parallel corpus from the original construction plan for the crossing structure and the standard text reviewed and polished by experts, and preprocess the constructed parallel corpus. The preprocessing process of the parallel corpus adopts the following standardized procedure:

[0008] in, Represents the original text. This represents the standard text that has been polished by experts, where N is the size of the corpus. S12. A multi-task learning strategy is adopted to fine-tune the T5 pre-trained model. The fine-tuning process takes text polishing as the core main task, while introducing grammatical error detection and professional terminology recognition as two auxiliary tasks.

[0009] As a further aspect of the present invention: In S11, the types of schemes and common language expression paradigms covered by the parallel corpus include, but are not limited to, grammatical error correction, text simplification, style transfer, and fluency improvement.

[0010] As a further aspect of the present invention: In S12, the T5 pre-trained model adopts the Span Corruption pre-training objective, and achieves multi-task unification through task prefixes, specifically including: The T5 pre-trained model uses a pure Transformer Encoder-Decoder architecture for generative tasks; The T5 pre-trained model adds a task prefix, and all tasks are identified by the prefix. The prefix can be used as a regular token input to the encoder to achieve unification across multiple tasks. For the pre-training objective, the T5 pre-trained model randomly selects 15% of the tokens in the input text, groups these tokens into several consecutive "spans", replaces each span with a single sentinel token, and then the T5 pre-trained model generates the content of the masked spans and splices them together in order.

[0011] As a further aspect of the present invention: S2 specifically includes: S21. Use the T5 pre-trained model, which has been fine-tuned through multiple tasks, as the core generation engine, and introduce a term constraint decoding mechanism in the key decoding stage. The S22 and T5 pre-trained models have a built-in dictionary of professional terms related to the construction of bridging structures. , where each term All have been reviewed by domain experts and include standardized definitions, applicable contexts, and common collocation information. Candidate words are generated at each step of the T5 pre-trained model's decoding process. At that time, dynamically determine whether it belongs to the terminology dictionary. ; S23. The term constraint decoding mechanism, in addition to operating at the level of individual lexical units, also incorporates n-gram matching and sub-word alignment strategies to handle the complete recognition and preservation of compound terms.

[0012] As a further aspect of the present invention: in S22, the generated candidate words... When making dynamic judgments, if If, then retain its original generation probability; if If the probability is reduced exponentially, this constraint mechanism can be formally expressed as:

[0013] in, This indicates that the T5 pre-trained model, in a given context and input text The original word generation probability is as follows. An adjustable penalty coefficient is used to control the tendency to generate non-terminal vocabulary.

[0014] As a further aspect of this invention: In S23, the sub-word alignment strategy achieves seamless integration between terms and the model vocabulary, specifically including: During the initialization phase, the dictionary... Each term in Perform preprocessing with the T5 tokenizer to obtain its corresponding subword sequence. This forms a mapping table of "terms → sub-word sequences"; During the decoding process, a term matching state machine is maintained to guide the next decoding step. This means obtaining the term currently being generated, prioritizing the generation of the next sub-word of the current term, suppressing other sub-words unrelated to the current term, and ensuring that once a term starts generating, it is output completely without being interrupted or replaced midway. To avoid erroneous truncation in the middle of terms, a sub-word boundary awareness mechanism is introduced, that is, for each sub-word to be generated... If it belongs to a certain term If the middle subword is a term, check if the preceding subword is a predecessor of the term. If not, significantly reduce its generation probability to prevent "patchwork" mismatches. Conversely, if it is a term continuation state, increase the probability of the target subword.

[0015] As a further aspect of the present invention: S3 specifically includes: S31. The engine operates based on rule configuration files written in YAML format obtained from the rule base, which can be dynamically loaded and updated, to build a complete standardized processing system. S32, Each rule in the rule base Formal representation as a pair ,in, For pattern matching expressions, This corresponds to the standardized replacement operation; S33. A weighted priority-based hierarchical conflict resolution mechanism is adopted to handle multi-rule conflicts, and the weight coefficients are dynamically configured according to the rule type.

[0016] As a further aspect of this invention: the standardized processing system comprehensively covers punctuation mark standardization, unit of measurement unification and format conversion, numerical expression standardization, and compliance checks and adjustments for specific fixed sentence patterns; The hierarchical conflict resolution mechanism uses the following formula:

[0017] in This indicates the standardized operation that is ultimately performed; This represents a variable that uses operators commonly used in mathematical optimization to maximize the target expression. Represents the set of rules that match the input text; This refers to the rule weight coefficient; the rule weight coefficient is dynamically configured according to the rule type, and all correction operations are recorded in the structured audit log, which includes the rule identifier, text position, original content, correction result, and timestamp fields.

[0018] As a further aspect of the present invention: S4 specifically includes: S41. Output the polished construction plan for the crossing frame; S42. The solution enters the structured generation stage, injecting standardized document metadata through the template engine, including project number, document version, generation timestamp, and review status information; S43. Output standardized formats to meet the actual business needs of construction plan preparation, project acceptance reports, technical disclosure documents, and project archiving. S44. The final construction plan is verified by the quality assessment module. The quality assessment module employs multi-dimensional quantitative indicators, as detailed below: Terminology consistency coefficient:

[0019] in, This indicates an indicator function; it takes the value 1 if the term is in the standard library, and 0 otherwise. The terminology in the document is represented by T, where T represents the standard terminology library and k represents the total number of terms. Grammar accuracy:

[0020] in, Indicates the number of syntax errors. Indicates the total number of sentences; Formatting compliance:

[0021] Where m represents the total number of format checkpoints. Describe whether the j-th format checkpoint passed; The overall quality score uses a weighted fusion algorithm:

[0022] The weighting coefficients satisfy α+β+γ=1 and are dynamically adjusted according to the type of construction document for the gantry crossing (e.g., increasing the α value when emphasizing terminology consistency); the output includes three components: a standardized construction plan, a revised audit report, and a quality assessment certificate.

[0023] The beneficial effects of this invention are: 1) This invention can maintain the original technical content as the core objective without changing it as much as possible. It adopts the principle of "minimum necessary modification" in the T5 pre-trained model, so that the system only optimizes the syntax, expression and format, avoids adding or tampering with the original factual information to the greatest extent, improves the language standardization and readability of the document, and significantly improves the fidelity and reliability of the polishing. 2) This invention maintains the consistency of professional terminology and numerical units through linguistic constraints and rule verification, avoids referencing unverified technical information, introduces an industry terminology dictionary and dynamic constraint mechanism in the generation stage to ensure the accuracy of key terms, units of measurement and engineering expressions; at the same time, it combines the YAML rule engine to perform post-processing of punctuation, format, units, etc., to achieve a two-layer verification of "model generation + explicit rules" to ensure the consistency of terminology and format. 3) This invention meets different needs from strict fidelity to style optimization through configurable polishing intensity. While outputting the polishing results, it generates modification difference and audit logs, recording model modification points, rule triggers and replacement content, and realizing full traceability. It supports polishing intensity adjustment and manual review, which is convenient for flexible application in different scenarios. 4) Compared with purely manual polishing, this method significantly improves processing efficiency and reduces labor costs. Compared with single rule or single statistical methods, this method ensures accuracy while also taking into account language fluency and format compliance. It is easy to deploy in engineering and promote in the industry. In terms of fidelity, professionalism, standardization and auditability, it is superior to traditional manual or general AI polishing solutions. It achieves the unity of improving the language quality of construction technical documents and optimizing production efficiency, and has strong practicality and promotion value. 5) This invention adopts a modular and containerized architecture design with a moderate model size; it supports dynamic updates and version control of dictionaries and rules, which facilitates rapid integration and continuous optimization in enterprise production environments, and improves the efficiency and standardization of engineering document processing. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Examples, such as Figure 1 As shown, this embodiment provides a text polishing method for auxiliary generation schemes of power transmission line crossing structure construction based on a T5 pre-trained model. This text polishing method constructs a language intelligent polishing system for crossing structure construction text by deeply integrating the T5 pre-trained model, domain-specific dictionaries, and configurable rule system. This language intelligent polishing system achieves semantic understanding and grammatical error correction based on the multi-task fine-tuned T5 pre-trained model, combines a built-in terminology constraint decoding mechanism to ensure the accuracy and consistency of professional vocabulary usage, and relies on a dynamically loadable rule configuration engine to automatically standardize text structure, punctuation marks, and units of measurement. The aim is to improve the language standardization and professionalism of crossing structure construction schemes and achieve high-quality, automated polishing of non-standardized technical texts.

[0027] This text editing method specifically includes the following steps: First, the construction of a high-quality parallel corpus and the fine-tuning of a multi-task model.

[0028] 1) A large-scale, high-quality parallel corpus was constructed from a massive amount of original construction plans for cross-passage structures and standard texts that had been reviewed and polished by experts. This parallel corpus covers various scheme types and common language expression paradigms, specifically including grammatical error correction, text simplification, style transfer, and fluency improvement. The preprocessing of the parallel corpus adopts the following standardized workflow:

[0029] in Represents the original text. This represents the standard text that has been reviewed and polished by experts, where N is the size of the corpus.

[0030] 2) Based on this, a multi-task learning strategy is adopted to fine-tune the T5 pre-trained model. The fine-tuning process takes text polishing as the core main task, while introducing grammatical error detection and professional terminology recognition as two auxiliary tasks.

[0031] 21) The T5 pre-trained model adopts a pure Transformer Encoder-Decoder architecture, which is suitable for generative tasks; 22) The T5 pre-trained model adds a task prefix. All tasks are identified by the prefix, which can be used as a regular token input to the encoder, thus achieving unification across multiple tasks. 23) For the pre-training objective, the T5 pre-trained model proposes Span Corruption, which involves randomly selecting 15% of the tokens in the input text, grouping these tokens into several consecutive "spans", replacing each span with a single sentinel token, and then the model needs to generate the content of the masked spans and concatenate them in order. Essentially, it is a denoising autoencoder that encourages the model to understand the context and reconstruct missing segments. It is naturally adapted to the encoder-decoder architecture, and the target of generation is continuous text, which is beneficial to downstream generation tasks. These designs significantly enhance the model's deep perception of the linguistic characteristics of technical texts and its generalization performance, enabling it not only to correct grammatical and spelling errors, but also to better understand the technical context and improve semantic coherence and professionalism.

[0032] It should be noted that the T5 pre-trained model (Text-To-Text Transfer Transformer) is a unified text generation framework proposed by Google Research in 2019. Its core idea is to model all NLP tasks in a unified "text-to-text" format.

[0033] Second, terminology constraint decoding and model polishing.

[0034] To improve the quality and professionalism of automatic editing of technical texts, a T5 pre-trained model with multi-task fine-tuning is used as the core generation engine. An innovative terminology-constrained decoding (TCD) mechanism is introduced in its key decoding stage. This mechanism aims to strictly ensure the accuracy, standardization and consistency of the use of professional terms while maintaining the fluency and semantic coherence of the language.

[0035] Specifically: 1) The system (intelligent language polishing system) has a built-in structured, high-quality dictionary of professional terms in the field of bridge construction. , where each term All have been reviewed by domain experts and include standardized definitions, applicable contexts, and common collocation information. Candidate words are generated at each step of the T5 pre-trained model's decoding process. At that time, the system dynamically determines whether it belongs to the terminology dictionary. :like If, then retain its original generation probability; if If the probability is reduced exponentially, this constraint mechanism can be formally expressed as:

[0036] in, This indicates that the T5 pre-trained model, in a given context and input text The original word generation probability is as follows. An adjustable penalty coefficient is used to control the tendency to generate non-terminal words, which can be achieved through reasonable settings. The system can significantly suppress the use of non-standard, colloquial, or potentially erroneous terminology while preserving the naturalness of the language. This method effectively ensures the standardized use of professional terminology.

[0037] 2) The term constraint decoding mechanism not only operates at the level of a single word, but also combines n-gram matching and sub-word alignment strategies to handle the complete recognition and preservation of compound terms; that is, when the T5 pre-trained model detects multi-word expressions in the term dictionary during the generation process, the system will temporarily lock several subsequent decoding steps to force the completion of the entire term output, thereby avoiding the term being truncated or mis-split. 3) Because the T5 pre-trained model uses Sentence Piece as its underlying word segmenter, its vocabulary consists of sub-words, while terminology dictionaries are typically stored as complete words or phrases. This inconsistency in granularity leads to a sub-word alignment strategy designed to seamlessly integrate terminology with the model's vocabulary. Specifically, during system initialization, the dictionary... Each term in Perform preprocessing with the T5 tokenizer to obtain its corresponding subword sequence. This creates a mapping table of "terms → subword sequences." During decoding, the system maintains a terminology state machine to guide the next decoding step. This involves acquiring the currently generated term, prioritizing the generation of the next subword of the current term, and suppressing other subwords unrelated to the current term. This ensures that once term generation begins, it is output completely without interruption or replacement. Furthermore, to avoid erroneous truncation in the middle of terms, a subword boundary awareness mechanism is introduced. This mechanism is used for each subword to be generated. If it belongs to a certain term If the middle sub-word (neither the first nor the last) is a term, check if the preceding sub-word is a predecessor of the term. If not, significantly reduce its generation probability to prevent "patchwork" mismatches. Conversely, if it is in a term continuation state, increase the probability of the target sub-word.

[0038] In summary, this terminology constraint decoding mechanism not only significantly enhances the text polishing capabilities of the T5 pre-trained model in highly specialized scenarios such as scaffolding construction, but also provides a transferable technical paradigm for applications in other vertical fields—that is, while maintaining the ability to generate general language, it enables the accurate expression and standardized output of professional content through lightweight, pluggable domain knowledge constraint modules.

[0039] Third, rule post-processing and standardized compliance checks.

[0040] The text, after being decoded based on the aforementioned terminology constraints, is further refined through a highly configurable rule post-processing engine to ensure that the text fully complies with the specifications of the cross-passage construction technical document.

[0041] The engine operates based on a rule configuration file written in YAML format that can be dynamically loaded and updated. It has built a complete standardized processing system that comprehensively covers punctuation standardization (such as the use of parentheses and comma separation), unit of measurement standardization and format conversion (such as standardizing "mpa" to "MPa" and "kilometer" to "km"), standardization of numerical expression (such as date format and decimal places), and compliance checks and adjustments for specific fixed sentence structures.

[0042] Each rule in the rule base It can be formally represented as a pair of tuples ,in, For pattern matching expressions, This corresponds to the standardized replacement operation.

[0043] For example, the standardization rules for pressure units can be expressed as follows:

[0044] The unified rule for date format is:

[0045] Matching mode Used to identify date strings separated by forward slashes, where Representing a four-digit year, two These represent the month and day, respectively, using one or two digits; parentheses are used to capture groups for reference during replacement. The corresponding replacement expression... By referencing the aforementioned three groups, the dates that were originally separated by slashes are reorganized into a standard format connected by hyphens, that is, "YYYY / MM / DD" is converted into "YYYY-MM-DD".

[0046] Rule application is achieved through cascading processing:

[0047] Where Input is the input text; Indicates the first The content matched by this rule; Indicates the first The rule replacement operation; Output is the final output after the rule chain is processed.

[0048] For scenarios involving multiple rule conflicts, the system adopts a hierarchical conflict resolution mechanism based on weighted priority:

[0049] in It is the standardized operation that is ultimately executed. This is a commonly used operator in mathematical optimization; its function is to select the variable that maximizes the objective expression. This represents the set of rules that match the input text. This is the rule weight coefficient, which is dynamically configured according to the rule type (w=0.6 for syntax rules, w=0.3 for format rules, and w=0.1 for semantic rules).

[0050] All corrective actions are recorded in a structured audit log, which includes fields such as rule identifier, text location, original content, corrective result, and timestamp, ensuring traceability and verifiability throughout the entire process and providing a data foundation for subsequent quality assessments.

[0051] This step further enhances the standardization and overall normalization of the text in terms of structure and format.

[0052] Fourth, standardized construction plan generation and output.

[0053] After the above series of processing steps, the final output is a polished construction plan for the scaffolding, characterized by fluent and standardized language, accurate terminology, and a clear and rigorous structure. This plan then enters the structured generation stage, where the system injects standardized document metadata through a template engine, including key information such as project number, document version, generation timestamp, and review status. A standardized format is output to meet the needs of practical business scenarios such as construction plan preparation, project acceptance reports, technical disclosure documents, and project archiving. The final output construction plan must pass the quality assessment module's review, which uses multi-dimensional quantitative indicators as follows.

[0054] Terminology consistency coefficient:

[0055] in, This is an indicator function; it takes the value 1 if the term is in the standard library, and 0 otherwise. This represents the terms in the document, where T is the standard terminology library and k is the total number of terms.

[0056] Grammar accuracy:

[0057] in, The number of syntax errors. This represents the total number of sentences.

[0058] Formatting compliance:

[0059] Where m is the total number of format checkpoints. Describe whether the j-th format checkpoint passed.

[0060] The overall quality score uses a weighted fusion algorithm:

[0061] The weighting coefficients satisfy α+β+γ=1 and are dynamically adjusted based on the type of construction document for the scaffolding (e.g., increasing the α value when emphasizing terminology consistency). The system output includes three components: a standardized construction plan, a revised audit report, and a quality assessment certificate.

[0062] The output documents fully comply with enterprise technical document management standards and industry standards such as the "Specifications for the Compilation of Construction Documents for Power Engineering" (DL / T5191-2004). They can be directly applied to scenarios such as project filing, technical briefings, and project acceptance for crossing bridge construction. Together with the aforementioned corpus construction, model fine-tuning, and terminology constraint decoding, they constitute a complete automated technical document processing solution, significantly improving the writing efficiency, quality level, and management effectiveness of crossing bridge construction technical documents. The final output construction plan must pass the quality assessment module's verification.

[0063] This embodiment can achieve the following: 1) To address the issues of "over-generation" and factual illusion in large models, this implementation uses a T5-base Chinese pre-trained model, focusing on "minimum necessary modifications" rather than free creation. It uses high-quality parallel corpora for supervised fine-tuning, guiding the model to learn "fidelity-preserving embellishment" behavior, avoiding fabrication, ensuring that the model output is strictly limited to the distribution of training data, without introducing external knowledge illusions, and ensuring the consistency of technical facts. 2) To address the issue of general models lacking domain-specific expertise, the training data in this embodiment is derived from real engineering technical documents, such as construction reports, design specifications, and acceptance records. These documents naturally contain professional terms such as "crossing frame," "load-bearing cable," and "wind deflection angle." During the fine-tuning stage, the usage patterns of these terms are explicitly modeled, enabling the model to internalize domain-specific language standards. At the same time, during the polishing process, an industry terminology dictionary is dynamically loaded to enhance the identification and retention capabilities of key terms. 3) To address the challenge of balancing performance and controllability in lightweight models, this implementation uses the T5 pre-trained model with a Seq2Seq architecture, which combines strong contextual modeling capabilities with controllable generation characteristics. By clarifying instructions through task prefixes, the model's understanding of the polishing intent is enhanced. The fine-tuning targets focus on the three aspects of "standard language + clear logic + accurate terminology" to avoid generalization to irrelevant content. 4) To address the issue of pure data-driven models ignoring rule constraints, a terminology constraint decoding mechanism is introduced to enforce consistency of key terms. At the same time, a rule filtering module is designed to perform format verification on the model output, thereby achieving rule-based protection against high-frequency hard specifications and enabling the model to handle complex semantic optimization. This ensures that the output text not only conforms to language fluency but also strictly follows industry formats and terminology standards, thus compensating for the "soft constraint" defects of pure deep learning models.

[0064] During implementation, internal construction plans, design specifications, and expert review texts from enterprises were first collected and anonymized to build a high-quality parallel corpus. The corpus was standardized in terms of punctuation, units, and numerical formats, and segmented or coded into words for subsequent processing. The training phase adopted a strategy of minimizing necessary modifications, prioritizing the preservation of original facts and making only necessary adjustments to grammar, expression, and format. Auxiliary tasks such as terminology recognition and format discrimination were incorporated to enhance the system's perception of engineering text characteristics. In the generation phase, a terminology constraint mechanism and dynamic vocabulary were introduced, prioritizing or forcibly retaining key terms. After output, a YAML-driven rule engine performed hard corrections on punctuation, units, and templates. The system simultaneously outputs modification differences and audit logs, recording the triggering rules and replacement details for each modification, supporting manual rollback and the reverting of manual modifications to subsequent optimization data.

[0065] This solution significantly improves the efficiency and standardization of technical documentation, possesses excellent scalability and engineering applicability, and can be widely applied to the generation and quality improvement of cross-passage construction schemes in engineering fields such as power construction and transportation construction. It effectively supports the automatic production and knowledge management of high-standard engineering documents.

[0066] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0067] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for refining text related to auxiliary generation schemes for power transmission line crossing frame construction based on a T5 pre-trained model, characterized in that... The method for refining the text of the auxiliary generation scheme for the construction of power transmission line crossing frames includes the following steps: S1. Construct a high-quality parallel corpus and fine-tune the T5 pre-trained model through multiple tasks; S2. Use a terminology constraint decoding mechanism to refine the output of the T5 pre-trained model; S3. Perform standardized compliance checks through a rule post-processing engine; S4. Generate and output standardized construction plans.

2. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 1, characterized in that, S1 specifically includes: S11. Construct a parallel corpus from the original construction plan for the crossing structure and the standard text reviewed and polished by experts, and preprocess the constructed parallel corpus. The preprocessing process of the parallel corpus adopts the following standardized procedure: ; in, Represents the original text. N represents the standard text that has been reviewed and polished by experts, where N is the size of the corpus. S12. A multi-task learning strategy is adopted to fine-tune the T5 pre-trained model. The fine-tuning process takes text polishing as the core main task, while introducing grammatical error detection and professional terminology recognition as two auxiliary tasks.

3. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 2, characterized in that: In S11, the parallel corpus covers scheme types and common language expression paradigms, including but not limited to grammatical error correction, text simplification, style transfer, and fluency improvement.

4. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 3, characterized in that, In S12, the T5 pre-trained model adopts the Span Corruption pre-training objective and achieves multi-task unification through task prefixes, specifically including: The T5 pre-trained model uses a pure Transformer Encoder-Decoder architecture for generative tasks; The T5 pre-trained model adds a task prefix, and all tasks are identified by the prefix. The prefix can be used as a regular token input to the encoder to achieve unification across multiple tasks. For the pre-training objective, the T5 pre-trained model randomly selects 15% of the tokens in the input text, groups the tokens into several consecutive "spans", replaces each span with a single sentinel token, and then the T5 pre-trained model generates the content of the masked spans and splices them together in order.

5. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 1, characterized in that, S2 specifically includes: S21. Use the T5 pre-trained model, which has been fine-tuned through multiple tasks, as the core generation engine, and introduce a term constraint decoding mechanism in the key decoding stage. The S22 and T5 pre-trained models have a built-in dictionary of professional terms related to the construction of bridging structures. , where each term All have been reviewed by domain experts and include standardized definitions, applicable contexts, and common collocation information. Candidate words are generated at each step of the T5 pre-trained model's decoding process. At that time, dynamically determine whether it belongs to the terminology dictionary. ; S23. The term constraint decoding mechanism, in addition to operating at the level of individual lexical units, also incorporates n-gram matching and sub-word alignment strategies to handle the complete recognition and preservation of compound terms.

6. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 5, characterized in that, In step S22, the generated candidate words When making dynamic judgments, if If, then retain its original generation probability; if If the probability is then exponentially decayed, the constraint mechanism is formally expressed as: ; in, This indicates that the T5 pre-trained model, in a given context and input text The original word generation probability is as follows. An adjustable penalty coefficient is used to control the tendency to generate non-terminal vocabulary.

7. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 5, characterized in that, In step S23, the sub-word alignment strategy achieves seamless integration between terms and the model vocabulary, specifically including: During the initialization phase, the dictionary... Each term in Perform preprocessing with the T5 tokenizer to obtain its corresponding subword sequence. This forms a mapping table of "terms → sub-word sequences"; During the decoding process, a term matching state machine is maintained to guide the next decoding step. This involves obtaining the term currently being generated, prioritizing the generation of the next sub-word of the current term, suppressing other sub-words unrelated to the current term, and ensuring that once a term starts generating, it is output completely without being interrupted or replaced. To avoid erroneous truncation in the middle of terms, a sub-word boundary awareness mechanism is introduced, that is, for each sub-word to be generated... If it belongs to a certain term If the middle subword is a term, check if the preceding subword is a predecessor of the term. If not, significantly reduce its generation probability to prevent "patchwork" mismatches. Conversely, if it is a term continuation state, increase the probability of the target subword.

8. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 1, characterized in that, The S3, which performs standardized compliance checks through a rule post-processing engine, specifically includes: S31. The engine operates based on rule configuration files written in YAML format obtained from the rule base, which can be dynamically loaded and updated, to build a complete standardized processing system. S32, Each rule in the rule base Formal representation as a pair ,in, For pattern matching expressions, This corresponds to the standardized replacement operation; S33. A weighted priority-based hierarchical conflict resolution mechanism is adopted to handle multi-rule conflicts, and the weight coefficients are dynamically configured according to the rule type.

9. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 8, characterized in that: The standardized processing system comprehensively covers punctuation mark standardization, unit of measurement unification and format conversion, numerical expression standardization, and compliance checks and adjustments for specific fixed sentence structures; The hierarchical conflict resolution mechanism adopts the following formula: ; in This indicates the standardized operation that is ultimately performed; This represents a variable that uses commonly used operators in mathematical optimization to maximize the target expression. Represents the set of rules that match the input text; This refers to the rule weight coefficient; the rule weight coefficient is dynamically configured according to the rule type, and all correction operations are recorded in the structured audit log, which includes the rule identifier, text position, original content, correction result, and timestamp fields.

10. The method for refining the text of the auxiliary generation scheme for power transmission line crossing frame construction according to claim 1, characterized in that, In step S4, generating and outputting a standardized construction plan specifically includes: S41. Output the polished construction plan for the crossing frame; S42. The construction plan for the gantry crossing has entered the structured generation stage. Standardized document metadata, including project number, document version, generation timestamp, and review status information, is injected through the template engine. S43. Output standardized formats to meet the actual business needs of construction plan preparation, project acceptance reports, technical disclosure documents, and project archiving. S44. The final output of the gantry construction plan passes the quality assessment module's inspection. The quality assessment module employs multi-dimensional quantitative indicators, as detailed below: Terminology consistency coefficient: ; in, This indicates an indicator function; it takes the value 1 if the term is in the standard library, and 0 otherwise. The terminology in the document is represented by T, where T represents the standard terminology library and k represents the total number of terms. Grammar accuracy: ; in, Indicates the number of syntax errors. Indicates the total number of sentences; Formatting compliance: ; Where m represents the total number of format checkpoints. Describe whether the j-th format checkpoint passes; The overall quality score uses a weighted fusion algorithm: ; The weighting coefficients satisfy α+β+γ=1 and are dynamically adjusted according to the type of construction document for the gantry crossing. The output consists of three parts: a standardized construction plan, a revised audit report, and a quality assessment certificate.

Images