A semi-automatic labeling method for a geological text knowledge base based on copper ore mineralization knowledge guidance

Abstract

This invention is applicable to the field of intelligent copper exploration, providing a semi-automatic annotation method based on a geological text knowledge base guided by copper metallogenic knowledge. The method includes the following steps: constructing a three-dimensional copper metallogenic knowledge system; cleaning, professionally segmenting, and structuring the original copper geological text using a copper-specific terminology database to obtain a pre-processed set of structured copper geological text fragments; embedding the three-dimensional copper metallogenic knowledge system into an improved semi-supervised model to automatically annotate the set of structured copper geological text fragments, obtaining preliminary automatic annotation results; filtering highly ambiguous samples from the preliminary automatic annotation results, correcting the selected samples based on the three-dimensional copper metallogenic knowledge system, and outputting corrected and high-precision automatic annotation results; and standardizing the high-precision automatic annotation results. This invention, guided by copper metallogenic knowledge, achieves excellent results in annotation accuracy, efficiency, engineering adaptability of results, and system iteration capabilities through a comprehensive process design.

Description

Technical Field

[0001] This invention relates to the field of intelligent copper exploration, specifically to a semi-automatic annotation method for a geological text knowledge base guided by copper mineralization knowledge. Background Technology

[0002] Artificial intelligence (AI) mineral exploration technology has become a core means to overcome challenges in deep and blind mineral exploration. Geological text knowledge bases, as the core carriers of ore-controlling elements, mineralization patterns, and exploration experience, directly determine the learning effectiveness of AI mineral exploration models in understanding mineralization logic through their annotation quality. However, copper ore geological texts are characterized by dense technical terms, long sentences implying mineralization logic, and numerous semantic ambiguities, placing extremely high demands on the professionalism and accuracy of text annotation.

[0003] Currently, there are three main types of geological text annotation technologies, but none of them can meet the specific needs of copper mining scenarios. First, the general semi-supervised annotation method based on BERT, while employing a semi-supervised learning approach of a small amount of manual annotation and a large number of pseudo-labels, completely fails to integrate copper mineralization knowledge and cannot recognize copper-specific semantics, resulting in a high false positive rate in copper mine text annotation. Second, the general geological terminology annotation system, although incorporating a general geological terminology database, lacks typical copper mineralization patterns and specific rules, and can only annotate isolated terms, failing to capture the mineralization correlation logic between rock masses, faults, alteration, and mineralization. Third, the simple rule-based annotation method for mineral texts relies solely on keyword string matching without any semantic understanding, easily mislabeling non-mineralized information as positive samples, resulting in extremely poor annotation accuracy.

[0004] In summary, existing technologies suffer from several shortcomings: a lack of adaptation to copper mineralization knowledge leads to annotation logic that is detached from exploration realities; shallow text semantic understanding fails to handle the specialized characteristics of copper ore texts; the human-machine collaboration mechanism is inefficient, making it difficult to balance the accuracy and efficiency of large-scale annotation; and insufficient standardization of annotation results results in poor compatibility with subsequent AI-based mineral exploration models. These shortcomings collectively constrain the quality and efficiency of constructing a copper ore geological text knowledge base, becoming a core bottleneck restricting the intelligent upgrading of copper ore exploration. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a semi-automatic annotation method for geological text knowledge base guided by copper mineralization knowledge, so as to solve the problems existing in the above-mentioned background technology.

[0006] This invention is implemented as follows: a semi-automatic annotation method for a geological text knowledge base guided by copper ore-forming knowledge, the method comprising the following steps: A three-dimensional copper mineralization knowledge system is constructed, which consists of a copper mineralization rule base, a copper-specific terminology base, and a copper mineralization correlation map. The original copper mine geological text was cleaned, professionally segmented, and structured using a copper mine-specific terminology database, resulting in a pre-processed set of structured copper mine geological text fragments. By embedding a three-dimensional copper mineralization knowledge system into an improved semi-supervised model, preliminary automatic annotation results are obtained by automatically annotating a set of structured copper geological text fragments. The preliminary automatic annotation results are screened for highly ambiguous samples. Based on the three-dimensional copper metallogenic knowledge system, the screened samples are corrected and the corrected and high-precision automatic annotation results are output. The high-precision automatic annotation results are standardized into three categories: basic annotation, quantitative annotation, and correlation annotation, so that they can be used as training data for AI mineral exploration models.

[0007] As a further aspect of the present invention: the step of constructing a three-dimensional copper ore-forming knowledge system specifically includes: By selecting representative deposit cases from various typical copper mines, and using a structured format that combines conditions to conclusions, a copper mineralization rule library is constructed. By extracting terms from copper mining literature, classifying them into mineral type terms, alteration terms, ore-controlling element terms, and element anomaly terms, and assigning weights to each type of terms related to mineralization, a copper mining-specific terminology library is constructed. Using mineralization events as the core nodes, the terms controlling ore elements are mapped to two types of nodes: ore-forming rock bodies and ore-controlling structures. The terms controlling mineral types are mapped to mineralization type nodes. Alteration type nodes corresponding to alteration terms and element anomaly nodes corresponding to element anomaly terms are connected, and the correlation weights between nodes are labeled to construct a copper mineralization correlation map.

[0008] As a further aspect of the present invention: the step of constructing the copper ore mineralization rule base specifically includes: Obtain representative deposit case data from various typical copper mines; Based on the representative mineral deposit case data, a structured format from condition combination to conclusion is adopted to form mineralization rules. The condition combination includes rock mass type, alteration combination, element content threshold, and tectonic location parameters. The mineralization rules are manually verified from four dimensions: rationality, accuracy, completeness, and adaptability. After verification, the mineralization rules are entered into the copper mineralization rule database, and an index is created for searching by copper ore type and condition combination.

[0009] As a further aspect of the present invention: the steps of cleaning, professionally segmenting, and structuring the original copper mine geological text using a copper mine-specific terminology database specifically include: The original copper mine geological text was converted into plain text format, page marks, table and image marks and editing marks were removed, redundant information was deleted, and a copper mine-specific terminology library was used in conjunction with a spelling correction algorithm to jointly correct optical character recognition errors. A hybrid word segmentation model that embeds features from a copper mine-specific terminology database is used for word segmentation to ensure the integrity of copper mine-specific compound terms. The segmented text is divided into four types of structured fragments: description of ore-controlling structures, description of alteration, description of elemental content, and description of mineralization phenomena. These fragments are then uniformly labeled and classified for storage, resulting in the set of structured copper ore geological text fragments.

[0010] As a further aspect of the present invention: the step of using a copper mine-specific terminology database in conjunction with a spelling correction algorithm to jointly correct optical character recognition errors specifically includes: The terms to be corrected are extracted from the cleaned original copper mine geological text. The terms to be corrected are compared word by word with the copper mine-specific terminology database. If they match perfectly, they are automatically corrected. If the match is not perfect, the text similarity is calculated using the edit distance algorithm. When the similarity reaches a set threshold, the contextual semantics are considered to determine whether to automatically correct the text. The terms that cannot be corrected are sorted according to their term weight priority and marked as items to be reviewed.

[0011] As a further aspect of the present invention: the step of embedding a three-dimensional copper mineralization knowledge system into an improved semi-supervised model to automatically annotate a set of structured copper geological text fragments to obtain preliminary automatic annotation results specifically includes: The three-dimensional copper ore-forming knowledge system is embedded into the improved semi-supervised model to obtain the initialized model; Based on a set of structured copper mine geological text fragments, copper mine-specific terms in the text are identified through an initialized model, and corresponding candidate labels are generated based on the identification results. The candidate tags are precisely verified condition by condition based on the copper mineralization rule library, and the implicit mineralization association tags are supplemented based on the copper mineralization association map to generate enhanced pseudo tags. Obtain the confidence score and knowledge matching score generated by the initialized model for each pseudo-label; The system uses a strategy that the confidence level is greater than or equal to a first preset threshold and the knowledge matching level is greater than or equal to a second preset threshold to filter out pseudo-labels and generate preliminary automatic labeling results.

[0012] As a further aspect of the present invention: the step of embedding the three-dimensional copper ore-forming knowledge system into the improved semi-supervised model specifically includes: Each structured rule in the copper ore mineralization rule base is converted into a logical expression; During the semi-supervised model training process, the predicted labels of the intermediate output of the model are extracted, and the predicted labels are compared with the logical expressions one by one to calculate the matching degree. Gradient penalty is applied to the model parameters corresponding to the predicted labels with matching degrees lower than a preset threshold. The normalization coefficients are obtained from the copper mine-specific terminology database. The normalization coefficients are then multiplied dimension-by-dimensionally with the basic word vectors in the word embedding layer of the semi-supervised model to obtain weighted word vectors. The normalization coefficients are obtained by normalizing the numerical value of the correlation between each term in the copper mine-specific terminology database and copper mineralization. A composite loss function is constructed, which includes a basic cross-entropy loss and a mineralization rule constraint loss term, wherein the mineralization rule constraint loss term is dynamically calculated based on the matching degree.

[0013] As a further aspect of the present invention: the step of screening highly ambiguous samples from the preliminary automatic annotation results and correcting the screened samples based on the three-dimensional copper mineralization knowledge system specifically includes: Highly ambiguous samples requiring manual correction are selected from the initial automatic annotation results and sorted according to preset priority rules. The system receives correction instructions for highly ambiguous samples after language sorting through human-computer interaction. The correction instructions include confirmation, rejection, supplementation, or modification of the automatic annotation results. The corrected samples are used to update the three-dimensional copper mineralization knowledge system and to train the improved semi-supervised model.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a three-dimensional copper mineralization knowledge system, providing professional support for subsequent annotation that aligns with the metallogenic laws of copper deposits. This addresses the fundamental shortcomings of existing technologies, such as the lack of copper-specific knowledge and the inability to identify mineralization logic. By utilizing a copper-specific terminology database for cleaning, word segmentation, and structuring, it effectively corrects optical character recognition errors, ensures that composite copper-related terms are not fragmented, and divides the text into four types of structured segments, laying a high-quality data foundation for accurate annotation. The three-dimensional knowledge system is embedded into an improved semi-supervised model for automatic annotation, enabling the model to understand copper mineralization logic. The generated preliminary annotation results significantly outperform general models in terms of terminology recognition and mineralization association capture. Through screening highly ambiguous samples and correcting them based on the knowledge system, only a small number of key samples require manual intervention, significantly shortening the annotation cycle. The correction results can also be fed back to update the knowledge system and model, achieving continuous iterative optimization. The annotation results are standardized into three categories: basic annotation, quantitative annotation, and correlation annotation, directly adapting to the training needs of AI mineral exploration models. This establishes a complete link from text annotation to engineering applications, enhancing the practical value of the results. In summary, this invention, guided by copper mineralization knowledge, has achieved excellent results in terms of annotation accuracy, efficiency, engineering adaptability of results, and system iteration capabilities through a complete process design, providing practical and effective support for intelligent copper exploration. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the overall process of a semi-automatic annotation method for a geological text knowledge base guided by copper mineralization knowledge. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] like Figure 1 As shown, this embodiment takes the geological text of a typical copper mine exploration report in my country as the processing object, and details the specific implementation process of the semi-automatic annotation method of the geological text knowledge base guided by copper metallogenic knowledge.

[0018] Step 1: Constructing a three-dimensional knowledge system of copper mineralization (1) Constructing a copper ore mineralization rule base 1) Obtain representative deposit case data of various typical copper mines. Collect core data such as metallogenic geological background, ore-controlling elements, alteration assemblages, and elemental anomaly characteristics of various copper mines to form a raw metallogenic data set.

[0019] 2) A structured format of "condition combination → conclusion" is adopted to form mineralization rules. The condition combination covers rock mass type, alteration combination, Cu content threshold, and tectonic location parameters, while the conclusion clarifies the mineralization potential level and mineralization type. The specific rules are shown in the table below (only some rules are shown; the rest are analogous):

[0020] 3) Manually verify the compiled metallogenic rules. In practice, a verification team of three or more senior engineers in the field of copper exploration can be invited to conduct verification from four dimensions: rationality, accuracy, completeness, and suitability, based on the actual metallogenic characteristics of 12 typical copper deposits in my country, publicly available metallogenic theories, and representative deposit exploration data. For the rationality verification, each engineer, in conjunction with typical deposit cases of the corresponding copper ore type (such as porphyry type corresponding to Jiangxi Dexing Copper Mine, skarn type corresponding to Hubei Daye Copper Mine), reviews the logical relationship between the "condition combination" and "conclusion" in a single rule, and judges whether the combination of conditions such as rock mass type, alteration combination, Cu content threshold, and tectonic location conforms to the actual mineralization law of the copper ore. For accuracy verification, engineers checked the accuracy of the core parameters in the rules one by one against the standard manual for copper mining, typical deposit exploration reports and authoritative metallogenic literature, including Cu content threshold, adaptability of alteration combination, and accuracy of rock mass type. Completeness verification: The verification team held a collective discussion to review whether the mineralization rules of the 12 types of copper mines fully cover the core mineralization scenarios of each type of copper mine, and whether there are any missing mineralization conditions or conclusions. Adaptability verification involves reviewing whether the "condition combination" of each rule facilitates semantic matching of the text and whether the "conclusion" can directly support the generation of annotation tags, taking into account the needs of subsequent knowledge-guided annotation.

[0021] 4) The verified mineralization rules (a total of 86) are entered into the rule base, and a rule index is established to support quick retrieval and retrieval by copper ore type and condition parameters. It should be noted that, taking porphyry copper deposits (Jiangxi Dexing Copper Mine) as an example, eight mineralization rules can be constructed to cover mineralization scenarios with different rock mass subtypes, alteration assemblages, Cu content thresholds, and tectonic locations. Due to space limitations, only one rule is used as an example here. The specific distribution is shown in the table below, which can fully support accurate matching in the subsequent knowledge-guided annotation stage.

[0022]

[0023] (2) Construct a copper mine-specific terminology database 1) Using the TF-IDF algorithm, the terminology was initially extracted from copper ore literature using eight categories of keywords: "copper ore, mineralization, alteration, rock mass, structure, element, ore deposit, and mineral type".

[0024]

[0025] 2) The initially extracted terms were systematically classified and organized. The classification logic strictly follows the core elements of the entire copper mineralization process. Combining the functional positioning of the terms in the mineralization law and annotation scenarios, they are clearly divided into four major categories: mineral type terms, alteration terms, ore-controlling element terms, and element anomaly terms. The boundaries of each category are clear, comprehensive, and without overlap or redundancy, with a total of over 1200 terms. The specific classification and details are as follows: Mineral type terms focus on copper ore types and deposit characteristics, clarifying the genesis and deposit attributes of different copper ores; Alteration terms revolve around alteration processes in copper mineralization, covering various alteration types and intensity descriptions related to copper mineralization; Ore-controlling element terms cover the core geological elements affecting copper mineralization, including key carriers such as rock masses, structures, and strata; Element anomaly terms focus on the element content and anomaly characteristics related to copper mineralization, providing support for the determination of mineralization potential. The specific number, core definitions, and examples of each type of term are shown in the table below.

[0026]

[0027] 3) Systematically weight each type of term, with a uniform weight range of 0.1-1.0. The overall principle is that "the closer the connection to copper mineralization, the higher the priority of the labeling, and the greater the weight value." Based on the functional positioning of each term in copper mineralization patterns and its practical value in the labeling scenario, different weight ranges are defined: Mineral type terms directly relate to copper ore type and core mineralization characteristics, and are the core focus of labeling, with a weight range of 0.7-1.0; Alteration terms and elemental anomaly terms are direct indicators of copper mineralization, corresponding to mineralization processes and mineralization potential assessments, respectively, with a weight range of 0.6-0.9; Ore-controlling element terms are the basic support for copper mineralization, covering core carriers such as rock masses and structures, with a weight range of 0.5-0.8; For general geological terms that are not specific to copper ore and have a weaker connection to mineralization (such as common sandstone, regional stratigraphic naming, etc.), the weight is uniformly set to 0.1-0.3. The detailed weight settings for each type of term are shown in the table below:

[0028] 4) The terminology database was deduplicated and corrected to create a copper mine-specific terminology database containing over 1200 terms, stored in XML format, and supporting dynamic adjustment of term weights. The deduplication process employed a dual approach of "automatic matching and deduplication + manual review and confirmation."

[0029] (2) Constructing a metallogenic correlation map of copper deposits 1) Taking the mineralization event as the core node, the mineralization rock mass, ore-controlling structure, alteration type, element anomaly, and mineralization type are used as secondary nodes, and each node is associated with the corresponding term in the terminology library.

[0030] 2) Based on the copper ore mineralization rule base (86 structured rules), the relationships between nodes were analyzed, clearly distinguishing between direct and indirect relationships. Each relationship was labeled with a weight (0.1-1.0). Direct relationships (weight ≥ 0.7) are core mineralization relationships; indirect relationships (weight 0.3-0.6) are auxiliary mineralization relationships; and unrelated elements (weight < 0.3) were directly eliminated. This resulted in over 320 valid relationship edges, covering the mineralization logic of 12 types of copper deposits. See the table below: 3) A visual association graph is constructed using the Neo4j graph database, and the graph is synchronously stored in XML format to support real-time retrieval in subsequent automatic annotation stages.

[0031] 4) Using 12 typical copper mines in my country (such as the Dexing porphyry copper mine in Jiangxi and the Daye skarn copper mine in Hubei) as verification samples, the nodes and relationships in the graph were matched one by one. The rationality of the association logic was verified, erroneous associations were corrected, missing associations were supplemented, association weights were optimized, and weight values ​​were adjusted to ensure that the graph can accurately reflect the internal logic of copper mine formation and support the identification and extraction of the implicit mineralization logic in the text.

[0032] Step 2: Processing the original copper mine geological text (1) Text cleaning 1) Standardize the format of the original copper mine geological text by converting PDF, DOCX and other formats to TXT format, removing formatting marks (page-related marks, table and image-related marks, editing-related marks and redundant control marks) and extracting the plain text content.

[0033] 2) Remove redundant information and accurately select the core text content directly related to copper mineralization. Thoroughly delete all redundant paragraphs, sentences, and fragmented information unrelated to copper mineralization. Based on the actual characteristics of copper geological texts (exploration reports, borehole logging, etc.), redundant information specifically includes administrative redundancy (administrative and management content unrelated to copper mineralization technology), personnel and equipment redundancy (personnel and equipment-related content unrelated to copper mineralization information), irrelevant technical redundancy (technical content not belonging to the field of copper mineralization and unrelated to copper geological characteristics), and repetitive and invalid redundancy (content that appears repeatedly in the text and fragmented information without practical meaning). Each category has a clearly defined core scope and examples to ensure that redundant information is identified unambiguously and deleted without omission.

[0034] 3) Correcting OCR recognition errors: A combined error correction approach using a "copper mine-specific terminology database + SpellChecker spelling correction algorithm" is employed. Detailed error correction rules are defined to ensure the accuracy and operability of error correction. Specific correction rules and execution logic are as follows: Terminology matching rules: Full matching is prioritized, comparing identified suspected erroneous terms with 1200+ categorized terms in the copper mine-specific terminology database word by word. If a completely identical term is found (e.g., the misidentified "huangtong kuang" perfectly matches "huangtong kuang" in the terminology database, ignoring the typo), it is directly determined as an error that can be corrected. If full matching fails, fuzzy matching is performed, calculating the text similarity between the suspected erroneous term and the terms in the database (using the edit distance algorithm). A similarity threshold of ≥85% is set as a valid fuzzy match; if it is lower than this threshold, it is determined that it cannot be corrected by terminology matching. Automatic correction judgment rules: Suspected erroneous terms that are completely matched will be automatically corrected directly, and a correction log (including the original erroneous text, the corrected term, and the matching type) will be recorded synchronously; suspected erroneous terms that are fuzzy matched (similarity 85%-94%) need to be further verified in conjunction with the context semantics. If the context contains copper mineralization elements related to the term (such as "Huangtong ore" containing expressions such as "Cu anomaly" and "alteration"), then automatic correction will be performed; otherwise, it will be marked as pending manual review; fuzzy matched terms with similarity ≥95% do not require context verification and will be automatically corrected directly. Manual review rules: Errors that cannot be automatically corrected (including fuzzy matching similarity < 85%, fuzzy matching 85%-94% with no effective contextual support, and SpellChecker algorithm identifying them as spelling errors but without corresponding terminology matches) are uniformly marked as items to be manually reviewed. At the same time, they are sorted according to the priority of term weight (high-weight terms (weight ≥ 0.7) are reviewed first, such as core terms like "porphyry copper deposit" and "skarnification"). During manual review, the terminology definitions and relationships in the copper mineralization knowledge system must be referenced to confirm the correction plan. After the review is completed, the correction results and review opinions are entered to form a complete error correction closed loop. Among them, the error of automatic correction must meet the correction accuracy rate ≥ 99%, and the error of manual review must achieve 100% correction to ensure the overall OCR error correction quality.

[0035] 3) The cleaned text is systematically deduplicated. Based on the professional characteristics of copper mine geological text, clear and executable deduplication rules are formulated to ensure the uniqueness of the text while retaining core mineralization information.

[0036] (2) Professional word segmentation optimization 1) Construct a professional word segmentation model A professional word segmentation model adapted to copper mine geological text annotation scenarios is built based on the BERT-BiLSTM-CRF hybrid model. It is specifically optimized for the characteristics of copper mine geological text, such as "dense technical terms, many long sentences related to mineralization, and blurred terminology boundaries". The functions of each module are all developed around the specific word segmentation requirements of copper mines. Based on the BERT model, the features of the copper mine-specific terminology database constructed in step 1 are incorporated. The weight information of more than 1,200 copper mine-specific terms (including four categories: mineral type, alteration, and ore-controlling elements) is embedded into the word vector representation of the BERT model. This enables the model to prioritize the identification of high-weight copper mine-specific terms (such as "porphyry copper mine" and "copper-molybdenum mineralization zone"), avoiding the misidentification of copper mine professional terms by general word segmentation models. The integration of the BiLSTM model addresses the challenge of segmenting long sentences in copper mine geological texts. Copper mine exploration reports and borehole logging texts often contain long sentences containing "rock type + alteration combination + element content" (such as "quartz diorite rock body is surrounded by sericitization and potassium feldspar alteration, with associated Cu content ≥100ppm anomaly"). The BiLSTM model captures long-distance semantic dependencies and accurately segments mineralization-related terms and non-term components in such long sentences, avoiding segmentation bias caused by excessively long sentences. Integrating the CRF model focuses on optimizing the boundary recognition of copper mine-specific terms. Addressing the issue that copper mine terms are often complex structures (such as "skarn-porphyry composite copper mine" and "weakly silicified carbonatization") and easily segmented, the CRF model learns the boundary feature patterns of copper mine terms to constrain the rationality of the word segmentation results. This ensures that complex copper mine-specific terms are not mistakenly segmented into single Chinese characters or general terms, laying the foundation for accurate term recognition and rule matching in the subsequent knowledge-guided automatic annotation process.

[0037] 2) Model fine-tuning training The word segmentation model is deeply embedded with a copper mine-specific terminology database (containing over 1200 professional terms categorized and weighted across four main categories: mineral type, alteration, ore-controlling elements, and elemental anomalies). Targeted fine-tuning training is conducted to align with the core logic of this patent, "copper mine metallogenic knowledge-guided annotation." This differs from the generalized fine-tuning of general word segmentation models; the entire process is designed around the precise word segmentation requirements of copper mine-specific terminology. Training samples are rigorously selected from core processing objects of this patent, such as copper mine exploration reports and borehole logging texts, choosing copper mine geological text fragments that have already been manually and precisely segmented and annotated. (Sample size ≥ 1000 entries) The samples must comprehensively cover the core proprietary terms of 12 typical copper mines, with a focus on high-weight terms (weight ≥ 0.7, such as porphyry copper mines, skarnification, Cu anomalies, etc.) and composite copper mine terms (such as copper-molybdenum mineralization zones, skarn-porphyry composite copper mines), ensuring that the samples closely match the actual application scenarios of the patents; fine-tune the training parameters to specifically adapt to the word segmentation requirements of copper mine terms, set the number of training iterations to 80-100 rounds, and adopt a step-decay strategy for the learning rate (initial learning rate 1e-5, every 2... The training process employs a 50% decay rate (0 rounds), utilizing the AdamW optimizer to reduce overfitting. A copper ore terminology integrity constraint is incorporated into the training process, forcing the model to learn the completeness of copper ore-specific composite terms and preventing the splitting of complete terms. The training objective is clearly focused on improving the segmentation accuracy of copper ore-specific terms, addressing the pain points of general segmentation models that easily split copper ore composite terms and misclassify copper ore-specific terms with general geological terms. For example, it ensures that "copper-molybdenum mineralization zone" is correctly segmented as "copper-molybdenum mineralization zone" rather than "copper / molybdenum / mineralization / zone," and that "skarn-porphyry composite copper ore" is not split. Simultaneously, it accurately distinguishes between copper ore-specific terms (such as chalcopyrite) and general geological terms (such as ordinary sandstone), laying a core foundation for accurate terminology recognition and rule matching in subsequent knowledge-guided automatic annotation. During fine-tuning, validation is performed every 10 iterations. The validation set uses copper ore geological text segmentation samples that did not participate in the training (accounting for 20% of the total samples). The segmentation accuracy of copper ore-specific terms is monitored in real time until the model converges (segmentation accuracy stabilizes at ≥96%), at which point fine-tuning stops.

[0038] 3) Labeling of proprietary terms The cleaned text is segmented using a finely tuned word segmentation model, which breaks the text down into individual words and labels the part of speech (noun, verb, adjective, etc.) of each word, with a focus on copper mine-specific terms. The labeling process strictly relies on the existing copper mine-specific terminology database to achieve accurate matching and standardized labeling of terms.

[0039] 4) Word segmentation result verification A layered and refined verification process, combining automatic validation and manual review, was implemented for the word segmentation results. This focused on the accuracy of word segmentation for copper mining-specific terms, ensuring that the results could directly support subsequent knowledge-guided automatic annotation. The specific verification process and implementation details are as follows: The core objectives of the verification are: closely adhering to the core of this patent, "Guided Labeling of Copper Ore Knowledge," and focusing on verifying two key dimensions: First, the completeness of copper ore-specific terminology, ensuring that complex copper ore-specific terms such as "porphyry copper deposit," "copper-molybdenum mineralization zone," and "skarn-porphyry composite copper deposit" are not split, and that single copper ore terms such as "chalcopyrite" and "sericification" are accurately identified; Second, the accuracy of terminology classification, ensuring that copper ore-specific terms are accurately distinguished from general geological terms (such as ordinary sandstone and shale) and non-geological terms, without confusion or misclassification, laying the foundation for subsequent terminology weight matching and metallogenic rule verification.

[0040] Automatic verification implementation: Based on the constructed copper-specific terminology database (1200+ categorized terms) and metallogenic association map, an automatic verification engine is built to perform three-step verification: First, terminology integrity verification: the word segmentation results are fully matched with the copper-specific terminology database, and composite term fragments that do not match completely (such as "skarn-porphyry" mistakenly split into "skarn / porphyry") are screened out and marked as integrity errors; Second, terminology classification verification: the "copper-specific terms" labeled after word segmentation are verified a second time to confirm that their categories (mineral type, alteration, ore-controlling elements, elemental anomalies) are completely consistent with the terminology database, and general geological terms that are mistakenly labeled as copper-specific terms are removed (such as "ordinary sandstone" being mistakenly labeled as an ore-controlling element term); Third, semantic consistency verification: combined with the copper metallogenic association map, the semantic relevance between the word segmentation results and the context is verified (such as after segmenting "fault zone", the context should contain relevant copper metallogenic expressions such as ore-controlling and alteration, otherwise it is marked as semantic ambiguity error).

[0041] Manual review implementation: For error samples marked by automatic verification, a "high-value priority" principle is adopted to select samples for manual review. Focusing on the core needs of copper ore labeling, two types of samples are prioritized for review: First, word segmentation error samples containing high-weight copper ore terms (weight ≥ 0.7, such as porphyry copper ore, skarnification, Cu anomaly), as these samples directly affect the accuracy of subsequent mineralization rule matching; Second, word segmentation error samples from deep exploration reports and texts related to typical copper ore deposits (porphyry and skarn), as these samples have high mineralization value, and the accuracy of word segmentation directly determines the quality of the labeling results. During manual review, the error samples are corrected by referring to the copper ore mineralization knowledge system (terminology library definitions, association map relationships), such as merging split compound terms, correcting term classifications, clarifying semantic ambiguities, and simultaneously recording the error type (completeness error, classification error, semantic error) and the basis for correction.

[0042] Accuracy Verification and Feedback Iteration: After verification, the "accuracy rate of copper mine-specific terminology segmentation" is used as the core evaluation indicator (calculated as: number of accurately segmented copper mine-specific terms ÷ total number of all copper mine-specific terms in the text × 100%), and this accuracy rate must be ≥ 96%. For paragraphs that do not meet the standard, the above automatic verification + manual review process is repeated until the accuracy requirement is met. At the same time, the manually corrected segmented samples are used as new training data and fed back to the BERT-BiLSTM-CRF segmentation model constructed above for secondary fine-tuning (20-30 training iterations, learning rate 5e-6), optimizing the model's ability to recognize copper mine-specific terms, reducing subsequent segmentation errors, and forming a closed loop of "segmentation-verification-correction-model optimization", which is in line with the core design logic of "human-machine collaboration and iterative optimization" of this patent.

[0043] (3) Text structuring 1) Define the criteria for determining structured text fragments. First, based on the copper mineralization logic, four types of structured text fragments are pre-defined: ore-controlling structure description, alteration description, element content description, and mineralization phenomenon description. The criteria for judging each type of fragment are clarified. For example, the ore-controlling structure description includes relevant expressions such as fault zones, folds, and intrusive rock contact zones, while the alteration description includes various copper mineral alteration terms and relevant expressions such as alteration range and intensity. 2) Text structured segmentation A text classification algorithm (CNN text classification algorithm) is used, combined with a copper ore-forming association map, to segment the word-segmented text. The segmentation process follows the core logic of "feature extraction - association matching - type determination - threshold filtering" to ensure that the segmentation results accurately match the four preset structured segment types. The specific operation steps are as follows: Feature extraction: Based on the convolutional layer of the CNN text classification algorithm, local feature extraction is performed on the segmented text sequence, focusing on extracting core features related to copper mineralization, including copper-specific terms (including category and weight) annotated after segmentation, term combination features (such as "rock mass terminology + alteration terminology" and "element terminology + content description"), and semantic context features. At the same time, the correlation features in the copper mineralization correlation map in step 1 are introduced, and the correlation weight between terms (such as the correlation weight of "fault zone-mineralization zone" is 0.9) is used as auxiliary features and integrated into the feature vector to improve the professionalism and relevance of feature extraction and avoid extracting redundant features unrelated to copper mineralization.

[0044] Association Matching: The extracted text feature vectors are matched with feature templates of four preset types of structured fragments (description of ore-controlling structures, alteration description, elemental content description, and mineralization phenomenon description). At the same time, the correlation between the combination of terms in the feature vectors and the ore-forming logic is verified by combining the copper mineralization correlation map. For example, if the feature vector contains terms and related relationships such as "fault zone", "ore-conducting", and "ore-hosting", the "ore-controlling structure description" template is matched first; if it contains alteration terms and intensity descriptions such as "skarnification" and "sericification", the "alteration description" template is matched first; if it contains element-related terms such as "Cu content" and "ppm", the "elemental content description" template is matched first; if it contains mineralization-related terms such as "copper mineralization" and "porphyry type", the "mineralization phenomenon description" template is matched first.

[0045] Type Determination: Based on the matching results, a multi-classification determination rule is adopted to determine the type of each segmented text fragment. Four priority levels are set for fragment types, with "description of mineralization phenomenon" and "description of element content" having the highest priority (these fragments are directly related to the core information of copper mineralization), followed by "description of alteration" and "description of ore-controlling structures". If a text fragment matches multiple type templates at the same time, such as containing both alteration and mineralization terms, it is determined to be of higher priority based on the "alteration-mineralization" association logic in the copper mineralization association map, such as being determined as "description of mineralization phenomenon" first, and the alteration-related association information is added in the subsequent structured tagging. If a fragment has no copper mineralization-related features, it is determined to be a non-core fragment and is not included in the structured fragment set for the time being, and will be uniformly removed later.

[0046] Threshold filtering: A preset classification confidence threshold (e.g., ≥0.8) is set. The CNN text classification algorithm outputs the confidence score of each segment matching the corresponding type template, and only segments with a confidence score ≥0.8 are retained as valid structured segments. For suspected segments with a confidence score between 0.5 and 0.8, a secondary verification is performed in conjunction with the copper mineralization correlation map. If the terminology correlation in the segment conforms to the mineralization rules (e.g., the correlation between "fault zone + alteration" conforms to the mineralization control logic), it is included in the corresponding type of structured segment. If the confidence score <0.5 or the secondary verification does not conform to the mineralization logic, it is judged as an invalid segment and is removed. Finally, the valid segments are split into structured segments that conform to the preset type.

[0047] 3) Structured fragment annotation Each structured fragment is marked, and the marking operation strictly follows the logic of "verification first, marking then, and archiving" to ensure that the marked information is complete, standardized, and traceable, and that it fully conforms to the four types of structured fragments (ore-controlling structure description, alteration description, element content description, and mineralization phenomenon description) set in the previous text. This provides clear data support for subsequent knowledge-guided automatic marking, manual correction, and standardization of results.

[0048] 4) Structured fragment classification and storage Based on the results of word segmentation verification and fragment labeling, and according to the four types of structured fragments (description of ore-controlling structures, description of alteration, description of elemental content, and description of mineralization phenomena) preset in the previous text, standardized classification and storage operations are performed to form a set of structured copper ore geological text fragments that can be directly called, laying the foundation for accurate annotation in the subsequent knowledge-guided automatic annotation process.

[0049] Step 3: Knowledge-guided automatic annotation This step deeply integrates the constructed copper mineralization knowledge system with the improved semi-supervised model. Through knowledge injection, the model acquires the ability to understand copper mineral semantics, solving the problem of poor recognition and mislabeling of copper mineral terms in general models. Then, pseudo-labels are generated, and high-quality pseudo-labels are selected by dual thresholds. While ensuring the efficiency of automatic labeling, the false positive rate is significantly reduced, providing a reliable foundation for subsequent human-machine collaborative correction.

[0050] (1) Knowledge injection initialization An improved semi-supervised model (such as the improved BERT model) is selected, and a copper mineralization rule constraint layer is added. This constraint layer contains three modules: a rule parsing module that converts structured mineralization rules (condition combination → conclusion) into logical expressions; a constraint matching module that verifies the matching degree between the model's intermediate outputs and the logical expressions in real time; and a loss penalty module that applies gradient penalties to outputs with a matching degree lower than a preset threshold. Through this constraint layer, the model actively avoids generating false labels that contradict the copper mineralization rules during training. The term weights from the copper mine-specific terminology library are embedded into word vectors. After normalizing the term weights, they are multiplied dimension-wise with the base word vectors of the BERT model to obtain weighted word vectors. This operation prioritizes the identification of high-weight terms (such as "porphyry copper deposit") by the model, while weakening the influence of low-weight general terms, thereby improving the recognition accuracy of copper mine-specific semantics. Constructing a composite loss function ,in Based on cross-entropy loss, This is the loss term constrained by the mineralization rules (dynamically calculated based on the matching degree between the model output and the rules). A loss function of 0.8 was chosen. This loss function, while ensuring basic classification capabilities, imposes additional penalties on outputs that do not conform to mineralization patterns, guiding the model to output annotation results that better reflect the realities of copper mine exploration. Furthermore, the model was fine-tuned using 10%-20% of the total sample data, comprised of manually annotated copper-specific samples, enabling it to understand the mineralization rules, terminology, and the "rock mass-alteration-element-mineralization" relationships of 12 typical copper deposits.

[0051] (2) Pseudo-tag generation The structured text fragments are input into the initialized model. The model calls a copper-specific terminology database to identify copper-specific terms in the text and generates preliminary candidate labels. For example, if "chalcopyrite" and "Cu content 200ppm" are identified, candidate labels such as "copper mineralization" and "Cu anomaly" are generated. The candidate labels are then precisely validated condition by condition using a copper mineralization rule database: the mineralization rules corresponding to the candidate label type are located, and the rock mass type, alteration combination, element content threshold, and tectonic location in the text are matched point by point with the conditions in the rules. Candidate labels that are completely matched or whose core conditions match and whose secondary conditions are not contradictory are retained; otherwise, they are discarded. This step effectively filters out mislabeling caused by semantic ambiguity (such as mislabeling "iron mineralization" as "copper mineralization"). Implicit labels are supplemented based on the copper mineralization association map: through the logic of "terminology matching → association location → label adaptation," implicit mineralization information in the text is identified. For example, if the text mentions "fault zone," it adds the label "ore-controlling structure"; if it mentions "skarnification," it adds the label "copper mineralization-related alteration"; and if it mentions "quartz monzonite," it adds the label "porphyry copper mineralization-related rock mass." This ensures that the annotation results not only include explicit terms but also capture the inherent logical relationships between mineralization elements.

[0052] (3) Pseudo-label screening

[0053] A dual-threshold filter is preset: the pseudo-label confidence threshold is set to 0.85, and the copper ore mineralization knowledge matching threshold is set to 0.8. The confidence score is output by the model, reflecting the model's certainty regarding the label; the knowledge matching score is calculated based on the comprehensive matching degree between the text fragment and the rule base, terminology base, and association graph, reflecting the fit between the label and the copper ore mineralization logic. Pseudo-labels that simultaneously meet both "confidence score ≥ 0.85" and "knowledge matching score ≥ 0.8" are retained as high-quality pseudo-labels and directly included in the initial automatic labeling results; other pseudo-labels (such as those with confidence scores between 0.5 and 0.85, containing ambiguous expressions such as "weak mineralization," or those with knowledge matching scores between 0.5 and 0.8 that contradict the mineralization rules) are marked as low-quality and pushed to step 4 for human-machine collaborative correction. This dual-threshold strategy balances the reliability of the model output (confidence score) and the compliance of domain knowledge (knowledge matching score), achieving a balance between automation and labeling accuracy. The high-quality pseudo-labels after screening are divided into four categories: mineral type and mineralization, alteration, ore-controlling elements, and elemental anomalies. Each category of labels strictly conforms to the mineralization law of copper deposits and can be used directly as preliminary automatic labeling results.

[0054] Step 4: Lightweight Human-Machine Collaboration Correction This step addresses the low-quality pseudo-labels and highly ambiguous text in the automatic annotation results by designing a lightweight human-machine collaboration mechanism that prioritizes high-value samples. By automatically filtering highly ambiguous samples, it guides human intervention to correct only a small number of key samples, avoiding large-scale manual intervention.

[0055] (1) Screening and ranking of highly ambiguous samples Two types of highly ambiguous samples are automatically selected from the initial automatic annotation results: samples with pseudo-label confidence in the middle range (e.g., 0.5-0.85), typically including semantically ambiguous expressions such as "weak mineralization" and "suspected copper mineralization"; and samples with copper mineralization knowledge matching in the middle range (e.g., 0.5-0.8), typically including texts containing some copper mineralization terms but lacking core mineralization conditions, or with ambiguous terminological relationships. A multi-dimensional weighted summation algorithm is used to prioritize highly ambiguous samples. The ranking indicators include: core mineralization term matching (weight 0.4), sample source priority (weight 0.3, e.g., deep exploration reports have the highest priority), copper mineralization knowledge matching (weight 0.2), and ambiguity type priority (weight 0.1, e.g., mineralization type ambiguity has the highest priority). Samples are sorted from high to low based on their comprehensive scores to ensure that samples with the highest mineralization value are processed manually first.

[0056] (2) Manual Interaction Correction A visual annotation interface was developed, comprising a text fragment display area, a current pseudo-label display area, a label editing area, and a metallogenic knowledge reference area. The metallogenic knowledge reference area calls upon the 3D copper deposit metallogenic knowledge system in real time, providing professional support for manual correction. Manual correction is only performed on key samples within a preset percentage range (typically 10%-15%). Correction operations include: confirming correct pseudo-labels, rejecting incorrect pseudo-labels, supplementing missing labels, and modifying inaccurate labels. All correction operations must be annotated with their basis (such as referencing specific metallogenic rule numbers or terminology weights) to ensure traceability.

[0057] (3) Knowledge and model iteration The manually corrected results are categorized and used to update the 3D copper deposit metallogenic knowledge system. If new metallogenic patterns or unreasonable parameters are found, the rules in the rule base are supplemented or corrected; if terms are missing or weights are improperly set, the terminology base is updated; if relationships between terms are missing or weights are unreasonable, the association graph is supplemented or adjusted. The corrected samples are used as new training samples and input into the annotation model for fine-tuning training, allowing the model to learn the logic of the manually corrected results and improve the accuracy of subsequent automatic annotations. Through the above closed-loop iteration, the proportion of highly ambiguous samples decreases significantly after each iteration, and the annotation accuracy continues to improve, achieving the co-evolution of the knowledge system and the annotation model.

[0058] Step 5: Standardize the output of labeled results This step addresses the training needs of AI-based mineral exploration models by standardizing the corrected, high-precision annotation results to construct three types of outputs: basic annotations, quantitative annotations, and correlation annotations. By unifying the format, supplementing quantitative information, and improving correlation relationships, the annotation results are made directly compatible with mainstream AI frameworks without requiring secondary processing.

[0059] (1) Standardization of basic annotation results The corrected high-precision annotation results were organized into a unified basic annotation format. Each record includes at least the text fragment identifier, copper mine-specific label type, label confidence level, original text location, and correction marker. The confidence level of manually corrected samples was uniformly marked as the highest value. The basic annotation results were saved in a structured format compatible with mainstream AI frameworks (such as JSON and XML) for easy loading and training.

[0060] (2) Standardization of quantitative labeling results Based on a copper mineralization rule base, quantitative values ​​for mineralization intensity and mineralization potential are generated for each sample. Mineralization intensity is divided into multiple preset levels and assigned corresponding quantitative values ​​based on parameters such as element content and alteration combination in the text. Mineralization potential is divided into multiple preset levels and assigned corresponding quantitative values ​​based on the degree of matching between the text and the mineralization rules. The quantitative annotation results are stored in association with the basic annotation results, and the basis for the quantification is noted to ensure the traceability of the results.

[0061] (3) Standardization of association annotation results Based on the copper mineralization association map, the terminological relationships and association weights between the labels in each text segment are extracted to generate an association list (format: text segment identifier—term 1—term 2—relationship—association weight). The associations of all segments are then summarized to form a standardized terminological association map, stored in a graph database-compatible format and a general structured format, which can be directly used for training AI graph learning models.

[0062] Finally, the above three types of results are packaged together and accompanied by a result description document (annotation format specifications, quantitative standard definitions, and association rule descriptions) to verify the compatibility of the results with mainstream AI frameworks, ensuring that they can be used for training AI mineral exploration models without secondary processing.

[0063] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A method for semi-automatic annotation of geological text knowledge base based on copper mineralization knowledge guidance, characterized in that, The method includes the following steps: A three-dimensional copper mineralization knowledge system is constructed, which consists of a copper mineralization rule base, a copper-specific terminology base, and a copper mineralization correlation map. The original copper mine geological text was cleaned, professionally segmented, and structured using a copper mine-specific terminology database, resulting in a pre-processed set of structured copper mine geological text fragments. By embedding a three-dimensional copper mineralization knowledge system into an improved semi-supervised model, preliminary automatic annotation results are obtained by automatically annotating a set of structured copper geological text fragments. The preliminary automatic annotation results are screened for highly ambiguous samples. Based on the three-dimensional copper metallogenic knowledge system, the screened samples are corrected and the corrected and high-precision automatic annotation results are output. The high-precision automatic annotation results are standardized into three categories: basic annotation, quantitative annotation, and correlation annotation, so that they can be used as training data for AI mineral exploration models.

2. The semi-automatic annotation method for a geological text knowledge base guided by copper ore mineralization knowledge according to claim 1, characterized in that, The steps for constructing a three-dimensional copper deposit metallogenic knowledge system specifically include: By selecting representative deposit cases from various typical copper mines, and using a structured format that combines conditions to conclusions, a copper mineralization rule library is constructed. By extracting terms from copper mining literature, classifying them into mineral type terms, alteration terms, ore-controlling element terms, and element anomaly terms, and assigning weights to each type of terms related to mineralization, a copper mining-specific terminology library is constructed. Using mineralization events as the core nodes, the terms controlling ore elements are mapped to two types of nodes: ore-forming rock bodies and ore-controlling structures. The terms controlling mineral types are mapped to mineralization type nodes. Alteration type nodes corresponding to alteration terms and element anomaly nodes corresponding to element anomaly terms are connected, and the correlation weights between nodes are labeled to construct a copper mineralization correlation map.

3. The semi-automatic annotation method for a geological text knowledge base guided by copper mineralization knowledge according to claim 2, characterized in that, The steps for constructing the copper ore mineralization rule base specifically include: Obtain representative deposit case data from various typical copper mines; Based on the representative mineral deposit case data, a structured format from condition combination to conclusion is adopted to form mineralization rules. The condition combination includes rock mass type, alteration combination, element content threshold, and tectonic location parameters. The mineralization rules are manually verified from four dimensions: rationality, accuracy, completeness, and adaptability. After verification, the mineralization rules are entered into the copper mineralization rule database, and an index is created for searching by copper ore type and condition combination.

4. The semi-automatic annotation method for a geological text knowledge base guided by copper mineralization knowledge according to claim 1, characterized in that, The steps of cleaning, professionally segmenting, and structuring the original copper mine geological text using a copper mine-specific terminology database specifically include: The original copper mine geological text was converted into plain text format, page marks, table and image marks and editing marks were removed, redundant information was deleted, and a copper mine-specific terminology library was used in conjunction with a spelling correction algorithm to jointly correct optical character recognition errors. A hybrid word segmentation model that embeds features from a copper mine-specific terminology database is used for word segmentation to ensure the integrity of copper mine-specific compound terms. The segmented text is divided into four types of structured fragments: description of ore-controlling structures, description of alteration, description of elemental content, and description of mineralization phenomena. These fragments are then uniformly labeled and classified for storage, resulting in the set of structured copper ore geological text fragments.

5. The semi-automatic annotation method for a geological text knowledge base guided by copper mineralization knowledge according to claim 4, characterized in that, The step of using a copper mine-specific terminology database in conjunction with a spelling correction algorithm to jointly correct optical character recognition errors specifically includes: The terms to be corrected are extracted from the cleaned original copper mine geological text. The terms to be corrected are compared word by word with the copper mine-specific terminology database. If they match perfectly, they are automatically corrected. If the match is not perfect, the text similarity is calculated using the edit distance algorithm. When the similarity reaches a set threshold, the contextual semantics are considered to determine whether to automatically correct the text. The terms that cannot be corrected are sorted according to their term weight priority and marked as items to be reviewed.

6. The semi-automatic annotation method for a geological text knowledge base guided by copper mineralization knowledge according to claim 1, characterized in that, The steps of embedding a three-dimensional copper mineralization knowledge system into an improved semi-supervised model to automatically annotate a set of structured copper ore geological text fragments to obtain preliminary automatic annotation results specifically include: The three-dimensional copper ore-forming knowledge system is embedded into the improved semi-supervised model to obtain the initialized model; Based on a set of structured copper mine geological text fragments, copper mine-specific terms in the text are identified through an initialized model, and corresponding candidate labels are generated based on the identification results. The candidate tags are precisely verified condition by condition based on the copper mineralization rule library, and the implicit mineralization association tags are supplemented based on the copper mineralization association map to generate enhanced pseudo tags. Obtain the confidence score and knowledge matching score generated by the initialized model for each pseudo-label; The system uses a strategy that the confidence level is greater than or equal to a first preset threshold and the knowledge matching level is greater than or equal to a second preset threshold to filter out pseudo-labels and generate preliminary automatic labeling results.

7. The semi-automatic annotation method for a geological text knowledge base guided by copper ore mineralization knowledge according to claim 6, characterized in that, The steps of embedding the three-dimensional copper ore-forming knowledge system into the improved semi-supervised model specifically include: Each structured rule in the copper ore mineralization rule base is converted into a logical expression; During the semi-supervised model training process, the predicted labels of the intermediate output of the model are extracted, and the predicted labels are compared with the logical expressions one by one to calculate the matching degree. Gradient penalty is applied to the model parameters corresponding to the predicted labels with matching degrees lower than a preset threshold. The normalization coefficients are obtained from the copper mine-specific terminology database. The normalization coefficients are then multiplied dimension-by-dimensionally with the basic word vectors in the word embedding layer of the semi-supervised model to obtain weighted word vectors. The normalization coefficients are obtained by normalizing the numerical value of the correlation between each term in the copper mine-specific terminology database and copper mineralization. A composite loss function is constructed, which includes a basic cross-entropy loss and a mineralization rule constraint loss term, wherein the mineralization rule constraint loss term is dynamically calculated based on the matching degree.

8. The semi-automatic annotation method for a geological text knowledge base guided by copper mineralization knowledge according to claim 1, characterized in that, The steps of filtering highly ambiguous samples from the initial automatic annotation results and correcting the filtered samples based on the three-dimensional copper deposit metallogenic knowledge system specifically include: Highly ambiguous samples requiring manual correction are selected from the initial automatic annotation results and sorted according to preset priority rules. The system receives correction instructions for highly ambiguous samples after language sorting through human-computer interaction. The correction instructions include confirmation, rejection, supplementation, or modification of the automatic annotation results. The corrected samples are used to update the three-dimensional copper mineralization knowledge system and to train the improved semi-supervised model.

Images