An infrared transparent conductive material document data extraction and database construction method based on a large language model

Abstract

A kind of infrared transparent conductive material literature data extraction and database construction method based on large language model, the present application aims at solving the problems such as the dispersion of infrared transparent conductive material literature performance data, the difficulty of material and performance attribute association.The database construction method is as follows:1, the literature abstract is used as information extraction text source, and standardization processing is carried out;2, task instruction, to-be-extracted field and its data type are set, and information extraction prompt template is established;3, the standardized literature corpus and information extraction prompt template are combined and then input into large language model;4, the structured extraction record is subjected to structure mode verification and standardization processing;5, the structured extraction record is subjected to original text evidence backtracking and confidence determination;6, repeated record determination and manual review;7, the screened structured extraction record is written into database.The present application can improve the batch extraction efficiency of literature performance data, realize the association among material, performance value and test condition and original text review.

Description

Technical Field

[0001] This invention belongs to the fields of materials informatics and natural language processing, specifically relating to a method for extracting, verifying, and constructing a database of performance data from literature on infrared transparent conductive materials based on a large language model. Background Technology

[0002] Infrared transparent conductive materials need to have high transmittance in the infrared band while also possessing a certain carrier transport capacity. Related research typically involves various performance parameters such as material composition, doping elements, fabrication process, film thickness, test wavelength, band gap, transmittance, conductivity, mobility, effective mass, dielectric constant, and sheet resistivity.

[0003] Currently, the aforementioned material performance data are mainly scattered across journal articles, conference papers, and other scientific literature. Different publications use inconsistent names, units, and numerical expressions for the same performance parameters, and the same material may also have different performance values ​​under different preparation processes, testing temperatures, testing wavelengths, or film thicknesses. Therefore, directly entering the values ​​from the literature into the database can easily lead to the loss of the correspondence between materials, performance parameters, and testing conditions.

[0004] Traditional data acquisition methods primarily rely on manual reading and data entry. However, with the increasing volume of materials and documents, manual processing suffers from low efficiency, inconsistent annotation standards, and difficulty in tracing data sources.

[0005] Information extraction methods based on named entity recognition can identify entities such as material names, numerical values, and units. However, for doped materials, cross-sentence attribute associations, dimensionless performance parameters, and complex numerical expressions, there may be a problem where the entity is correctly identified but the association between the material and the performance attribute is incorrect.

[0006] Large language models possess contextual understanding and structured text generation capabilities, making them suitable for extracting information from materials and literature. However, without restrictions on output fields, the correlation between materials and performance, and the original text evidence, large language models may generate data not explicitly stated in the literature or incorrectly correlate performance values ​​of different materials and under different testing conditions.

[0007] Therefore, a database construction method is needed that can batch process literature on infrared transparent conductive materials and combine structured extraction of large language models, backtracking of original evidence, standardization of material names, unit normalization, rule verification, and manual review. Summary of the Invention

[0008] This invention aims to address the problems of scattered performance data in literature on infrared transparent conductive materials, low efficiency of manual extraction, difficulty in associating material properties with performance attributes, and difficulty in tracing data sources. This invention utilizes a large language model, structured output constraints, backtracking of original evidence, and data verification rules in the field of materials to establish a method for extracting literature data and constructing a database for infrared transparent conductive materials.

[0009] The present invention, a method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model, is implemented according to the following steps:

[0010] I. Literature Data Acquisition and Standardization:

[0011] Keywords and search terms related to infrared transparent conductive materials are entered into the literature database to retrieve literature data. Literature abstracts are used as the source of information extraction texts, and literature identifiers are saved for retrospective analysis. The literature data is then standardized to obtain standardized literature corpus.

[0012] II. Construction of Information Extraction Prompt Template:

[0013] For the standardized literature corpus obtained in step one, set the task description, fields to be extracted, and data types of each field to be extracted, then set rules and establish an information extraction prompt template; the fields to be extracted include infrared transparent conductive materials, performance type, performance value, unit, test conditions, and text source evidence;

[0014] III. Structured Extraction from Large Language Models:

[0015] The standardized literature corpus is combined with the information extraction prompt template and then input into the large language model to output the structured extracted records.

[0016] IV. Structured Record Validation:

[0017] The structured extracted records are first subjected to structure pattern verification, which includes at least one of field hierarchy verification, required field verification, data type verification, and enumeration value verification. The structured extracted records that pass the structure pattern verification are then subjected to material name and performance name normalization, performance value parsing, and unit normalization in sequence to obtain the verified structured extracted records.

[0018] V. Backtracking of Source Evidence and Determination of Confidence:

[0019] The infrared transparent conductive material name, chemical formula, performance value and unit in the verified structured extraction record are matched with the information extraction text source by string matching, normalization matching or position matching, so as to complete the text source evidence backtracking, and then the confidence score is performed on the verified structured extraction record.

[0020] VI. Duplicate Record Detection and Manual Review:

[0021] Duplicate record determination is performed based on at least one of the following: document identifier, material identifier, performance type, performance value, unit, and test conditions; records with the same value under the same document, material, performance type, and test conditions are merged; records that fail to verify the structure pattern, fail to match the text source evidence, have abnormal numerical ranges, abnormal unit conversions, or have a confidence level lower than a preset threshold are manually reviewed to obtain the filtered structured extracted records.

[0022] VII. Building the Database:

[0023] The filtered structured extracted records are written into the infrared transparent conductive materials database, thus completing the method for extracting literature data and constructing the database of infrared transparent conductive materials based on a large language model.

[0024] This invention integrates a field-constrained prompt template for infrared transparent conductive materials, structured extraction of large language models, anchoring of text source evidence, correlation between materials and performance and test conditions, rule verification, manual review, and parallel storage of original and standardized fields into an integrated document data processing workflow.

[0025] This invention utilizes the contextual understanding capabilities of a large language model to map unstructured document text into structured performance records; it restricts field attribution by using the binding relationships between materials, performance, test conditions, and text source evidence; it identifies anomalous records by using structural patterns, material names, numerical ranges, units, and evidence matching rules; it processes low-confidence and conflicting records through manual review; and it enables result traceability and rule updates by using original fields, standardized fields, and model and prompt template version information.

[0026] Compared with existing methods that involve manual data entry or only entity recognition, this invention can improve the efficiency of batch extraction of performance data from documents, reduce erroneous associations between materials, performance values, and test conditions, improve the comparability of performance names and units in different documents, and enable each performance record to be verified based on document identifiers and text source evidence; it also supports manual processing of abnormal records, control of duplicate records, and incremental updates to the database. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of the method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model, as described in this invention.

[0028] Figure 2 This is the interface of the manual annotation software for literature data in the embodiment;

[0029] Figure 3This is a comparison chart of the information extraction performance of the large language model and the named entity recognition method under zero-sample prompting conditions in the embodiments. Figure 3 (a) is a comparison chart of precision, recall, and F1 score for each method. Figure 3 (b) is a comparison chart of the number of incorrectly extracted records on data summaries without a target. Figure 3 (c) A radar comparison chart of precision, recall, F1 score, anti-hallucination rate, and doping information identification indicators. Figure 3 (d) is a comparison chart of F1 values ​​for four types of properties: transmittance, sheet resistivity, resistivity, and mobility.

[0030] Figure 4 This is a comparison chart of information extraction performance between zero-sample and few-sample prompts in the embodiment, where, Figure 4 (a) is a comparison chart of precision, recall, and F1 score. Figure 4 (b) is a comparison chart of the number of incorrectly extracted records on data summaries without a target. Figure 4 (c) A radar comparison chart of precision, recall, F1 score, anti-hallucination rate, and doping information identification indicators. Figure 4 (d) is a precision-recall distribution graph of the results of each method and literature comparison;

[0031] Figure 5 This is a schematic diagram illustrating the data relationships within the database of infrared transparent conductive materials. Detailed Implementation

[0032] Specific Implementation Method 1: This implementation method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model is carried out according to the following steps:

[0033] I. Literature Data Acquisition and Standardization:

[0034] Keywords and search terms related to infrared transparent conductive materials are entered into the literature database to retrieve literature data. Literature abstracts are used as the source of information extraction texts, and literature identifiers are saved for retrospective analysis. The literature data is then standardized to obtain standardized literature corpus.

[0035] II. Construction of Information Extraction Prompt Template:

[0036] For the standardized literature corpus obtained in step one, set the task description, fields to be extracted, and data types of each field to be extracted, then set rules and establish an information extraction prompt template; the fields to be extracted include infrared transparent conductive materials, performance type, performance value, unit, test conditions, and text source evidence;

[0037] III. Structured Extraction from Large Language Models:

[0038] The standardized literature corpus is combined with the information extraction prompt template and then input into the large language model to output the structured extracted records.

[0039] IV. Structured Record Validation:

[0040] The structured extracted records are first subjected to structure pattern verification, which includes at least one of field hierarchy verification, required field verification, data type verification, and enumeration value verification. The structured extracted records that pass the structure pattern verification are then subjected to material name and performance name normalization, performance value parsing, and unit normalization in sequence to obtain the verified structured extracted records.

[0041] V. Backtracking of Source Evidence and Determination of Confidence:

[0042] The infrared transparent conductive material name, chemical formula, performance value and unit in the verified structured extraction record are matched with the information extraction text source by string matching, normalization matching or position matching, so as to complete the text source evidence backtracking, and then the confidence score is performed on the verified structured extraction record.

[0043] VI. Duplicate Record Detection and Manual Review:

[0044] Duplicate record determination is performed based on at least one of the following: document identifier, material identifier, performance type, performance value, unit, and test conditions; records with the same value under the same document, material, performance type, and test conditions are merged; records that fail to verify the structure pattern, fail to match the text source evidence, have abnormal numerical ranges, abnormal unit conversions, or have a confidence level lower than a preset threshold are manually reviewed to obtain the filtered structured extracted records.

[0045] VII. Building the Database:

[0046] The filtered structured extracted records are written into the infrared transparent conductive materials database, thus completing the method for extracting literature data and constructing the database of infrared transparent conductive materials based on a large language model.

[0047] In step two of this implementation method, the fields to be extracted, i.e., the field definitions in the information extraction prompt template, include multiple fields such as material identifier, infrared transparent conductive material composition, infrared transparent conductive material name, dopant parent, dopant element, performance type, performance value, unit, test conditions, preparation conditions, text source evidence, and literature identifier. The material identifier is a unique identifier within the database used to uniquely identify a standardized material record. It can be generated based on standardized chemical formulas, doping relationships, and preset numbers, and is used to link the material information table and the performance record table. The performance type includes at least one of band gap, transmittance, conductivity, resistivity, mobility, effective mass, dielectric constant, sheet resistance, film thickness, and test wavelength.

[0048] In step two of this implementation method, the data type of the field to be extracted is specified as follows: records is an array, material, property, unit, conditions, and context are strings, and value is a numerical value. The rules include rules for the association between materials and properties and test conditions, unit rules, missing value rules, rules prohibiting speculation, and rules for identifying doped materials. Specifically, the rules for the association between materials and properties and test conditions stipulate that records should be created separately for different materials, different properties, or different test conditions; the unit rules stipulate that the original units should be retained and converted in post-processing; the missing value rules and rules prohibiting speculation stipulate that when the text source is not explicitly given, a null value should be output and should not be filled in; the rules for identifying doped materials stipulate that the material field should retain both the parent material and the dopant element; the structured output format stipulates that only a JSON object containing the records array should be output, and an empty array should be output when there is no target data.

[0049] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the literature data in step one includes at least two types of literature information, such as literature title, abstract, keywords, unique document identifier (DOI), publication year, author, and source information.

[0050] Specific Implementation Method 3: This implementation method differs from Specific Implementation Method 1 in that the standardization process in step one includes at least one of the following processing methods: duplicate record determination, language identification, integrity check, relevance filtering, and character standardization.

[0051] The standardization process in this implementation also includes handling methods for Unicode equivalent characters, subscripts and superscripts, hyphens, scientific notation, and invisible spaces as described in the unified documentation.

[0052] Specific Implementation Method Four: This implementation method differs from one of the specific implementation methods one to three in that the task description in step two is limited to extracting only the intrinsic optical, electrical, and transport properties of the infrared transparent conductive material.

[0053] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the large language model in step three is a cloud-based large language model called through a remote interface or a large language model deployed on a local electronic device.

[0054] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the structured extraction of records in step three uses JSON format, XML format, or database record format.

[0055] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the structured extraction record obtained in step three includes at least the following fields: infrared transparent conductive material field, performance type field, performance value field, unit field, test condition field, text source evidence field, and document identifier field.

[0056] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that, in step four, records that fail the structural pattern verification or fail to normalize are transferred to manual review.

[0057] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that the enumeration value verification in step four refers to checking whether the values ​​of the performance type, unit category, data source type, and audit status fields of the infrared transparent conductive material belong to the preset allowed set.

[0058] The field hierarchy verification described in step four of this implementation refers to checking whether the structured extraction result is a preset root object, whether the root object contains a records array, whether each element in the records array is a structured performance record, and whether each field is located at a preset record level. If the level is missing, the nesting position is incorrect, or the records is not an array, the field hierarchy verification is determined to fail.

[0059] Specific Implementation Method 10: This implementation method differs from Specific Implementation Methods 1 to 9 in that the confidence score in step 5 is determined based on the structural pattern verification results, material field verification results, performance value field rationality verification results, unit field verification results, and text source evidence field verification results.

[0060] The structural pattern verification described in this embodiment refers to checking whether the structured extraction result contains a preset records array, whether each element in the array is a structured performance record, whether required fields exist, and whether the data types of each field meet preset requirements. The material field verification refers to checking whether the material field is not empty and conforms to preset material name or chemical formula expression rules. The performance value field rationality verification refers to checking whether the performance value can be parsed into a numerical value and whether it is within the preset reasonable range of the corresponding performance type. The unit field verification refers to checking whether the unit matches the corresponding performance type and can be standardized according to preset unit conversion rules. The text source evidence field verification refers to checking whether the text source evidence is not empty and whether it can be located in the corresponding information extraction text source. The results of the above five verifications are recorded as 0 or 1, with 1 for passing the verification and 0 for failing the verification. The average of the five verification results is used as the confidence score of the structured extracted record.

[0061] Example: This example demonstrates a method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model, implemented according to the following steps:

[0062] I. Literature Data Acquisition and Standardization:

[0063] This embodiment is executed in a Python environment. Keywords related to infrared transparent conductive materials, such as "transparent conducting," "infrared transparent," "transparent electrode," and "conducting oxide," along with their synonyms, are input into a literature database to construct a search query. The search yields 8317 initial literature records. After DOI or standardized title deduplication, English language filtering, removal of records without abstracts, and relevance filtering, 7974 English abstract records are retained. These records include the literature title, abstract, DOI, publication year, and source information. Unicode equivalent characters, subscripts, superscripts, hyphens, scientific notation, and invisible spaces in the abstracts are standardized to obtain a standardized literature corpus.

[0064] II. Construction of Information Extraction Prompt Template:

[0065] An information extraction prompt template is established based on the data types in the literature on infrared transparent conductive materials. The target fields are a set of fields in the information extraction prompt template used to limit the model output content, including at least material, property, value, unit, conditions, and context; performance types include band_gap, optical_band_gap, refractive_index, transmittance, reflection, haze, conductivity, resistance, sheet_resistance, carrier_concentration, mobility, effective_mass, work_function, dielectric_constant, and figure_of_merit. The prompt template stipulates: only the intrinsic optical, electrical, or transport properties of the material explicitly given in the text source are extracted; missing values ​​must not be filled in; values ​​for the same material under different conditions are recorded separately; doping information is retained in the material field; non-target indicators such as device efficiency, photocurrent, IPCE, and grain size are not extracted; an empty records array is returned when there is no target data.

[0066] A positive example is: Input containing "The resistivity of Al-doped ZnO thin films is 4.5 × 10⁻⁶". -4 Ω·cm, carrier concentration of 5.2×1020 cm -3 The summary is "average transmittance is 88%". Output three records for ZnO:Al: resistance, carrier_concentration and transmission, and retain the numerical value, unit and test conditions respectively.

[0067] A negative example is: the input is a summary of a perovskite solar cell with ITO electrodes, showing a photoelectric conversion efficiency of 18.2% and a stabilization time of 500 h. Since the metrics are device performance rather than intrinsic properties of the target material, the output is an empty records array.

[0068] III. Structured Extraction from Large Language Models:

[0069] The standardized literature corpus is combined with the information extraction prompt template and then input into the large language model. This embodiment uses two implementations: a cloud-based large language model called through a remote interface and a large language model deployed on a local electronic device via Ollam. The traditional named entity recognition model is only used as a baseline for comparing information extraction performance. The large language model outputs structured extracted records.

[0070] The cloud-based large language model uses the Deepseek-v4-flash model, with a temperature set to 0, a maximum output length of 2500 tokens, a JSON object as the mandatory output format, and 8 parallel worker threads. Each abstract can only have a maximum of the first 6000 characters input. After parsing the model output into JSON, the document identifier, DOI, original abstract, and records array are written together into a JSONL file, and a pre-built set of document identifiers is used to enable breakpoint resumption.

[0071] The local electronic device's large language model adopts the deepseek-r1:7b model deployed via Ollam, with temperature set to 0, maximum generation length set to 3000 tokens, maximum input of the first 5000 characters for each summary, and a maximum of 3 retries if the call fails.

[0072] IV. Structured Record Validation:

[0073] The structured extracted records undergo structural pattern validation. This embodiment employs field hierarchy validation, required field validation, data type validation, and enumeration value validation: checking if the `records` array exists in the root object and whether its elements are record objects; checking if the `material`, `property`, and `value` fields exist and are not empty; parsing the `value` to a floating-point number and removing records that cannot be parsed; matching the `property` with a preset performance name mapping table, marking unmapped performance names as unknown categories and transferring them to manual review. After the above validations, leading and trailing spaces are removed from the `material` field while retaining doping representations, and the performance names, numerical representations, and units are standardized to obtain the validated structured extracted records.

[0074] V. Backtracking of Source Evidence and Determination of Confidence:

[0075] In this embodiment, text source evidence refers to the short text sentence in the abstract that corresponds to the structured extracted record. The complete abstract is saved as the source text of the evidence. During evidence backtracking, string matching is first performed in the context field. If no direct match is found, normalized matching is performed after unifying Unicode characters, subscripts, superscripts, spaces, hyphens, scientific notation, and unit aliases, requiring that the material name, performance value, and unit be within the same evidence fragment. This embodiment uses a five-rule scoring method to calculate the confidence level, which determines the results of five checks: structure pattern verification, material field verification, performance value field rationality verification, unit field verification, and text source evidence field verification. A pass is recorded as 1, and a fail is recorded as 0. The average of the five check results is used as the confidence score of the structured extracted record. Records with a confidence level below 0.80 or no text source evidence match are transferred to manual review.

[0076] VI. Duplicate Record Detection and Manual Review:

[0077] Duplicate record determination is performed based on at least one of the following: document identifier, material identifier, performance type, performance value, unit, and test conditions; records with the same value under the same document, material, performance type, and test conditions are merged; records that fail to verify the structure pattern, fail to match the text source evidence, have abnormal numerical ranges, abnormal unit conversions, or have a confidence level lower than a preset threshold are manually reviewed to obtain the filtered structured extracted records.

[0078] VII. Building the Database:

[0079] The filtered structured extraction records are written into the infrared transparent conductive material database.

[0080] In this embodiment, a filtered structured extraction record includes the document identifier source_id, DOI, original material name material, normalized performance type property, core performance marker is_core, normalized value value, standard unit unit, unit verification marker unit_flag, test conditions conditions, textual source evidence context, and original value value_raw, thereby completing the method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model.

[0081] In step three of this embodiment, the precision, recall, and F1 score of the large language model are calculated according to the following formulas:

[0082]

[0083] Where TP is the number of correctly extracted records, FP is the number of incorrectly extracted records, and FN is the number of records that exist in the standard test set but were not extracted.

[0084] In this embodiment, the cloud-based model achieves precision, recall, and F1 score of 0.571, 0.596, and 0.584 with zero-shot hints, respectively; after adding the two demonstrations, these scores are 0.605, 0.632, and 0.618. The local model achieves these scores of 0.425, 0.149, and 0.221 with zero-shot hints, and 0.409, 0.412, and 0.410 after adding demonstrations; the baseline scores for traditional named entity recognition are 0.112, 0.105, and 0.109. For identifying miscellaneous information, the cloud-based few-shot model achieves 0.88, while the baseline score for traditional named entity recognition is 0.43.

[0085] Figure 3 This indicates that, on the same manually labeled test set, the precision, recall, and F1 score of the cloud-based large language model are all higher than those of the local large language model and the traditional named entity recognition baseline. Furthermore, the large language model has a higher accuracy in identifying associations of doped materials. Figure 4 The results show that after adding two complete sample examples and strengthening the fields, prohibiting speculation, and imposing output format constraints, the precision, recall, and F1 score of the cloud model improved from 0.571, 0.596, and 0.584 to 0.605, 0.632, and 0.618, respectively; while the precision of the local model improved from 0.425 to 0.409, the recall improved from 0.149 to 0.412, and the F1 score improved from 0.221 to 0.410. This indicates that the prompt template optimization significantly improved the recall and overall extraction performance of the local model, but did not improve its precision.

[0086] In this embodiment, when new literature data is acquired, the language model version and prompt template version registered in the database are used to generate new structured extraction records. Based on the material identifier, performance type, test conditions, and literature identifier, it is determined whether the new structured extraction record is a duplicate of an existing record in the database. Non-duplicate records are written to the database, and their text source evidence, language model version, prompt template version, confidence level, and review status are saved. When the material name standardization rules, unit dictionary, or numerical range rules are updated, standardized fields are regenerated based on the original fields retained in the database, without changing the original data and text source evidence.

[0087] The results of this embodiment show that few-sample demonstration and prohibition inference rules can improve the overall extraction performance of the cloud model and are significantly better than the traditional named entity recognition baseline used. The joint preservation of original fields, standardized fields and text source evidence enables the extraction results to be reviewed, traced and updated incrementally with the rule version.

Claims

1. A method for extracting infrared transparent conductive material literature data based on a large language model and constructing a database, characterized in that The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model is implemented according to the following steps: I. Literature Data Acquisition and Standardization: Keywords and search terms related to infrared transparent conductive materials are entered into the literature database to retrieve literature data. Literature abstracts are used as the source of information extraction texts, and literature identifiers are saved for retrospective analysis. The literature data is then standardized to obtain standardized literature corpus. II. Construction of Information Extraction Prompt Template: For the standardized literature corpus obtained in step one, set the task description, fields to be extracted, and data types of each field to be extracted, then set rules and establish an information extraction prompt template; the fields to be extracted include infrared transparent conductive materials, performance type, performance value, unit, test conditions, and text source evidence; III. Structured Extraction from Large Language Models: The standardized literature corpus is combined with the information extraction prompt template and then input into the large language model to output the structured extracted records. IV. Structured Record Validation: The structured extracted records are first subjected to structure pattern verification, which includes at least one of field hierarchy verification, required field verification, data type verification, and enumeration value verification. The structured extracted records that pass the structure pattern verification are then subjected to material name and performance name normalization, performance value parsing, and unit normalization in sequence to obtain the verified structured extracted records. V. Backtracking of Source Evidence and Determination of Confidence: The infrared transparent conductive material name, chemical formula, performance value and unit in the verified structured extraction record are matched with the information extraction text source by string matching, normalization matching or position matching, so as to complete the text source evidence backtracking, and then the confidence score is performed on the verified structured extraction record. VI. Duplicate Record Detection and Manual Review: Duplicate records are determined based on at least one of the following: document identifier, material identifier, performance type, performance value, unit, and test conditions; records with the same value under the same document, material, performance type, and test conditions are merged. Records that fail to verify structural patterns, fail to match text source evidence, have abnormal numerical ranges, have abnormal unit conversions, or have confidence levels below a preset threshold are manually reviewed to obtain the filtered structured extraction records. VII. Building the Database: The filtered structured extracted records are written into the infrared transparent conductive materials database, thus completing the method for extracting literature data and constructing the database of infrared transparent conductive materials based on a large language model. 2.The method of claim 1, wherein the method further comprises: determining a plurality of data fields from the plurality of documents; and constructing a database based on the plurality of data fields. The literature data in step one includes at least two types of literature information, such as the literature title, abstract, keywords, unique identifier of the literature, publication year, author, and source information.

3. The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model according to claim 1, characterized in that... The standardization process in step one includes at least one of the following methods: duplicate record determination, language identification, integrity check, relevance filtering, and character standardization.

4. The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model according to claim 1, characterized in that... The task description in step two limits the extraction to the intrinsic optical, electrical, and transport properties of infrared transparent conductive materials.

5. The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model according to claim 1, characterized in that... In step three, the large language model is either a cloud-based large language model invoked via a remote interface or a large language model deployed on a local electronic device.

6. The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model according to claim 1, characterized in that... In step three, the structured extraction records are extracted using JSON, XML, or database record formats.

7. The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model according to claim 1, characterized in that... The structured extracted records obtained in step three include at least the following fields: infrared transparent conductive material field, performance type field, performance value field, unit field, test condition field, text source evidence field, and document identifier field.

8. The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model according to claim 1, characterized in that... In step four, records that fail the structural pattern verification or fail normalization are transferred to manual review.

9. The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model according to claim 1, characterized in that... Step four, enumeration value verification, refers to checking whether the values ​​of the performance type, unit category, data source type, and audit status fields of the infrared transparent conductive material belong to the preset allowed set.

10. The method for extracting literature data and constructing a database for infrared transparent conductive materials based on a large language model according to claim 1, characterized in that... In step five, the confidence score is determined based on the results of structural pattern verification, material field verification, performance value field rationality verification, unit field verification, and text source evidence field verification.

Images