Environmental survey report structured information extraction method based on large language model

Abstract

The invention discloses an environment survey report structured information extraction method based on a large language model, and the method comprises the steps: obtaining a PDF file of an environment survey report, constructing an environment survey report chapter tree, and carrying out the splitting processing, and obtaining a lightweight environment survey report; introducing a field layer, a fragment layer and a dialogue layer, constructing a hierarchical Prompt large language model, and extracting the lightweight environment survey report to obtain a plurality of candidate results; and performing confidence comprehensive score calculation on the plurality of candidate results, performing screening, and performing adaptive optimization on the screening result to realize extraction of the structured information of the environmental survey report. According to the method, invalid Token calculation can be reduced, the model calling cost is reduced, and meanwhile, the extraction precision of the key field is remarkably improved. The environmental survey report structured information extraction method based on the large language model can be widely applied to the technical field of artificial intelligence.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for extracting structured information from environmental survey reports based on a large language model. Background Technology

[0002] Environmental contaminated site investigation reports contain a wealth of important information about pollution sources, pollution extent, pollutants, and environmental risks. These reports are typically presented as unstructured text and are mostly in PDF format. Traditional information extraction methods are largely based on rule matching or shallow machine learning models. These methods require extensive manual rule writing or preparation of large amounts of labeled data, resulting in low efficiency and poor scalability, making them ill-suited to the diverse and complex nature of environmental report texts. In recent years, deep learning, especially pre-trained language models based on Transformers (such as BERT and GPT), has demonstrated powerful capabilities in natural language processing, but currently, no effective specialized extraction scheme for environmental investigation reports has been developed. Summary of the Invention

[0003] To address the aforementioned technical problems, the purpose of this invention is to provide a method for extracting structured information from environmental survey reports based on a large language model.

[0004] The first technical solution adopted in this invention is: a method for extracting structured information from environmental survey reports based on a large language model, comprising the following steps: Obtain the PDF file of the environmental investigation report, construct the chapter tree of the environmental investigation report, and split it to obtain a lightweight environmental investigation report; By introducing field layer, fragment layer and dialogue layer, a large language model of layered Prompt is constructed and extracted from lightweight environment survey report to obtain several candidate results; A comprehensive confidence score is calculated for several candidate results and then filtered. The filtered results are then adaptively optimized to achieve structured information extraction from environmental investigation reports.

[0005] Furthermore, the step of obtaining a lightweight environmental investigation report by constructing a chapter tree from the PDF file of the environmental investigation report and splitting it into sections specifically includes: Obtain the PDF file of the environmental investigation report and perform structural chapter division processing to construct an environmental investigation report chapter tree with annotation information; Based on the structured fields and relationships in the PDF file of the environmental survey report, an environmental survey domain pattern is introduced to construct a readable environmental survey knowledge graph; Based on the readable environmental survey knowledge graph, the chapter tree of the environmental survey report with labeled information is split to construct a lightweight environmental survey report.

[0006] Furthermore, the step of obtaining the PDF file of the environmental investigation report, dividing it into structural chapters, and constructing an environmental investigation report chapter tree with annotation information specifically includes: Obtain the PDF file of the environmental survey report and parse it using a layout parsing tool to extract the text content and layout structure information; Based on the text content and page layout information, structural identification and chapter division are performed to construct a chapter tree for the environmental survey report; Based on the chapter tree of the environmental survey report, each chapter node is labeled with a chapter type label and its text position in the PDF file, resulting in an environmental survey report chapter tree with labeled information.

[0007] Furthermore, the step of constructing a readable environmental survey knowledge graph by introducing environmental survey domain patterns based on the structured fields and relationships in the PDF file of the environmental survey report specifically includes: Obtain the structured fields and relationships from the PDF file of the environmental survey report to obtain the environmental survey domain schema; Based on the environmental survey field model, an environmental survey knowledge graph is obtained by mapping it with the preset environmental standard risk judgment threshold. Obtain entities, attributes, and relationships from the environmental survey knowledge graph, assign them machine-readable identifiers, and construct a readable environmental survey knowledge graph.

[0008] Furthermore, the step of splitting the chapter tree of the environmental survey report with labeled information based on the readable environmental survey knowledge graph to construct a lightweight environmental survey report specifically includes: Based on the readable environment survey knowledge graph, keyword positioning is performed on the chapter tree of the environment survey report with labeled information to determine candidate text fragments; For table-type regions in candidate text fragments, the table structure recognition algorithm is used to parse the row and column structure into quadruples, and the original cell content is associated with the semantic tags of the chapter to obtain the associated quadruples, which are sampling point-index-value-unit. Candidate text fragments are split into several "field-candidate fragment groups" and combined with the associated quadruples to construct a lightweight environmental survey report.

[0009] Furthermore, the hierarchical Prompt large language model specifically includes a field layer, a fragment layer, and a dialogue layer, wherein: The field layer is based on the environmental survey domain model and constructs a Prompt template consisting of "field definition + example + output JSON template"; The fragment layer designs several semantically equivalent but differently expressed Prompt templates based on candidate text fragments; The dialogue layer is used to automatically ask follow-up questions based on incomplete results and complete them within the same dialogue context.

[0010] Furthermore, the step of calculating and filtering a comprehensive confidence score for several candidate results, and adaptively optimizing the filtering results to achieve structured information extraction from the environmental investigation report, specifically includes: Several candidate results are standardized and their confidence scores are calculated and then filtered to obtain the final extraction results. The final extraction results are then subjected to cross-report entity alignment and information aggregation to obtain the aggregated extraction results. Outlier detection and optimization are performed on the aggregated extraction results to obtain consistent extraction results; The extraction results after consistency calibration are written into the environmental survey knowledge graph and output, realizing the extraction of structured information from the environmental survey report.

[0011] Furthermore, the step of standardizing several candidate results, calculating and screening based on confidence scores, and obtaining the final extraction result specifically includes: Several candidate results are aligned and formatted to obtain standardized candidate results. The standardized candidate results are scored in multiple dimensions to obtain a multi-dimensional consistency metric. The multi-dimensional scores include semantic matching degree, cross-template consistency, knowledge graph constraint satisfaction degree, and cross-chapter logical self-consistency degree. The multidimensional consistency measure values ​​of the candidate results are weighted according to preset weights to obtain the comprehensive confidence score of the candidate results; Candidates with a confidence score higher than a preset threshold are selected as the final extraction results.

[0012] The beneficial effects of the method of this invention are as follows: This invention constructs a chapter tree of the environmental investigation report by acquiring the PDF file of the environmental investigation report and performs splitting processing to obtain a lightweight environmental investigation report. It can lock key fields into corresponding chapters or tables and then send them into a large language model for processing, which significantly reduces the interference of irrelevant text. Furthermore, it introduces a field layer, a fragment layer, and a dialogue layer to construct a hierarchical Prompt large language model and extract several candidate results from the lightweight environmental investigation report. It pre-constructs environmental investigation domain patterns and environmental knowledge graphs, encodes pollutant-medium-limit and risk level determination rules into computable constraints, and enforces checks during confidence calculation and result calibration. Finally, it calculates and filters the confidence scores of several candidate results and adaptively optimizes the filtered results to achieve structured information extraction from the environmental investigation report. This can reduce invalid token calculations, reduce model call costs, and significantly improve the extraction accuracy of key fields. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of a method for extracting structured information from an environmental survey report based on a large language model, as described in this invention. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0015] First, it should be noted that the embodiments of this invention address the problems of high maintenance costs and poor generalization ability of traditional rule-based methods, heavy reliance on large-scale labeled data for supervised learning methods, lack of domain-specific hints and confidence reliability guarantees when directly applying general-purpose large language models, and low efficiency and lack of unified and scalable solutions for batch processing of PDF documents. Therefore, considering the characteristics of environmental pollution site investigation reports—long, complex in format (mostly PDF), diverse field types, and strong regulatory constraints (pollutant limits, risk level determination rules, etc.)—an overall architecture of "format-aware parsing - environmental domain knowledge model - constraint-driven consistency calibration" is constructed. This architecture not only utilizes the language understanding capabilities of large language models for extraction but also performs logical verification and calibration of the extraction results through environmental domain knowledge graphs and regulatory constraints, thereby obtaining highly confident structured results.

[0016] Reference Figure 1 This invention provides a method for extracting structured information from environmental survey reports based on a large language model. The method includes the following steps: S100. Obtain the PDF file of the environmental investigation report, construct the chapter tree of the environmental investigation report, and split it to obtain a lightweight environmental investigation report; S110. Obtain the PDF file of the environmental investigation report and perform structural chapter division processing to construct an environmental investigation report chapter tree with annotation information; Specifically, the PDF file of the environmental investigation report is obtained and parsed using a layout parsing tool to obtain the text content and layout structure information; based on the text content and layout structure information, the structure is identified and chapters are divided to construct an environmental investigation report chapter tree; based on the environmental investigation report chapter tree, each chapter node is labeled with a chapter type label and its text position in the PDF file to obtain an environmental investigation report chapter tree with labeled information.

[0017] In this embodiment, a PDF file of the environmental survey report is obtained; further, the PDF is parsed using a layout parsing tool to extract the text content while retaining the layout structure information, including: 1) Page numbering, text block coordinates, font size, bold / heading style, etc.; 2) Structured layout elements such as table areas, figure titles, headers and footers.

[0018] Furthermore, based on layout features and keywords, the report is divided into chapters and its structure is identified, resulting in a chapter tree that reflects the unique structure of an environmental investigation report, including: 1) Site overview, land use history, and potential pollution sources; 2) Sampling point layout and monitoring plan; 3) Test result table; 4) Risk assessment and risk management recommendations.

[0019] Finally, in the chapter tree, each chapter node is labeled with a chapter type tag and its text position range in the original PDF. This embodiment, by explicitly modeling PDF layouts and common chapter structures in environment reports, solves the problems of input redundancy and semantic confusion caused by long texts and complex layouts in reports from the source, providing targeted contextual constraints for efficient and accurate extraction by subsequent large language models.

[0020] S120. Based on the structured fields and relationships in the PDF file of the environmental survey report, introduce environmental survey domain patterns and construct a readable environmental survey knowledge graph; Specifically, the structured fields and relationships in the PDF file of the environmental investigation report are obtained to obtain the environmental investigation domain pattern; based on the environmental investigation domain pattern, it is mapped with the preset environmental standard risk judgment threshold to obtain the environmental investigation knowledge graph; the entities, attributes and relationships in the environmental investigation knowledge graph are obtained, assigned machine-readable labels, and a readable environmental investigation knowledge graph is constructed.

[0021] In this embodiment, the structured fields and relationships to be extracted from the environmental survey report are pre-analyzed to form an environmental survey domain schema, including: 1) Physical entities: plots of land, sampling points, media (soil, groundwater, etc.), pollutants, monitoring indicators, risk levels, remediation measures, etc. 2) Attribute fields: plot name, geographical location, land use, pollutant name, detection concentration, unit, judgment criteria, risk conclusion, etc.; 3) Relationship types: plot-pollutant, sampling point-test result, pollutant-environmental medium, test result-risk level, etc.

[0022] By mapping the above models to relevant national / local environmental standards and risk assessment guidelines, limits and judgment rules are established to construct an environmental survey knowledge graph, including: 1) A ternary group of pollutant-media-standard limit; 2) The correspondence rules between risk levels, concentration ranges, and exceedance multiples; 3) Knowledge of the relationship between typical industries, processes and characteristic pollutants.

[0023] A unified, machine-readable identifier is generated for entities, attributes, and relationships in the knowledge graph, which is used for subsequent prompt word design, extraction result alignment, and logical constraint checks. This embodiment not only predefines extraction fields but also explicitly encodes environmental regulations and decision rules into a computable knowledge graph, which will be used later to perform "physical / regulatory constraint verification" on the output of the large language model.

[0024] S130. Based on the readable environmental survey knowledge graph, the chapter tree of the environmental survey report with labeled information is split to construct a lightweight environmental survey report.

[0025] Specifically, based on the readable environment survey knowledge graph, keyword positioning is performed on the chapter tree of the environment survey report with labeled information to determine candidate text fragments; for table-type areas in the candidate text fragments, the table structure recognition algorithm is used to parse the row and column structure into quadruples, and the original cell content is associated with the semantic tags of the chapter to obtain the associated quadruples, which are sampling point-indicator-value-unit; the candidate text fragments are split into several "field-candidate fragment groups", and combined with the associated quadruples, a lightweight environment survey report is constructed.

[0026] In this embodiment, candidate text fragments are located based on the chapter tree and domain pattern for different fields / relationships, for example: 1) Pollution source-related fields can only be retrieved in the "Land Use History", "Process Flow", and "Pollution Source Identification" sections; 2) Numerical fields such as concentration and unit should be extracted from the "Detection Results Table" and "Monitoring Data" sections. 3) Risk conclusion fields should be located primarily in the "Risk Assessment" and "Conclusions and Recommendations" sections; For table-type areas, a table structure recognition algorithm is used to parse the row and column structure into candidate quadruples of "sampling point-indicator-value-unit", and the original cell content is associated with the semantic tags of the chapter. The entire long report is split into several "field-candidate fragment groups" to generate a compact context window for subsequent large language model extraction, avoiding interference from irrelevant content.

[0027] This embodiment significantly reduces the noise impact of irrelevant text on the generated results by adopting a task splitting strategy of "first locate, then ask the model", which reduces the computational load of model calls and provides a structured input basis for subsequent multi-round extraction and consistency calibration.

[0028] S200 introduces field layer, fragment layer and dialogue layer, constructs a layered Prompt large language model and extracts from lightweight environment survey report to obtain several candidate results; In this embodiment, the large language model is not simply used as a "question-answering tool," but is embedded in a hierarchical extraction framework, specifically including: 1) Field layer Prompt design.

[0029] For each entity / field type, construct a Prompt template consisting of "field definition + example + output JSON template". The template explicitly references the field names and meanings of the domain schema, for example:

Field Definition

[0030] For the same field, several semantically equivalent but differently expressed Prompt templates (multiple templates) are designed for combinations of candidate fragments, including: a template for "extracting numerical values ​​and units"; a template for "explaining the origin of risk conclusions"; and a template for "judging against standard limits".

[0031] The large language model is called for each "field-candidate fragment group" to obtain multiple candidate structured results.

[0032] 3) Dialogue layer follow-up questions and completion.

[0033] For incomplete results provided by the model (such as missing units or standard numbers), a follow-up prompt is automatically generated and completed within the same dialogue context; for multiple possible values ​​provided by the model, an additional prompt is added to request the model to provide an "uncertainty explanation" to provide a basis for subsequent confidence assessment.

[0034] This embodiment deeply integrates Prompt design with domain patterns and knowledge graphs, and employs a layered extraction structure of "field layer - fragment layer - dialogue layer" instead of a simple single-round cloze test or single-round question-and-answer format. The multi-template approach in this embodiment not only pursues diversity in relational representation but also emphasizes explicit constraints on environmental field semantics and regulatory semantics.

[0035] S300: Calculate and screen the confidence scores of several candidate results, and adaptively optimize the screening results to achieve structured information extraction from the environmental investigation report.

[0036] Specifically, several candidate results are standardized and their confidence scores are calculated and filtered to obtain the final extraction results; the final extraction results are then aligned across report entities and aggregated to obtain aggregated extraction results; outlier detection and optimization are performed on the aggregated extraction results to obtain consistent calibrated extraction results; the consistent calibrated extraction results are written into the environmental survey knowledge graph and output to achieve structured information extraction from the environmental survey report.

[0037] The process involves aligning and standardizing the fields of several candidate results to obtain standardized candidate results; calculating multi-dimensional scores for the standardized candidate results to obtain multi-dimensional consistency metrics, including semantic matching degree, cross-template consistency, knowledge graph constraint satisfaction degree, and cross-chapter logical consistency degree; weighting the multi-dimensional consistency metrics of the candidate results according to preset weights to obtain a comprehensive confidence score for the candidate results; and selecting candidate results with a comprehensive confidence score higher than a preset threshold as the final extraction results.

[0038] In this embodiment, candidate results are first standardized by aligning fields and standardizing formats of the multiple candidate results, including unit conversion (e.g., unifying mg / kg and μg / g) and standardizing Chinese aliases (e.g., unifying “total petroleum hydrocarbons” and “TPH”). Then, a multidimensional consistency measure is performed. For each candidate structured record, the following multidimensional score is calculated: 1) Semantic matching degree: The semantic similarity between the candidate result and its corresponding original segment is calculated using a text similarity model; 2) Cross-template consistency: Statistically analyze the consistency of output for the same field under different Prompt templates; 3) Knowledge graph constraint satisfaction: Determine whether the combination of pollutants, media, standard numbers, and risk levels in the candidates is consistent with the knowledge graph, such as: whether the concentration is within the reasonable physical range of the pollutant in the corresponding media; whether the conclusion of exceeding the standard is consistent with the information such as "multiple of exceeding the standard" and "number of monitoring points exceeding the standard" in the description; 4) Cross-chapter logical consistency: Whether contradictory conclusions are given for the same plot of land and the same pollutant in different chapters (such as "monitoring results" and "risk conclusions").

[0039] Further calculate the comprehensive confidence score by weighting the above multidimensional scores according to preset weights to obtain the comprehensive confidence score in the 0–1 interval: [ Score = \alpha \cdot S_{\text{semantic}} + \beta \cdotS_{\text{template}} + \gamma \cdot S_{\text{kg_constraint}} + \delta \cdot S_{\text{logic}} ].

[0040] If a candidate result violates mandatory regulatory constraints (e.g., the concentration is far below the detection limit but it is judged as "severely exceeding the standard"), its confidence level is directly set to 0; the final result selection and manual review mark, and the result with the highest comprehensive confidence level among multiple candidate results is selected as the final extraction result; when the confidence level of all candidate results is lower than the preset threshold, the field is marked as "requires manual review", and the possible reasons are attached (e.g., "exceeds the range of known pollutants in the knowledge graph" or "there is an obvious contradiction in the original text").

[0041] This embodiment incorporates "environmental knowledge graph," "regulatory limit constraints," and "cross-chapter logic verification" into the confidence calculation process, forming a constraint-driven consistency calibration mechanism for environmental survey scenarios. This mechanism can automatically identify model outputs that "appear reasonable in language but are unreasonable in environmental physics / regulations," thereby significantly improving the environmental credibility of the results.

[0042] Then, cross-report entity alignment and information aggregation are carried out. Multiple survey reports from the same region, the same plot of land, or the same enterprise are aligned. Based on a comprehensive matching strategy including geographical location, plot name, and enterprise name, the extraction results from multiple reports are merged into a unified plot entity. Through time series analysis, the changes in pollutant concentration and risk level in different years are compared to identify potential abnormal trends.

[0043] Further anomaly-based adaptive optimization is performed. For cases where there are still obvious anomalies after cross-report aggregation (such as the risk level of the same plot suddenly changing from "safe" to "severely exceeding the standard" within a year without corresponding support from monitoring data), the original text fragments in the corresponding reports are automatically traced back. These anomaly samples are used to construct "difficult example Prompts" or weakly labeled samples to update the Prompt template or fine-tune the sub-model used for confidence calculation, thereby achieving incremental learning with small samples.

[0044] Finally, we continuously improve extraction performance. After processing a new batch of reports each time, we feed the extraction results and manual correction results back to the adaptive optimization module to iteratively update the domain pattern, knowledge graph, and Prompt design.

[0045] This embodiment introduces "consistency across reports" into the model structure, extending a single extraction task into a closed-loop system that continuously self-corrects and strengthens as data accumulates, significantly improving the stability and maintainability of long-term operation in real regulatory scenarios.

[0046] Then, structured output and environmental knowledge graph updates are performed. The extracted results, after consistency calibration, are output in a unified JSON or tabular format, with fields including site attributes, pollutant list, concentration and unit, standard number, risk level and its basis, etc. The extracted results are incrementally written into the environmental survey knowledge graph to form a queryable and reasonable environmental information infrastructure. This provides a structured data interface for upper-level applications (such as environmental risk maps, contaminated site management systems, and project approval auxiliary tools).

[0047] In summary, the embodiments of the present invention have the following advantages compared with the prior art: 1) It has a stronger sense of layout and structure for environmental survey reports (PDF).

[0048] This invention performs layout parsing and chapter structure recognition specifically on environmental survey report PDFs before traditional text extraction, preserving layout information such as tables, titles, paragraphs, headers, and footers, and constructing a document representation of "chapter tree + table structure". This allows fields such as "pollution source", "monitoring results", and "risk conclusions" to be locked in their corresponding chapters / tables before being fed into a large language model for processing, significantly reducing interference from irrelevant text.

[0049] 2) By introducing environmental knowledge graphs and regulatory constraints, the results are more consistent with the logic of environmental physics and regulation.

[0050] This invention pre-constructs environmental survey models and environmental knowledge graphs, encoding pollutant-medium-limit and risk level determination rules into computable constraints. During confidence level calculations and result calibration, it enforces checks on the rationality of the "concentration-unit-medium-standard-risk level" combination. This not only corrects common identification errors but also eliminates outputs that "semantically seem reasonable but violate regulations / physical common sense."

[0051] 3) Candidate fragment localization + hierarchical prompt, balancing high accuracy and high efficiency.

[0052] This invention first uses a chapter tree and table structure to precisely locate candidate segments. Then, for each type of field, it employs a layered Prompt design of "field-level definition + segment-level multiple templates + dialogue-level follow-up questions," breaking down long reports into several smaller tasks and assigning them to a large language model. Compared to directly extracting general Prompts from the entire report, this method ensures that the model always operates in a context strongly related to the target field, reducing invalid token calculations, lowering model call costs, and significantly improving the extraction accuracy of key fields (such as pollutant names, concentrations, and risk conclusions).

[0053] 4) The confidence system based on multi-source consistency and knowledge constraints makes it easier to identify "high-risk errors".

[0054] The confidence level of this invention is not based on a single source, but rather integrates multiple dimensions such as semantic matching degree, cross-Prompt template consistency, knowledge graph constraint satisfaction, and cross-chapter logical consistency to form a multi-source consistency scoring mechanism for environmental investigation scenarios. For results that clearly violate regulatory limits or are contradictory across chapters, the system can automatically determine them as having low confidence and mark them as "requiring manual review," prioritizing reminders to users to check fields that may lead to incorrect regulatory decisions.

[0055] Finally, let's explain this in the context of actual verticality: 1) Data preparation: Collect 100 environmental survey reports (PDF format) and extract the text using pdfplumber or PyMuPDF.

[0056] 2) Prompt template construction: For the fields to be extracted, such as land parcel name and pollution source, construct at least two different prompt templates and add domain examples for few-shot to learn from.

[0057] 3) Information extraction and consistency voting: For each field, multiple templates are called and input into the ChatGLM model. If the results returned by multiple Prompts are consistent, the result is output; if they are inconsistent, multiple rounds of queries are performed to reconfirm, or manual intervention is required.

[0058] 4) Confidence score: Based on model probability, text matching and consistency results, calculate the confidence score for each field. If the confidence score is higher than 0.8, it will be automatically approved; if it is lower than 0.8, it will be marked for manual review.

[0059] 5) Results output and verification: Output in JSON format and Markdown tables; low-confidence results are marked by the user interface for manual verification.

[0060] Comparative analysis results show that the field extraction accuracy reaches over 90%, significantly better than the traditional method's approximately 50%; the efficiency of batch processing 100 reports is improved by about 20 times, which is significantly better than manual extraction one report at a time.

[0061] Therefore, the embodiments of the present invention have the following technical differences compared to the prior art: 1) Layout-aware analysis and chapter structure modeling for environmental survey report PDFs; This invention addresses the common issue of complex PDF formats in environmental investigation reports by proposing a layout-aware parsing and structural modeling technique. This technique identifies layout elements such as pages, text blocks, titles, table areas, headers, and footers within the PDF document, constructing a chapter tree reflecting the report's logical structure. Chapters such as site overview, pollution source identification, monitoring results, risk assessment, and conclusions and recommendations are explicitly divided. Based on this, candidate segments are located for target fields such as site name, pollution source, pollutant concentration, and risk conclusions according to a pre-defined "field-chapter mapping rule." Contextual information is extracted from local text or tables strongly related to these fields and input into the large language model. This layout awareness and chapter modeling significantly reduces the interference of irrelevant content in long documents on model judgment, improving the accuracy and stability of the extraction results. Furthermore, it reduces the size of the effective text to be processed, minimizing the computational overhead of large model calls. This provides a high-quality structured input foundation for subsequent hierarchical prompts and consistency calibration—an effect that existing methods based solely on pure text processing cannot achieve.

[0062] 2) Environmental survey patterns and environmental knowledge graphs are constructed and used as a constraint layer in result calibration; This invention pre-analyzes typical entities, attributes, and relationships in environmental investigation reports before extraction, constructing an environmental investigation domain model including land parcels, sampling points, media, pollutants, detection indicators, standard limits, and risk levels. It organizes knowledge such as national and local pollutant limit standards, risk assessment rules, and industry-specific pollutants into an environmental knowledge graph, uniformly encoded using entity, attribute, and rule triples. After initial extraction from the large language model, this invention utilizes the knowledge graph to verify the legality of combinations such as "pollutant-media-concentration-unit-standard number-risk level." Candidate results that violate regulatory limits, physical common sense, or judgment rules are directly judged as unreliable or low-confidence, thus introducing environmental expertise and regulatory rules into the extraction process as hard / soft constraint layers. In this way, this invention can effectively identify and eliminate erroneous outputs that "seem semantically reasonable but are not valid in environmental regulations and physical sense," ensuring that the structured results are not only textually correct but also usable and reliable in environmental management and compliance decision-making, significantly outperforming existing technologies that rely solely on general text similarity or error mapping tables for post-processing.

[0063] 3) Confidence calculation and low-confidence item marking mechanism based on multidimensional consistency and environmental constraints; After obtaining multiple candidate extraction results, this invention does not select solely based on the output probability of a single model. Instead, it constructs a multi-dimensional consistency and constraint-driven confidence calculation mechanism for environmental survey scenarios: First, it evaluates the semantic matching degree of candidate results, measuring the semantic similarity between the structured results and the corresponding original text fragments. Second, it calculates consistency across Prompt templates, statistically analyzing whether the output of the same field tends to be consistent under different templates. Third, it combines an environmental knowledge graph to check whether the candidate results meet professional constraints such as pollutant limits, exceedance multiples, and risk level determinations. Finally, it examines the logical consistency between chapters such as "Monitoring Results," "Risk Assessment," and "Conclusions and Recommendations" on the same site and the same pollutant from a cross-chapter perspective. The above multi-dimensional scores are combined according to preset weights to form a confidence score in the range of 0 to 1. Results that seriously violate the hard constraints of the domain knowledge graph are directly reduced to 0 and marked as abnormal. When the confidence scores of all candidate results for a certain field are below the threshold, this invention automatically marks it as "requiring manual review," prioritizing prompting users to pay attention to potentially high-risk error fields. This mechanism enables the system to effectively distinguish between "highly reliable and automatically accepted results" and "suspicious results that require manual verification." While improving the overall extraction accuracy, it significantly reduces the workload of manual review, which is of great practical value, especially in environmental regulation, where the tolerance for errors is extremely low.

[0064] 4) The evidence chain output of the extraction results and the incremental update mechanism of the environmental knowledge graph; This invention, while outputting structured extraction results, also provides a corresponding evidence chain output and knowledge graph incremental update mechanism: for each structured record, it not only provides field values ​​such as site information, pollutant name, concentration value, unit, and risk conclusion, but also simultaneously outputs the position identifier of its corresponding original sentence or table cell, as well as the environmental standard number and judgment rule on which it is based. This allows users to establish a traceable causal chain from "result – original evidence – standard rule". For extraction results with high confidence, this invention can incrementally write them into the environmental knowledge graph, gradually enriching environmental knowledge such as pollutant distribution, exceedance status, and risk evolution, which can then be used to constrain and support reasoning in subsequent extraction tasks. Through this mechanism, this invention not only meets the business needs of environmental protection departments for "explainable and traceable" extraction results, but also organically combines the extraction process with knowledge base construction. This enables the system to have good integrability and scalability when supporting upper-level applications such as environmental risk assessment, contaminated site management, and project approval, thereby further amplifying the comprehensive application value of the technical solution of this invention in actual environmental management.

[0065] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for extracting structured information from environmental survey reports based on a large language model, characterized in that, Includes the following steps: Obtain the PDF file of the environmental investigation report, construct the chapter tree of the environmental investigation report, and split it to obtain a lightweight environmental investigation report; By introducing field layer, fragment layer and dialogue layer, a large language model of layered Prompt is constructed and extracted from lightweight environment survey report to obtain several candidate results; A comprehensive confidence score is calculated for several candidate results and then filtered. The filtered results are then adaptively optimized to achieve structured information extraction from environmental investigation reports.

2. The method for extracting structured information from environmental survey reports based on a large language model according to claim 1, characterized in that, The step of obtaining a lightweight environmental investigation report by constructing a chapter tree from the PDF file of the environmental investigation report and splitting it into sections specifically includes: Obtain the PDF file of the environmental investigation report and perform structural chapter division processing to construct an environmental investigation report chapter tree with annotation information; Based on the structured fields and relationships in the PDF file of the environmental survey report, an environmental survey domain pattern is introduced to construct a readable environmental survey knowledge graph; Based on the readable environmental survey knowledge graph, the chapter tree of the environmental survey report with labeled information is split to construct a lightweight environmental survey report.

3. The method for extracting structured information from environmental survey reports based on a large language model according to claim 2, characterized in that, The step of obtaining the PDF file of the environmental investigation report, dividing it into structural chapters, and constructing an environmental investigation report chapter tree with annotation information specifically includes: Obtain the PDF file of the environmental survey report and parse it using a layout parsing tool to extract the text content and layout structure information; Based on the text content and page layout information, structural identification and chapter division are performed to construct a chapter tree for the environmental survey report; Based on the chapter tree of the environmental survey report, each chapter node is labeled with a chapter type label and its text position in the PDF file, resulting in an environmental survey report chapter tree with labeled information.

4. The method for extracting structured information from environmental survey reports based on a large language model according to claim 3, characterized in that, The step of constructing a readable environmental survey knowledge graph by incorporating structured fields and relationships from the PDF file of the environmental survey report, and introducing environmental survey domain patterns, specifically includes: Obtain the structured fields and relationships from the PDF file of the environmental survey report to obtain the environmental survey domain schema; Based on the environmental survey field model, an environmental survey knowledge graph is obtained by mapping it with the preset environmental standard risk judgment threshold. Obtain entities, attributes, and relationships from the environmental survey knowledge graph, assign them machine-readable identifiers, and construct a readable environmental survey knowledge graph.

5. The method for extracting structured information from environmental survey reports based on a large language model according to claim 4, characterized in that, The step of splitting the chapter tree of the environmental survey report with labeled information based on the readable environmental survey knowledge graph to construct a lightweight environmental survey report specifically includes: Based on the readable environment survey knowledge graph, keyword positioning is performed on the chapter tree of the environment survey report with labeled information to determine candidate text fragments; For table-type regions in candidate text fragments, the table structure recognition algorithm is used to parse the row and column structure into quadruples, and the original cell content is associated with the semantic tags of the chapter to obtain the associated quadruples, which are sampling point-index-value-unit. The candidate text fragments are split into several "field-candidate fragment groups", and the associated quadruples are combined to construct a lightweight environmental survey report.

6. The method for extracting structured information from environmental survey reports based on a large language model according to claim 5, characterized in that, The hierarchical Prompt large language model specifically includes a field layer, a fragment layer, and a dialogue layer, wherein: The field layer is based on the environmental survey domain model and constructs a Prompt template consisting of "field definition + example + output JSON template"; The fragment layer designs several semantically equivalent but differently expressed Prompt templates based on candidate text fragments; The dialogue layer is used to automatically ask follow-up questions based on incomplete results and complete them within the same dialogue context.

7. The method for extracting structured information from environmental survey reports based on a large language model according to claim 6, characterized in that, The step of calculating and filtering a number of candidate results based on their confidence scores, and then adaptively optimizing the filtered results to extract structured information from the environmental investigation report, specifically includes: Several candidate results are standardized and their confidence scores are calculated and then filtered to obtain the final extraction results. The final extraction results are then subjected to cross-report entity alignment and information aggregation to obtain the aggregated extraction results. Outlier detection and optimization are performed on the aggregated extraction results to obtain consistent extraction results; The extraction results after consistency calibration are written into the environmental survey knowledge graph and output, realizing the extraction of structured information from the environmental survey report.

8. The method for extracting structured information from environmental survey reports based on a large language model according to claim 7, characterized in that, The step of standardizing several candidate results, calculating a comprehensive confidence score, and filtering them to obtain the final extraction result specifically includes: Several candidate results are aligned and formatted to obtain standardized candidate results. The standardized candidate results are scored in multiple dimensions to obtain a multi-dimensional consistency metric. The multi-dimensional scores include semantic matching degree, cross-template consistency, knowledge graph constraint satisfaction degree, and cross-chapter logical self-consistency degree. The multidimensional consistency measure values ​​of the candidate results are weighted according to preset weights to obtain the comprehensive confidence score of the candidate results; Candidates with a confidence score higher than a preset threshold are selected as the final extraction results.

Images