Professional process enhancement generation method, electronic equipment and computer program product

Abstract

The invention provides a professional process enhancement generation method, electronic equipment and a computer program product. The professional process enhancement generation method comprises the following steps: acquiring associated context information of a paragraph to be generated from a memory according to reported state information of the paragraph to be generated; constructing a cue word of a paragraph to be generated, wherein the cue word comprises the associated context information; inputting the cue word into a large language model, and enabling the large language model to generate the content of the paragraph to be generated according to the associated context information; and extracting context information according to the content of the paragraph to be generated, and updating the context information into the memory. According to the method, the stability of the hierarchical structure of the paragraph of the generated content can be ensured, the style consistency and the logic coherence in long text generation are improved, and the quality of a vertical professional field report generated by a large language model is improved.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and natural language processing technology, and in particular to a method for enhancing professional processes, an electronic device, and a computer program product. Background Technology

[0002] In certain specialized fields, such as chemicals, energy, and steel, companies are required to submit numerous well-structured and detailed technical reports, such as carbon emission assessment reports, during the application, verification, and operational phases. These reports typically require the use of standardized methodologies, descriptions of complete business processes, and integration of multi-source heterogeneous data. They also necessitate expert writing and review according to industry standards, heavily relying on manual processes, including manually reviewing project materials, selecting assessment boundaries and emission factors, and drafting standardized expressions. This process is time-consuming, costly, involves repetitive sections, and relies heavily on expert experience, making it difficult to guarantee efficiency and consistency. With the development of large language models, using them to generate these technical reports to reduce reliance on manual labor, improve report writing efficiency, and enhance report quality has garnered increasing attention.

[0003] However, the reliability of current general-purpose large language models in vertical professional fields is still difficult to guarantee. Summary of the Invention

[0004] This disclosure provides a professional process enhancement generation method, an electronic device, and a computer program product.

[0005] In a first aspect, embodiments of this disclosure provide a professional process enhancement generation method, comprising: obtaining the associated context information of the paragraph to be generated from memory based on the status information of the paragraph to be generated in the report; constructing prompt words for the paragraph to be generated, the prompt words containing the associated context information; inputting the prompt words into a large language model, causing the large language model to generate the content of the paragraph to be generated based on the associated context information; extracting context information based on the content of the paragraph to be generated, and updating it in the memory.

[0006] In some embodiments, the associated context information includes a knowledge graph and the content of already generated paragraphs related to the paragraph to be generated; obtaining the associated context information of the paragraph to be generated from memory based on the reported state information of the paragraph to be generated includes: obtaining the knowledge graph from memory based on the state information of the paragraph to be generated, wherein the knowledge graph contains at least one of terms and data; and obtaining the content of already generated paragraphs related to the paragraph to be generated from memory based on the state information of the paragraph to be generated.

[0007] In some embodiments, extracting context information based on the content of the paragraph to be generated and updating it in the memory includes: extracting update information of the knowledge graph based on the content of the paragraph to be generated, and updating the knowledge graph in the memory based on the update information; vectorizing the content of the paragraph to be generated and storing it in the memory.

[0008] In some embodiments, the memory includes external memory, and updating the knowledge graph in the memory according to the update information includes: for newly detected terms or data variables, creating graph nodes and labeling initial metadata; for terms that already exist in the knowledge graph, updating the occurrence frequency and citation records of the terms.

[0009] In some embodiments, the prompt word includes a phased execution state block and a semantic memory injection block; constructing the prompt word for the paragraph to be generated includes: updating the knowledge graph to the phased execution state block; and updating the content of the generated paragraphs related to the paragraph to be generated to the semantic memory injection block.

[0010] In some embodiments, the phased execution status block includes at least one of the following fields: a paragraph information field for storing information about the paragraph to be generated; a terminology field for storing used terms; a data field for storing data; and a previous paragraph summary field for storing a summary of the previous paragraph.

[0011] In some embodiments, the prompt words further include a style structure static instruction block; constructing the prompt words for the paragraph to be generated includes: extracting style-leading prompt content with identity semantics from the template report; and constructing the style structure static instruction block based on the style-leading prompt content.

[0012] In some embodiments, extracting style-leading prompts with identity semantics from a template report includes: identifying the structure of the template report and obtaining structure tags; quantitatively evaluating the writing style of the template report and determining the writing style category; and extracting contextual terms from the template report.

[0013] In some embodiments, the style structure static instruction block includes at least one of the following fields: an identity definition field for defining the role identity; an input material field for defining the structure and type of the input material; a structure field for defining chapter structure requirements; a writing style field for defining writing style requirements; and an output structure field for defining output format requirements.

[0014] In some embodiments, the professional process enhancement generation method further includes: performing structuring and semantic processing on the input materials.

[0015] In some embodiments, the input material is processed in a structured and semantic manner, including: structuring the non-text elements in the input material; and translating the terms in the input material.

[0016] In some embodiments, the non-text elements include at least one of tables, images, and formulas; structuring the non-text elements in the input material includes: extracting table regions from the input material, identifying data in the tables, and converting them into key-value pairs as a structured representation of the tables; and / or identifying images from the input material, extracting node text from the images, and converting it into a graph structure as a structured representation of the images; and / or extracting formulas from the input material, and converting the formulas into expression tuples as a structured representation of the formulas.

[0017] In some embodiments, the professional process enhancement generation method further includes: using an evaluation sub-model to assess the consistency and professionalism of the content of the generated paragraphs.

[0018] In some embodiments, the consistency and professionalism of the generated paragraphs are evaluated using an evaluation sub-model, including: determining an evaluation standard system, which includes at least one of the following evaluation dimensions: structural consistency, material consistency, standard compliance, common sense, professionalism, and readability; constructing training data based on the evaluation standard system, wherein the training data includes positive and negative examples corresponding to each evaluation dimension; training and fine-tuning an initial evaluation model using the training data to obtain the evaluation sub-model; and using the evaluation sub-model to evaluate the consistency and professionalism of the generated paragraphs.

[0019] In some embodiments, the report includes any one of a technical report, an industry report, a carbon emission assessment report, an energy assessment report, or an environmental assessment report.

[0020] In a second aspect, embodiments of this disclosure provide an electronic device, including: one or more processors; and a memory storing one or more programs thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the professional process enhancement generation method described in the first aspect of this disclosure.

[0021] Thirdly, embodiments of this disclosure provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the professional process enhancement generation method described in the first aspect of this disclosure.

[0022] The professional process enhancement generation method provided in this disclosure, when generating reports using a large language model, retrieves the associated context information of the paragraph to be generated from memory, dynamically updates prompt words based on the associated context information, and updates the memory after the content of the paragraph to be generated is generated. This ensures the stability of the paragraph hierarchy structure of the generated paragraph, improves the style consistency and logical coherence in long text generation, and avoids problems such as style drift, poor paragraph connection, and logical jumps in the multi-paragraph structure of long text. As a result, it can generate reports in vertical professional fields using a large language model, thereby improving the quality of vertical professional field reports generated by the large language model. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a professional process enhancement generation method according to an embodiment of this disclosure.

[0024] Figure 2 This is a schematic diagram of the structure of an electronic device according to an embodiment of this disclosure.

[0025] Figure 3 This is a schematic diagram of a professional process enhancement generation system architecture in an embodiment of this disclosure.

[0026] Figure 4 This is a schematic diagram of a control generation module architecture in an embodiment of this disclosure. Detailed Implementation

[0027] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of this disclosure will be described in detail below with reference to the accompanying drawings.

[0028] Exemplary embodiments will be described more fully below with reference to the accompanying drawings; however, these exemplary embodiments may be embodied in different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will enable those skilled in the art to fully understand the scope of this disclosure.

[0029] Where there is no conflict, the various embodiments of this disclosure and the features thereof in the embodiments may be combined with each other.

[0030] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0031] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. As used herein, the singular forms “a” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded.

[0032] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and this disclosure, and will not be interpreted as having an idealized or overly formal meaning, unless expressly so defined herein.

[0033] The inventors' research revealed in this disclosure indicates that general-purpose large language models are still difficult to directly apply to scenarios involving "vertical domains + long texts + high specialization." The reasons are mainly threefold: First, general-purpose large models lack a deep understanding of industry terminology, processes, and compliance rules; second, the long text generation process lacks consistency and contextual memory mechanisms, easily leading to logical jumps or style drift; and third, existing methods generally lack automated quality assessment and feedback systems, making it difficult to guarantee the reliability of the generated content.

[0034] Specifically, the general language model has the following shortcomings when generating reports in vertical professional fields: (1) Insufficient professional field knowledge: Most general language models are pre-trained or fine-tuned in general fields, lacking in-depth modeling of terminology and process specifications in vertical industries such as energy and environment, which may lead to misuse of terminology, inconsistent processes or logical confusion when generating professional reports; (2) Long text consistency problem: General language models generally perform well when generating short texts, but when generating long and complex reports, they are prone to context loss, logical jumps and inconsistent styles between generated content, affecting the quality and readability of the report; (3) Poor multimodal data integration capability: General language models cannot effectively process and integrate multiple types of data (such as tables, pictures, formulas, etc.), and when generating reports containing complex data, they often cannot achieve seamless data connection, resulting in a lack of professionalism and accuracy in the generated reports; (4) Lack of automated quality assessment mechanism: General language models are not capable of automated quality assessment and feedback, especially in compliance checks, terminology consistency and content accuracy, still relying on a large amount of manual review.

[0035] The embodiments disclosed herein aim to provide a systematic method that combines large model generation capabilities with professional process knowledge, connecting the entire process of material processing, style learning, controllable generation, and automated evaluation, thereby achieving high-quality intelligent generation and reliability assurance of serious long texts in vertical professional scenarios.

[0036] In view of this, the present disclosure provides an Expert Process Augmented Generation (EPAG) method, which aims to form a closed loop of material processing, style extraction, controllable generation and automatic evaluation to create a serious long text artificial intelligence (AI) assisted generation system that is feasible, reliable and industry-adaptable, based on the insufficient capabilities of general large language models.

[0037] In a first aspect, embodiments of this disclosure provide a professional process enhancement generation method, referring to... Figure 1 ,include: S1. Obtain the associated context information of the paragraph to be generated from memory based on the status information of the paragraph to be generated in the report; S2. Construct a prompt for the paragraph to be generated, wherein the prompt contains the associated context information; S3. Input the prompt words into the large language model, so that the large language model generates the content of the paragraph to be generated based on the associated context information; S4. Extract context information based on the content of the paragraph to be generated and update it in the memory.

[0038] In this embodiment of the disclosure, the report to be generated refers to a report in a vertical professional field, such as a professional report in the fields of energy and environment. This embodiment of the disclosure does not impose any special limitations on this.

[0039] In this embodiment of the disclosure, a paragraph generally refers to a relatively complete textual unit in a report that revolves around a single theme and has a logically coherent structure. In this embodiment, a paragraph can encompass various textual components with hierarchical structures, including but not limited to sections, chapters, and parts, and is not limited to explicit divisions by numbering, headings, or formatting. That is, in this embodiment, a paragraph can be a natural segment in language or text structure, or it can refer to a logical module of content organization. In some embodiments, a paragraph refers to a basic unit that possesses content integrity, can independently convey specific information, and is identifiable and referable within the overall textual structure of the report.

[0040] In this embodiment of the disclosure, the associated context information of the paragraph to be generated includes information about the generated paragraphs related to the paragraph to be generated, the terms and data involved in the paragraph to be generated, and so on.

[0041] In this embodiment of the disclosure, after generating the content of the paragraph to be generated, the context information is extracted and updated in memory, which may become the associated context information when the paragraph is generated later.

[0042] In the embodiments of this disclosure, the memory can be external memory, internal memory, or a combination of external and internal memory. This disclosure does not impose any special limitations on this aspect.

[0043] In this embodiment of the disclosure, a dynamic updating and recall mechanism for contextual information is established during the report generation process through steps S1 to S4. Specifically, when generating a new paragraph in the report, the associated contextual information related to the new paragraph is recalled from memory; after the new paragraph is generated, its contextual information is extracted and updated in memory. This ensures that when generating any paragraph, the prompts input into the large language model contain associated contextual information. With the prompts of associated contextual information, the paragraph hierarchy structure of newly generated paragraphs and already generated paragraphs can be stabilized, improving style consistency and logical coherence in long text generation. This avoids problems such as style drift, poor paragraph connection, and logical jumps in multi-paragraph structures of long texts, thereby enabling the generation of reports in vertical professional fields using the large language model and improving the quality of such reports.

[0044] In some embodiments, the associated context information includes a knowledge graph and the content of already generated paragraphs related to the paragraph to be generated; obtaining the associated context information of the paragraph to be generated from memory based on the reported state information of the paragraph to be generated includes: obtaining the knowledge graph from memory based on the state information of the paragraph to be generated, wherein the knowledge graph contains at least one of terms and data; and obtaining the content of already generated paragraphs related to the paragraph to be generated from memory based on the state information of the paragraph to be generated.

[0045] In this embodiment of the disclosure, the content of the already generated paragraph related to the paragraph to be generated can be the full text or a fragment of the already generated paragraph, or it can be a summary of the already generated paragraph. This embodiment of the disclosure does not impose any special limitations on this.

[0046] In some embodiments, the knowledge graph and the content of already generated paragraphs related to the paragraph to be generated are stored in memory in the form of semantic vectors. When retrieving the associated context information from memory, the state information of the paragraph to be generated is first encoded to obtain a semantic vector, and then the semantic vectors are compared to obtain the associated context information related to the paragraph to be generated.

[0047] In some embodiments, extracting context information based on the content of the paragraph to be generated and updating it in the memory includes: extracting update information of the knowledge graph based on the content of the paragraph to be generated, and updating the knowledge graph in the memory based on the update information; vectorizing the content of the paragraph to be generated and storing it in the memory.

[0048] In this embodiment, the knowledge graph can be stored in internal memory, in external memory, or a portion of the knowledge graph can be stored in internal memory while another portion is stored in external memory. This embodiment does not impose any particular limitations on this.

[0049] In some embodiments, the memory includes external memory, and updating the knowledge graph in the memory according to the update information includes: for newly detected terms or data variables, creating graph nodes and labeling initial metadata; for terms that already exist in the knowledge graph, updating the occurrence frequency and citation records of the terms.

[0050] In embodiments of this disclosure, when the memory includes external memory, it may also include internal memory or not. This disclosure does not impose any particular limitation on this. In some embodiments, when the memory includes both external and internal memory, a portion of the terms and data may be stored in internal memory, while another portion may be stored in external memory.

[0051] In some embodiments, the prompt word includes a phased execution state block and a semantic memory injection block; constructing the prompt word for the paragraph to be generated includes: updating the knowledge graph to the phased execution state block; and updating the content of the generated paragraphs related to the paragraph to be generated to the semantic memory injection block.

[0052] In some embodiments, the terms and data in the knowledge graph are stored in a phased execution state block. If the phased execution state block cannot hold all the terms and data, the excess terms and data are updated and stored in external memory.

[0053] In some embodiments, the phased execution status block includes at least one of the following fields: a paragraph information field for storing information about the paragraph to be generated; a terminology field for storing used terms; a data field for storing data; and a previous paragraph summary field for storing a summary of the previous paragraph.

[0054] Below is an example of a phased execution state block: Current segment number: Segment 4 (benchmarking analysis); Previous paragraph summary: The previous paragraph analyzed the total regional emissions and their source structure; Current cumulative terms: CO2 emissions, denitrification; Current cumulative data: Annual power generation: 7.83 billion kWh, unit coal consumption: 273 g / kWh, carbon content: 0.67 t / t.

[0055] Below is an example of a semantic memory injection block: Recall Item 1: Greenhouse Gas Compliance Assessment; Recall Item 2: Introduction to Greenhouse Gas Measurement Standards.

[0056] In some embodiments, the prompt words further include a style structure static instruction block; constructing the prompt words for the paragraph to be generated includes: extracting style-leading prompt content with identity semantics from the template report; and constructing the style structure static instruction block based on the style-leading prompt content.

[0057] In this embodiment, a style extraction module (PPE, Prompt PatternExtractor) is set in the system. This module stably generates style-leading prompts with identity semantics, which are stored in a static style structure instruction block, ensuring the stability of the generated report in terms of style, tone, and content consistency. This addresses the "long text consistency problem" and solves the issues of inconsistent content style and logical jumps in the generated content found in existing technologies.

[0058] In some embodiments, extracting style-leading prompts with identity semantics from a template report includes: identifying the structure of the template report and obtaining structure tags; quantitatively evaluating the writing style of the template report and determining the writing style category; and extracting contextual terms from the template report.

[0059] In some embodiments, the style extraction module can automatically extract structural semantic tags and writing style features based on existing template reports, and generate control prompt words with role identity semantics to drive the subsequent paragraph generation module. The process of constructing the style structure static instruction block includes: (1) Structure recognition, automatic recognition of the overall structure, section structure and paragraph type of the report in the template report. For example, extract the title of each part of the template report, such as "Chapter 1 Policy Compliance Analysis", "Chapter 2 Engineering Analysis, Greenhouse Gas Emission Evaluation, Comparison and Feasibility Demonstration, Emission Management and Monitoring Plan and Greenhouse Gas Emission Carbon Emission Assessment Conclusion", etc. (2) Style classification, quantitatively evaluate the writing style of typical texts based on syntactic features, writing length, terminology density, passive voice ratio, number of cited standards and other dimensions. (3) Extraction of terminology context and technical template tags, extracting structured expressions such as multiple repetitive frames. For example, "This project adopts xx technology to achieve xx". (4) Style leading prompt word fusion generation: embed structural tags, style categories, contextual terms and other content into predefined prompt word templates to construct style leading prompt words, which are used to control the output logic and expression mode of the large language model. Style-leading prompts, as static indicator blocks of style structure within the prompts, guide the large language model to generate long texts with stable style structures. In some embodiments, style-leading prompts integrate elements such as role identity definition, input material structure and type, chapter structure requirements, writing style requirements, and output format control.

[0060] In some embodiments, the style structure static instruction block includes at least one of the following fields: an identity definition field for defining the role identity; an input material field for defining the structure and type of the input material; a structure field for defining chapter structure requirements; a writing style field for defining writing style requirements; and an output structure field for defining output format requirements.

[0061] In some embodiments, the professional process enhancement generation method further includes: performing structuring and semantic processing on the input materials.

[0062] In some embodiments, structuring and semanticizing the input materials includes structuring tables, images, formulas, etc., and semantically translating terminology. The main purpose is to transform raw materials with complex data formats and proper nouns into expressions that are easy for large models to understand.

[0063] In some embodiments, the input material is processed in a structured and semantic manner, including: structuring the non-text elements in the input material; and translating the terms in the input material.

[0064] In some embodiments, the non-text elements include at least one of tables, images, and formulas; structuring the non-text elements in the input material includes: extracting table regions from the input material, identifying data in the tables, and converting them into key-value pairs as a structured representation of the tables; and / or identifying images from the input material, extracting node text from the images, and converting it into a graph structure as a structured representation of the images; and / or extracting formulas from the input material, and converting the formulas into expression tuples as a structured representation of the formulas.

[0065] A structured representation of a table, such as key-value pairs in JSON format, like: {"Indicator":"Annual Coal Consumption", "Value":267.08, "Unit":"g / kWh"}.

[0066] Structured representations of images, such as graph structures, such as: { "nodes": ["boiler", "denitrification unit", "dust collector", "induced draft fan", "chimney"], "edges": [ ["boiler", "denitrification device"], ["Denitrification device", "Dust collector"], ["Dust collector", "Exhaust fan"], ["Exhaust fan", "chimney"] ] }

[0067] Structured representations of formulas, such as expression tuples, like: CO2 emissions = Coal consumption × EF{"Factor":"Coal consumption", "Coefficient":"EF", "Result":"CO2 emissions"}.

[0068] In some embodiments, the professional process enhancement generation method further includes: using an evaluation sub-model to assess the consistency and professionalism of the content of the generated paragraphs.

[0069] This disclosure discloses a multimodal consistency detection structure that integrates multi-source inputs such as structural information, original materials, and terminology knowledge to support the training of automated quality scoring models, thereby improving the automated evaluation capability and multimodal consistency of generated content. This enhances the system's ability to process complex multimodal data and perform automated quality detection, reducing reliance on manual review.

[0070] In some embodiments, the consistency and professionalism of the generated paragraphs are evaluated using an evaluation sub-model, including: determining an evaluation standard system, which includes at least one of the following evaluation dimensions: structural consistency, material consistency, standard compliance, common sense, professionalism, and readability; constructing training data based on the evaluation standard system, wherein the training data includes positive and negative examples corresponding to each evaluation dimension; training and fine-tuning an initial evaluation model using the training data to obtain the evaluation sub-model; and using the evaluation sub-model to evaluate the consistency and professionalism of the generated paragraphs.

[0071] In some embodiments, the report includes any one of a technical report, an industry report, a carbon emission assessment report, an energy assessment report, or an environmental assessment report.

[0072] Secondly, referring to Figure 2 This disclosure provides an electronic device, which includes: One or more processors 101; The memory 102 stores one or more programs that, when executed by one or more processors, enable the one or more processors to implement the professional process enhancement generation method described in the first aspect of the present disclosure. One or more I / O interfaces 103 are connected between the processor and the memory and configured to enable information exchange between the processor and the memory.

[0073] The processor 101 is a device with data processing capabilities, including but not limited to a central processing unit (CPU); the memory 102 is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read / write interface) 103 is connected between the processor 101 and the memory 102, enabling information exchange between the processor 101 and the memory 102, including but not limited to a data bus (Bus).

[0074] In some embodiments, the processor 101, memory 102, and I / O interface 103 are interconnected via bus 104, and thus connected to other components of the computing device.

[0075] Thirdly, embodiments of this disclosure provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the professional process enhancement generation method described in the first aspect of this disclosure.

[0076] This disclosure presents a complete workflow of "material processing—style extraction—controllable generation—automatic evaluation," constructing a controllable long text generation system. It has the following significant beneficial effects: (1) Improve the consistency and professionalism of content generation quality.

[0077] This embodiment of the disclosure introduces a style extraction module, which extracts the structural style features and identity semantic tags from the template report to form reusable style-leading prompts, guiding the large model to express professional content with consistent style and standardized terminology. This mechanism significantly reduces style differences and expression confusion in human writing, ensuring that the generated content conforms to industry writing standards and stylistic logic.

[0078] (2) To achieve standardization and controllability of terminology expression.

[0079] This disclosed embodiment, through a terminology translation process chain (context extraction → prompt word control generation → compliance verification), can automatically convert terms into colloquial and compliant professional expressions, and through explicit calls from a terminology database, improve the accuracy, consistency, and readability of terminology expressions, and reduce compliance risks caused by misuse of terms.

[0080] (3) Improve the structural stability of generated content By constructing a segmented generation method, this embodiment of the disclosure supports dividing the entire report into logically clear chapters / paragraphs, generating content and tracking consistency segment by segment, and maintaining semantic consistency throughout.

[0081] (4) Introduce automated evaluation models for quality judgment This disclosure proposes an automated evaluation module that, based on small-sample expert scoring and the construction of positive and negative example pairs, performs supervised fine-tuning of a large language model, enabling it to possess multi-dimensional judgment capabilities such as structural regularity, terminological consistency, and grammatical logic. The evaluation model can complete the scoring and screening of generated content without human intervention, replacing the traditional high-cost process that relies on expert review, and improving the credibility and controllability of the generated content.

[0082] (5) Reduce labor costs and writing cycle, and improve industry adaptability. Combined with the complete automated process of the embodiments of this disclosure, compared to the 3-5 days required for manual drafting of a typical environmental impact assessment report, this system can generate compliant and high-quality draft text within hours when all input materials are available. This greatly shortens the compilation cycle, lowers the professional writing threshold, and effectively supports the needs of industrial digital transformation and large-scale production of environmental impact assessment business.

[0083] The professional process enhancement generation method proposed in this disclosure has functions such as terminology translation, style template matching, segment consistency generation, and automated scoring. It is applicable to multiple technology-based industries where professional texts are the core output. Its application scope is not limited to the field of sustainable development but can be widely extended to the following fields: Firstly, regarding the generation of carbon emission assessment reports, the system is planned to be deployed in key industries (such as thermal power, cement, and steel). It can automatically generate technical reports including carbon emission accounting, emission reduction technology pathways, and policy adaptation analysis, making it an ideal tool for scenarios such as carbon verification, carbon footprint assessment, and dual-carbon audit.

[0084] Secondly, the system has good versatility and configurability, and can be quickly adapted to different domain terminology libraries, report structures and style templates. Therefore, it has broad application prospects in evaluation documents such as ESG reports, energy assessments, safety assessments, occupational health, and soil and water conservation. It is particularly suitable for formal document generation scenarios that require standardized expression, terminology consistency and structural integrity.

[0085] Furthermore, the embodiments disclosed herein can be integrated with functional modules such as policy knowledge graphs, industry classification recommendations, and semantic search to build a one-stop AI-assisted decision-making and report writing platform, and further expand to sub-sectors such as policy retrieval, bidding document writing, regulatory compliance review, and risk control document generation.

[0086] To enable those skilled in the art to more clearly understand the technical solutions provided by the embodiments of this disclosure, the technical solutions provided by the embodiments of this disclosure will be described in detail below through specific embodiments: Example 1 Reference Figure 3 This embodiment proposes a serious and professional intelligent generation and evaluation process for long texts. The overall process adopts a modular design of material processing → semantic and style extraction → controllable generation → automated evaluation to ensure the compliance, multimodal consistency, professionalism and readability of the generated text.

[0087] 1. Material Processing Module The material processing module, as the core preprocessing workflow, involves two steps: structured processing of tables, images, and formulas, and semantic translation of terminology. Its main purpose is to transform the original materials with complex data formats and proper nouns into expressions that are easy for large models to understand, providing a high-quality, structured, and semantic input foundation.

[0088] 1.1 Structured processing of tables, images, and formulas 1.2. Terminological Semantic Translation This submodule consists of three steps: Step 1: Terminology Context Sampling and Semantic Acquisition Context window extraction: Locate typical sentences in which terms appear in multiple report templates, and extract 5-8 words before and after each term as context to ensure complete context.

[0089] Semantic vector encoding: Using a model such as Sentence-BERT to encode the extracted context to obtain semantic vectors.

[0090] Terminology Matching: By comparing the semantic vectors of the above terms with the semantic vectors of similar terms in different industry terminology tables, the system determines the industry field and typical usage context to which the term is more likely to belong. Example: The term "module" appears in the sentence "This system adopts a modular design, which facilitates independent maintenance and expansion of each functional unit." The system extracts the context, obtains the semantic vector, and performs semantic encoding on the entire sentence to understand that the context in which the module appears is related to system design, functional units, and maintenance. Then, it searches for matching module definitions in different industry terminology tables: For example, in the education industry, it refers to content units in teaching, such as grammar modules and reading modules; in the engineering / manufacturing industry, it refers to hardware or functional components, such as electronic control modules and filtering modules. The system finds that words like "functional unit," "maintenance," and "expansion" are closer to the usage habits of modules in the engineering industry, so it determines that the module belongs to the meaning of a functional component in the engineering industry, rather than a teaching module.

[0091] Step 2: Control the generation of a general description A controlled large language model is built using a general large language model and a PromptLayer. The input of the contextual information and semantic attribution results of terms guides the model to generate general explanations of terms that are stylistically standardized and structurally clear.

[0092] Step 3: Terminology Compliance Verification Abbreviation Explanation Check: Check for the presence of unexplained English abbreviations in the terms; Ambiguity detection: Avoid vague expressions, such as "SCR efficiency" which is ambiguous (it is recommended to change it to "SCR denitrification efficiency of the catalyst"). Standardization of expression score: The standardization of the term description is scored based on the rules of the terminology database and the template specifications to determine whether the term description is standard, fluent, and clear.

[0093] 2. Style Extraction Module To achieve a high degree of consistency between the generated text and expert writing habits in terms of structure, expression style, and writing style, this embodiment designs and implements a "style extraction module". This module can automatically extract structural semantic tags and writing style features based on existing template reports, and generate control prompt words with role identity semantics to drive the subsequent paragraph generation module.

[0094] This module is divided into four steps, as follows: Step 1: Structure Identification By utilizing regular expressions, template segmenters, and Sentence-BERT similarity comparison technology, the system automatically identifies the overall structure, section structure, and paragraph types of template reports. For example, it extracts the headings of each section in the template report, such as Chapter 1 Policy Compliance Analysis, Chapter 2 Engineering Analysis, Greenhouse Gas Emission Assessment, Comparison and Feasibility Study, Emission Management and Monitoring Plan, and Greenhouse Gas and Carbon Emission Assessment Conclusions.

[0095] Step 2: Style Classification By combining tools such as Spacy and Textstat, the writing style of typical texts can be quantitatively evaluated based on syntactic features (such as n-gram usage patterns), writing length, term density, passive voice ratio, and number of cited standards.

[0096] Step 3: Extracting Terminology Context and Technical Template Tags Based on TF-IDF and n-gram, it stably extracts structured expressions without relying on semantic reasoning. For example, n-gram can be used to extract repetitive frames in multiple sentences, such as "This project uses xx technology to achieve xx".

[0097] Step 4: Style-leading cue word fusion and generation By embedding structural tags, style categories, and contextual terms into predefined prompt templates, style-leading prompts are constructed to control the output logic and expression of the large language model. These prompts guide the generation of long texts with a stable style structure within the large model, which serves as the first part of the prompts—the static style structure indicator block. The prompts integrate elements such as: ① role definition (e.g., "You are a carbon emission assessment engineer"), ② input material structure and type, ③ chapter structure requirements (please organize the entire text according to the xx structure), ④ writing style requirements (passive voice, logical connectors), and ⑤ output format control (paragraph numbering, chapter concluding remarks). Specific examples are provided in the implementation section below.

[0098] 3. Controllable generation module This module is a large-scale model generation module, aiming to achieve structured, semantically consistent, and stylistically unified generation of long-form text content in vertical domains. Through a three-stage contextual prompting mechanism, combined with semantic memory retrieval and structured knowledge support, this module generates long-form reports segment by segment, ensuring the overall content conforms to a predetermined chapter structure, professional expression style, and contextual consistency. Simultaneously, by collaborating with external memory, analogous to the hierarchical memory model of an operating system, it "expands" the memory capacity of the large model. For example... Figure 4 As shown, its workflow is as follows: Solid lines indicate memory updates after paragraph generation: first, the generation module is called to generate paragraphs, then terms and key data are extracted from the newly generated paragraphs, and then the prompt word's phased execution state block and knowledge graph+ are updated and written to the context database. Dashed lines indicate prompt word updates before paragraph generation: similarity retrieval is performed, and the prompt word's phased execution state block and semantic memory injection block are updated.

[0099] 3.1 Three-part contextual clue words Before each piece of content is generated, the system dynamically constructs structured prompts consisting of the following three parts: (1) Style structure static instruction block: The content does not change during the generation process; This is derived from style guide hints in the upstream style extraction module. This includes requirements for chapter structure, language style, and fixed sentence patterns.

[0100] (2) Phased execution state block: Record the current task execution status in real time; This includes the current paragraph name, a list of terms and key data used, etc. The module is automatically maintained by the function selection module; Because the storage capacity of this block is limited, redundant terms and data will be written into a structured knowledge graph for unified management.

[0101] Example: Current paragraph number: Paragraph 4 (Benchmarking analysis) Previous paragraph summary: The previous paragraph analyzed the total regional emissions and their source structure; Current cumulative terms: CO2 emissions, denitrification.

[0102] Current cumulative data: Annual power generation: 7.83 billion kWh, coal consumption per unit: 273 g / kWh, carbon content: 0.67 t / t.

[0103] (3) Semantic memory injection block: Automatically retrieve historical paragraphs related to the current paragraph from the context memory database; Based on vector similarity retrieval, the Top-K content is returned to guide contextual consistency. The returned results are updated by injecting semantic memory into the result block as auxiliary information.

[0104] 3.2 External Memory (1) Structured knowledge graph: Used to store technical terms and key data that could not be fully contained in the phased execution status block; Whenever the system detects a new term or variable (such as the first appearance of "unit coal consumption" or "carbon content"), it automatically creates a graph node and labels the initial metadata; if the term already exists in the graph, it updates its frequency of occurrence and the most recently referenced segment record.

[0105] (2) Unstructured context database: Stores the complete original text of the currently generated paragraph and its metadata (such as paragraph number, topic, terminology list, paragraph summary, etc.); All text will be vectorized using an embedding model and then written to the database to support subsequent fast semantic retrieval; To avoid confusion, this database only stores the generated text of the current task.

[0106] 3.3 Function Selection Module This module is responsible for coordinating the dynamic updates and content retrieval tasks of the context storage system during paragraph generation. This module does not involve the generation and invocation of the large language model. Its core functions are: updating the generated content to external memory (including both structured and unstructured databases); and before the next round of generation, retrieving highly relevant content from the memory based on the paragraph topic to construct the "semantic memory injection block" for prompt words.

[0107] (1) Database linkage update mechanism After each piece of content is generated, the system automatically executes the following update process: ① Knowledge graph update: Used to store terms and key data that could not be contained in the phased execution status block; If a new term or key data appears for the first time, a node will be automatically created and added to the terminology graph; if it already exists, its frequency of occurrence and the latest cited paragraph will be updated.

[0108] ② Writing to unstructured semantic memory: The system stores the full text of the currently generated paragraph along with its metadata (paragraph name, terminology, etc.) into the context database; Text is embedded and vectorized to support fast similarity retrieval.

[0109] (2) Semantic Memory Retrieval Strategy: To enhance the contextual consistency and reusability of the generated text, a memory recall process is performed before each segment is generated: The search process includes: Input: The name of the next paragraph to be generated; Retrieval method: Vectorize and compare all historical paragraphs using an embedding model to calculate semantic similarity; Sorting method: Taking into account both semantic similarity and temporal relevance; Output: Returns the text of the Top-K paragraphs, updating the content of the "semantic memory injection block" in the prompt words.

[0110] 4. Automated evaluation module This module uses a finely tuned large language model to evaluate the consistency and professionalism of the generated content across multiple dimensions, including structure, materials, expression, and grammar. Its core advantage lies in its ability to train a lightweight evaluation model with assessment capabilities without requiring extensive manual annotation. This is achieved through expert scoring of a small number of samples and the construction of high-quality positive and negative example pairs.

[0111] The automated evaluation module includes two key subsystems: the subsystem for determining the evaluation criteria system and the subsystem for fine-tuning the large evaluation model.

[0112] 4.1 Determine the evaluation standard system To enable multi-dimensional structured evaluation of the generated text, this module defines six evaluation dimensions, each with a uniform scoring scale (0-5 points) and judgment criteria, as follows: Structural consistency: Determine whether the generated content follows the preset template structure, such as whether the overall structure is consistent with the definition of style-leading keywords.

[0113] Material consistency: Assess whether the content reasonably references the "processed materials", including text, tables, formulas, images, etc., and check for any missing, misplaced, or inconsistent references.

[0114] Compliance: Check whether legal clauses or industry standard expressions are used correctly to avoid rule-related errors.

[0115] Common sense: Determine whether the text is fluent and semantically coherent, and whether the expression conforms to common general knowledge, logical structure and reasoning ability in large language models.

[0116] Professionalism: Assess whether the wording is accurate, whether it reflects the logical integrity and level of expression in the professional field, and whether there are any inappropriate word choices or missing terms.

[0117] Readability: Check whether the text is fluent and clear, has a reasonable structure, and whether there are problems such as disordered word order or difficulty in understanding.

[0118] 4.2 Fine-tuning the large-scale evaluation model subsystem This subsystem uses supervised fine-tuning (SFT) to train an open-source large language model with a small sample of professional scoring data, enabling it to judge the quality of professional reports.

[0119] (1) Training data construction method The submodule is designed with multiple types of positive and negative examples, corresponding to the six evaluation dimensions mentioned above. The construction method is as follows: Structural consistency dimension: positive examples generate multi-level report paragraphs by following chapter templates; negative examples are constructed by shuffling paragraph order, deleting chapter logic, stacking duplicate information, etc.

[0120] Material consistency dimension: Positive examples strictly reference processed materials (such as tables, images, terminology explanations, etc.) to ensure contextual consistency; negative examples introduce irrelevant materials, omit or misuse data tables or terms, causing content confusion or factual errors.

[0121] Compliance dimension: Positive examples are achieved by clearly citing policies and regulations and their document numbers or target expressions; negative examples are achieved by using vague descriptions, incorrect abbreviations or incorrect units (such as missing concentration units in "NOx reduced to 230mg / m³").

[0122] General common sense dimension: Positive examples are generated through logical reasoning and in accordance with basic semantic reasoning; negative examples are constructed by introducing causal confusion, contradictions, and knowledge errors.

[0123] Professionalism dimension: Positive examples are generated using industry-standard terminology and professional expression style prompts; negative examples are generated by replacing them with colloquial, non-professional terms (e.g., "desulfurization equipment" is referred to as "device for removing alkaline gas").

[0124] Readability dimension: Positive examples are generated using standard language and clear paragraph prompts; negative examples are constructed by splicing multiple paragraphs, deleting connecting sentences, or creating logical breaks, making the content difficult to understand or the word order chaotic.

[0125] (2) Training process and methods Model selection: Mistral-7B, a medium-sized open-source language model that supports supervised fine-tuning, was selected; Training method: Use low-rank adaptation (LoRA) for efficient fine-tuning; Data strategy: Use the above positive and negative sample construction method, without relying on large-scale manual annotation; Evaluation metrics: Consistency assessment of small sample size using expert scoring + GPT-4 output.

[0126] Example 2 This embodiment verifies the practical application capability of the "EPAG Professional Process Enhancement Generation System" in professional long text generation tasks. Using project design specifications, policies and regulations as materials, and an environmental impact assessment report of a coal-fired power plant expansion project as a template report, a credible environmental impact assessment report paragraph is generated through four stages: "material processing → style extraction → control generation → automated evaluation". This demonstrates the core technical path and practical effect of this embodiment.

[0127] I. Material Processing Module The original input contains unstructured and semi-structured materials from multiple sources, such as: Project design specifications (unit type, installed capacity, emission equipment, fuel parameters); Table data (unit coal consumption, annual carbon dioxide emissions); Policies and regulations ("Dual Carbon Action Plan", "Feasible Technical Guidelines for Pollution Prevention and Control in Thermal Power Plants" (HJ2301-2017)); Report structure template (such as the "3.1 Analysis of Emission Reduction Measures" section).

[0128] The system will automatically perform the following processing steps: (1) Structured processing of tables, images, and formulas, for example, converting images into standard graph structures and storing only the latter. Taking the flue gas purification process diagram as an example, it can be simply structured as follows: { "nodes": ["boiler", "denitrification unit", "dust collector", "induced draft fan", "chimney"], "edges": [ ["boiler", "denitrification device"], ["Denitrification device", "Dust collector"], ["Dust collector", "Exhaust fan"], ["Exhaust fan", "chimney"] ] }

[0129] (2) Terminology translation. For example: You are a technical terminology explanation assistant. Please explain the following terms in plain, concise language without using technical abbreviations, and do not use the terms themselves.

[0130] Terminology: SCR denitrification; Contextual fragment: Combined with SCR denitrification unit...designed denitrification efficiency of 85%; Low-NOx combustion technology in boilers is the preferred technology for controlling NOx. Ensure NOx concentration; Precise ammonia injection and uniform flow field distribution are among the measures that ensure the stable and efficient operation of the denitrification facilities.

[0131] Attribution determination result: Industry Classification: Environmental Protection Industry - Flue Gas Treatment.

[0132] Application scenario: Technical specifications for flue gas purification systems in thermal power plant boilers.

[0133] Output example: Selective catalytic reduction (SCR) technology is a flue gas purification method that reduces nitrogen oxides to nitrogen and water by injecting a reducing agent (such as urea).

[0134] II. Style Extraction Module The style extraction module automatically performs the four steps described in Example 1 on the template report: structure recognition, style classification, extraction of terminology context and technical template tags, and fusion generation of style leading prompts. An example of generated style leading prompts is as follows: [Character Setting] You are a senior carbon emission assessment engineer, skilled in preparing various environmental impact assessment technical documents in accordance with national and industry standards. You have extensive policy understanding and compliance analysis capabilities, and are familiar with the characteristics, control technologies, and standard references of greenhouse gas emissions in the power generation industry.

[0135]

Task Objective

[0136] [Input materials] You will be writing based on the following materials: Project background and basic information (such as installed capacity, unit type, geographical location); Tabular data (such as unit coal consumption, annual power generation, carbon emissions, and benchmark). List of control technologies (such as SCR denitrification, low-NOx combustion, and energy efficiency improvement measures); Applicable policies (such as the "Guidelines for Feasible Technologies for Pollution Prevention and Control in Thermal Power Plants" and the "Dual Carbon Action Plan"). Template report chapter structure and writing style guidelines.

[0137] [Writing Structure] Please organize the entire text according to the following chapter structure: 1. Project Overview and Emissions Background; 2. Greenhouse gas emission inventory preparation and data explanation; 3. Greenhouse gas emission reduction measures and technological pathways; 4. Emission intensity analysis and industry comparison; 5. Compliance analysis and policy suitability assessment; 6. Conclusions and Recommendations.

[0138] Style Requirements Use formal, descriptive language; Primarily using the passive voice (e.g., "was measured as", "used to achieve..."). Maintain standardized terminology and prioritize the use of industry-standard expressions such as "unit carbon emission intensity", "tCO2 / MWh", and "control efficiency". Use logical connectors (such as "in conclusion", "through...it can be seen that...") to enhance the coherence of the argument; Each chapter should include: background / basis → technical and data explanation → conclusions or recommendations.

[0139] Output Format Output the data in sections according to chapter number (e.g., “4.2 Analysis of Emission Reduction Technologies”). It can automatically generate table reference suggestions (such as "see Table X...") based on the provided data; Each chapter should end with a logical summary sentence that reflects the main findings or basis for judgments made in that chapter.

[0140] [Generation Control] If the generated model contains duplicates, skipped segments, or common sense conflicts, please automatically correct or mark them. When citing standards or regulations, please include the document number (e.g., GBXXXX-XXXX) or the name of the policy source; All indicators must be output in a standardized manner (e.g., “kg / t”, “mg / Nm³”, “tCO2e”).

[0141] Example opening sentence "This project is in response to the national dual-carbon goals and the clean development policy of coal-fired power. It plans to expand two 1000MW coal-fired power units in XX City. It is a typical high-parameter, large-capacity thermal power project. This evaluation report is prepared to assess its greenhouse gas emission impact."

[0142] Please draft a complete environmental impact assessment report based on this.

[0143] III. Controlled Generation Module 1. Prompt Example When generating chapter content (such as "3.2 Greenhouse Gas Emission Calculation"), the three-part Prompt is as follows: [Example of Prompt structure]: (1) Style structure static instruction block: See the example of style-leading words above.

[0144] (2) Phased execution state block: Current segment number: Segment 4 (Greenhouse gas emission calculation); Terms already used: greenhouse gases, emission factors, unit coal consumption; Key data: Annual power generation = 7.83 billion kWh, unit coal consumption = 273 g / kWh, carbon content = 0.67 t / t; Previous paragraph summary: The previous paragraph analyzed the setting and accounting methods for emission limits.

[0145] (3) Semantic memory injection block (Top-K recall segment summary): Recall Item 1: Greenhouse Gas Compliance Assessment; Recall Item 2: Introduction to Greenhouse Gas Measurement Standards.

[0146] 2. Contextual consistency tracking mechanism To ensure consistency in terminology, key data, and stylistic expressions during the generation of long texts across multiple paragraphs, the system performs the following two synchronized updates after each paragraph is generated: (1) Updating structured knowledge graphs If new terms or key data appear in the current paragraph (such as the first appearance of "carbon-oxygen ratio" or "emission intensity = 0.672 t / MWh"), then: Automatically create terminology nodes; Update usage frequency and reference context segment.

[0147] (2) Writing to unstructured context database The system will store the full text (non-summary) generated in the current segment into the unstructured database.

[0148] 3. Recall logic before the next segment is generated: Before the next paragraph is generated, the system will perform a vector search from the context database based on the paragraph topic (such as "3.3 Identification of Key Emission Sources"), using keywords and paragraph numbers as conditions, sort the paragraphs, and return the Top-K related paragraphs to update the semantic memory injection block of the new round of Prompt.

[0149] IV. Automated Evaluation Module The final generated paragraphs will be scored by a finely tuned evaluation model, as shown in the following example: It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0150] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in connection with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in connection with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of this disclosure as set forth by the appended claims.

Claims

1. A method for enhancing and generating professional processes, comprising: The associated context information of the paragraph to be generated is obtained from memory based on the status information of the paragraph to be generated in the report; Construct prompt words for the paragraph to be generated, wherein the prompt words contain the associated context information; The prompt words are input into a large language model, which then generates the content of the paragraph to be generated based on the associated context information. Contextual information is extracted from the content of the paragraph to be generated and updated in the memory.

2. The professional process enhancement generation method according to claim 1, wherein, The associated context information includes a knowledge graph and the content of already generated paragraphs related to the paragraph to be generated; Based on the status information of the paragraph to be generated in the report, the associated context information of the paragraph to be generated is obtained from memory, including: Based on the state information of the paragraph to be generated, a knowledge graph is retrieved from the memory, wherein the knowledge graph contains at least one of terms and data; Based on the status information of the paragraph to be generated, the content of the already generated paragraphs related to the paragraph to be generated is retrieved from the memory.

3. The professional process enhancement generation method according to claim 2, wherein, Extracting context information from the content of the paragraph to be generated and updating it in memory includes: The knowledge graph update information is extracted based on the content of the paragraph to be generated, and the knowledge graph in the memory is updated based on the update information. The content of the paragraph to be generated is vectorized and stored in the memory.

4. The professional process enhancement generation method according to claim 3, wherein, The memory includes external memory; updating the knowledge graph in the memory according to the update information includes: For newly detected terms or data variables, create graph nodes and label initial metadata; For terms that already exist in the knowledge graph, update the frequency of occurrence and citation records of the terms.

5. The professional process enhancement generation method according to claim 2, wherein, The prompt words include a phased execution state block and a semantic memory injection block; Construct prompts for the paragraph to be generated, including: Update the knowledge graph to the phased execution state block; Update the content of the generated paragraphs related to the paragraph to be generated into the semantic memory injection block.

6. The professional process enhancement generation method according to claim 5, wherein, The phased execution status block includes at least one of the following fields: The paragraph information field is used to store information about the paragraph to be generated; The terminology field is used to store terms that have been used. Data fields are used to store data; The Previous Paragraph Summary field is used to store a summary of the previous paragraph.

7. The professional process enhancement generation method according to claim 5, wherein, The prompt words also include style structure static instruction blocks; Construct prompts for the paragraph to be generated, including: Extract style-leading prompts with identity semantics from the template report; Construct the style structure static instruction block based on the style guide prompt content.

8. The professional process enhancement generation method according to claim 7, wherein, Extract style-leading prompts with identity semantics from the template report, including: Identify the structure of the template report and obtain the structure tags; Quantitatively evaluate the writing style of the template report and determine the writing style category; Extract contextual terms from the template report.

9. The professional process enhancement generation method according to claim 8, wherein, The style structure static instruction block includes at least one of the following fields: The identity definition field is used to define the role identity; The input material field is used to define the structure and type of the input material; The structure field is used to define the chapter structure requirements; The writing style field is used to define writing style requirements; Output structure field, used to define output format requirements.

10. The professional process enhancement generation method according to any one of claims 1 to 9, wherein, The professional process enhancement generation method also includes: The input materials are processed in a structured and semantic manner.

11. The professional process enhancement generation method according to claim 10, wherein, The input materials are processed in a structured and semantic manner, including: The non-text elements in the input material are structured. The terminology in the input material is translated.

12. The professional process enhancement generation method according to claim 11, wherein, The non-text elements include at least one of tables, images, and formulas; structuring the non-text elements in the input material includes: Extract table regions from the input material, identify the data in the tables, and convert them into key-value pairs as a structured representation of the tables; and / or Identify images in the input material, extract node text from the images and convert it into a graph structure as a structured representation of the images; and / or Extract the formulas from the input materials and convert them into expression tuples as a structured representation of the formulas.

13. The professional process enhancement generation method according to any one of claims 1 to 9, wherein, The professional process enhancement generation method also includes: The consistency and professionalism of the generated paragraphs are evaluated using an evaluation sub-model.

14. The professional process enhancement generation method according to claim 13, wherein, The evaluation sub-model is used to assess the consistency and professionalism of the generated paragraphs, including: A system of evaluation criteria shall be established, which shall include at least one of the following evaluation dimensions: structural consistency, material consistency, standard compliance, common sense, professionalism, and readability. Training data is constructed based on the evaluation criteria system, wherein the training data includes positive and negative examples corresponding to each evaluation dimension; The initial evaluation model is trained and fine-tuned using the training data to obtain the evaluation sub-model; The evaluation sub-model is used to assess the consistency and professionalism of the generated paragraphs.

15. The professional process enhancement generation method according to any one of claims 1 to 9, wherein, The report includes any one of the following: technical report, industry report, carbon emission assessment report, energy assessment report, and environmental assessment report.

16. An electronic device comprising: One or more processors; A memory having stored one or more programs that, when executed by one or more processors, cause the one or more processors to implement the professional process enhancement generation method according to any one of claims 1 to 15.

17. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the professional process enhancement generation method as described in any one of claims 1 to 15.

Images