Geological literature key information extraction method based on large model

Abstract

The application discloses a geological literature key information extraction method based on a large model, which comprises collecting geological literature data in a preset area and preprocessing the geological literature data; obtaining literature information data by jointly extracting information from the geological multi-source data through a geological knowledge graph, obtaining a semantic alignment relationship by using a multi-granularity semantic alignment mechanism to correspond to the upper and lower relationships of the literature information data; setting an initial weight of the literature information data according to a geological rule base and the semantic alignment relationship, dynamically adjusting the initial weight based on a reinforcement learning decision routing to obtain an optimized weight, and obtaining fusion data by fusing heterogeneous knowledge graphs according to the literature information data and the optimized weight; constructing a geological literature key information extraction model according to the fusion data, and performing man-machine collaborative optimization on the geological literature key information extraction model according to the result deviation; and inputting to-be-extracted data into the geological literature key information extraction model to output an extraction result.

Description

Technical Field

[0001] This invention relates to the field of geological information extraction, and more particularly to a method for extracting key information from geological documents based on large models. Background Technology

[0002] In the field of geological research, geological literature contains various types of data, including text, maps, and tables, carrying core information such as regional geological characteristics, entity attributes, and relational relationships. It serves as crucial data support for geological exploration and resource assessment. However, the sheer volume and diverse formats of current geological literature present significant limitations to traditional information extraction methods: insufficient adaptability to multimodal data, making it difficult to simultaneously process technical terms in text, spatial coordinates in maps, and structured data in tables; a lack of domain-specific semantic understanding, easily confusing obscure geological terms with entities of the same name, and failing to effectively establish hierarchical relationships between entities of different granularities; and rigid data weighting, making dynamic adjustments based on data source credibility and information importance difficult, resulting in insufficient accuracy and completeness of extraction results. Furthermore, manual extraction of key information is time-consuming, labor-intensive, and inefficient, failing to meet the needs of geological research for rapid data acquisition and integration.

[0003] Therefore, there is an urgent need for a method that integrates domain knowledge and advanced technologies to solve problems such as collaborative extraction, semantic alignment, and dynamic weight optimization of multimodal geological data, so as to achieve efficient and accurate extraction of key information from geological literature and provide strong data support for geological research. Summary of the Invention

[0004] The purpose of this invention is to provide a method for extracting key information from geological literature based on large models.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0006] This invention includes the following steps:

[0007] Collect geological literature data from a preset area and preprocess the geological literature data; the geological literature data includes text data, map data, and tabular data.

[0008] Literature information data is obtained by jointly extracting information from the geological multi-source data using a geological knowledge graph, and semantic alignment relationship is obtained by performing corresponding hierarchical relationship on the literature information data using a multi-granularity semantic alignment mechanism.

[0009] The initial weights of the literature information data are set according to the geological rule base and the semantic alignment relationship. The initial weights are dynamically adjusted based on the decision routing of reinforcement learning to obtain optimized weights. The heterogeneous knowledge graph is fused according to the literature information data and the optimized weights to obtain fused data.

[0010] A key information extraction model for geological documents is constructed based on the fused data. The key information extraction model for geological documents is then optimized by human-computer collaboration based on the result deviation. The data to be extracted is input into the key information extraction model for geological documents, and the extraction results are output.

[0011] Furthermore, a method for obtaining literature information data by jointly extracting information from the aforementioned multi-source geological data using a geological knowledge graph includes:

[0012] For the text data: pre-trained word vectors based on geological word vectors, and the word vector space that is easily confused is corrected by a geological terminology dictionary; character-level BPE segmentation is used for obscure terms, and after the input text is encoded by BERT, the contextual dependencies are captured by a bidirectional long short-term memory network, and the entity label sequence is obtained by combining the conditional random layer with the geological entity transition probability matrix; the entity disambiguation algorithm is used to match geological knowledge graph nodes, and for entities with the same name, subject labels are assigned based on contextual similarity to obtain the text entity set;

[0013] For map data: DBNet algorithm is used to detect text regions in the map, CRNN model is used to identify text content, spatial coordinates of text detection boxes are extracted, and combined with map scale to convert to actual geographic coordinates or relative positions, establish spatial association between text and entities, and output map entity set;

[0014] For tabular data: Based on text density and table line features, the table is divided into high-density data area and low-density annotation area. For borehole tables, the three-dimensional data of depth, lithology and fossils are automatically extracted and arranged. The numerical data is standardized by unit, the qualitative description is mapped to quantitative attribute values, matched with the knowledge graph attribute dictionary, and the output is a table entity set.

[0015] The method combines rule base and statistical learning to extract entity pair relationships, calls the geological knowledge graph reasoning engine to verify the relationships, automatically marks contradictory relationships as pending confirmation, and outputs non-contradictory relationships as a set of verified relationships; rule-driven: when an entity is followed by ± value, an entity plus value attribute relationship is automatically established; statistical-driven: relationships are identified by combining co-occurrence frequency with BERT relationship classifier.

[0016] Output the text entity set, figure entity set, table entity set, and verification relation set as document information data.

[0017] Furthermore, a method for obtaining semantic alignment relationships by performing corresponding hierarchical relationships on the document information data using a multi-granularity semantic alignment mechanism includes:

[0018] The entities in the literature information data are labeled with granularity, and the rationality of the granularity is verified by the geological ontology database. Entities with mixed granularity are split or merged, and the entities in the literature information data are layered according to a 5-level granularity system, where the 5 levels are domain level, discipline level, type level, instance level, and attribute level.

[0019] Semantic association of entities at different granularities is performed based on dual verification of word vector similarity and domain rules;

[0020] Word vector similarity verification: The cosine similarity between entities is calculated using optimized geological word vectors. The similarity thresholds between domain-level and subject-level entities, and between subject-level and type-level entities are greater than or equal to 0.85. The similarity thresholds between type-level and instance-level entities, and between instance-level and attribute-level entities are greater than or equal to 0.78. Entities with hierarchical relationships but below the similarity threshold are subject to rule verification.

[0021] Domain rule validation: If the definition of entity A contains the core features of entity B, then establish a hierarchical relationship between A and B; if entity B inherits all the core attributes of entity A, then confirm the hierarchical relationship; entities with semantic conflicts are marked as parallel relationships.

[0022] By constructing complete hierarchical relationship chains for entities of different granularities through hierarchical transmission and conflict resolution, the structured semantically aligned relationship table is output. Hierarchical transmission automatically completes the relationship chain based on transitivity rules. Conflict resolution: when an entity has multiple potential superior concepts, subject tags are selected first, and intermediate nodes are automatically inserted for relationships that skip intermediate levels.

[0023] Furthermore, the method for setting the initial weights of the document information data based on the geological rule base and the semantic alignment relationship includes:

[0024] The rule base includes entity type rules, data source rules, and attribute importance rules, and the weight benchmark values ​​are determined through expert consensus.

[0025] The baseline weights are corrected by multi-granular semantic alignment. For hierarchical relationships, the subordinate entity inherits the weight of the superior entity and increases the weight. For synonym relationships, the weight of authoritative domain terms is given priority.

[0026] Based on the rule base, entity type matching and data source matching are performed on entities in the literature information data to obtain entity type value and data source weight. Semantic alignment correction value is obtained by correcting the semantic alignment baseline value. The entity type value and data source weight are weighted and summed, and the semantic alignment correction value is added to obtain the entity weight.

[0027] Attribute type matching is performed on the attributes associated with entities, and attribute weights are obtained by correcting the reliability of attribute values ​​based on quantitative attributes, qualitative data, precise numerical values, and fuzzy descriptions.

[0028] The system performs relationship type matching between entities, and obtains relationship weights by adjusting the evidence strength based on multi-source evidence, single-source evidence, knowledge graph verification passed, and relationships to be verified; the entity weights, attribute weights, and relationship weights are output as the initial weights.

[0029] Furthermore, the method for dynamically adjusting the initial weights to obtain optimized weights based on reinforcement learning-based decision routing includes:

[0030] The weight adjustment task is modeled as a Markov decision process, and dynamic decision-making is carried out through state, action, and reward loops. The geological entity type and literature topic to be extracted are the task context, the identified entity set and attributes are the extracted information, and the credibility score of each data source is the multi-source data reliability label.

[0031] The adjustment of the weights of multi-source data includes three basic actions and magnitude control. The three basic actions are to enhance text weight, increase map weight, and integrate multi-source weights. The weight adjustment range for enhancing text weight is +0.1 to +0.3, the weight adjustment range for increasing map weight is +0.1 to +0.3, and the weight adjustment range for integrating multi-source weights is ±0.05 for each source weight.

[0032] The reward value is calculated based on the consistency between the extraction results and the geological knowledge map, data reliability, and task objectives. The expression is:

[0033] ;

[0034] in The degree of matching between entity attributes and geological knowledge graphs. To ensure the weight allocation aligns with the credibility of the data source, To ensure the completeness of entity attribute extraction under the current weight, For matching degree weight, As a weight for fit, Integrity weight;

[0035] An evaluation network and a target network strategy are introduced to optimize the deep Q-network in reinforcement learning. The deep Q-network structure consists of three fully connected layers and an activation function. The evaluation network outputs the Q-value of each action in real time, with the current state S as input and the Q-value of each action in the action space as output. The target network freezes and copies parameters from the evaluation network periodically to calculate the target Q-value. The three fully connected layers are the input layer, hidden layer, and output layer.

[0036] Using initial weights based on a geological rule base as initial actions, the network is evaluated by inputting the current state S and outputting the Q-value of each action. The greedy strategy selects actions, adjusts weights, extracts entity attributes, calculates reward values, observes new states, and updates the Q-value. The expression is:

[0037] ;

[0038] in This is the current state. For the new state, As a discount factor, For the current action, For new actions, The evaluation network value under the new state and new action. The target network value under the current state and current action. To maximize the value of all possible new actions in the action space;

[0039] Every 100 training steps, the parameters of the evaluation network are copied to the target network. The temporal difference error is calculated based on the absolute value of the difference between the evaluation network and the target network. Sample priorities are assigned according to the importance of the samples, and the sample sampling probability is calculated. Importance sampling weights are introduced to correct the loss function. This process is iterated until the policy network converges, and the optimized weight set for the current task is output as the optimized weights.

[0040] Furthermore, the method for obtaining fused data by fusing heterogeneous knowledge graphs based on the literature information data and the optimized weights includes:

[0041] Based on the entities and relationships within the document information data, text subgraphs, image subgraphs, structured subgraphs, and external data subgraphs are obtained. A multimodal subgraph is constructed based on these subgraphs, and fusion priorities are assigned to each subgraph based on optimized weights. The text subgraph contains document text entities and attribute relationships, with weights initialized according to entity extraction confidence. The image subgraph contains spatial entities and coordinate attributes, with weights dynamically adjusted based on the recognition accuracy of the deep bounding box network and convolutional recurrent neural network. Structured subgraphs and external data subgraphs are also included.

[0042] The BERT-base-cased model is used to encode the entity attribute text. A 128-dimensional feature vector is obtained through max pooling. The embedding space is optimized by fusing it with geological word vectors. The spatial topological relationship of map entities is learned by using hyperbolic graph convolutional network. The entity features are trained by adaptive gradient algorithm, and the depth, lithology and fossil triples are transformed into structured embedding vectors.

[0043] By comparing the similarity of the embedding vectors of entities in different subgraphs, the geological knowledge graph reasoning engine is invoked to verify the rationality of the alignment, and subject labels are assigned to entities with the same name by combining optimized weights and contextual keywords.

[0044] The confidence level of multi-source evidence for the same relationship is calculated by weighting it with optimized weights. A relationship confidence threshold of 0.6 is set, and relationships with confidence levels below the threshold are marked as requiring manual verification.

[0045] Using verified relationships in the knowledge graph to infer missing relationships, and combining geophysical data subgraphs to complete relationships between entities across subgraphs;

[0046] Output the entity ID, name, type, multi-source attributes, weight value, relationship ID, head / tail entity ID, relationship type, confidence level, and weight source as fused data.

[0047] Furthermore, the method for constructing a key information extraction model for geological documents based on the fused data includes:

[0048] The TransE algorithm is used to map knowledge graph entities to vector space, so that the vector distance between semantically similar entities is less than the similarity threshold, and the distance between dissimilar entities is greater than the dissimilarity threshold.

[0049] For multi-source entities, feature vectors are fused by weighted averaging, with weights dynamically allocated proportionally to vectorized entity features, and geological relationship path features are introduced to enhance relationship features.

[0050] A framework for a key information extraction model from geological documents is constructed based on a multimodal feature encoder, a dynamic weight fusion layer, and a relational reasoning and information extraction layer.

[0051] Multimodal feature encoder: A geological-domain fine-tuned BERT input text fragment is used to output a contextual feature vector, obtaining text feature encoding; based on the spatial coordinates of map entities and text content, visual features are extracted through a convolutional neural network, concatenated with text features, and then weighted through an attention mechanism to obtain map feature encoding; for structured tabular data, Transformer-XL is used to capture long sequence dependencies, outputting a temporal feature vector to obtain table feature encoding; the weights of text feature encoding, map feature encoding, and table feature encoding are dynamically allocated through a multimodal attention layer;

[0052] Dynamic Weight Fusion Layer: Based on reinforcement learning decision routing, dynamic weight adjustment of multi-source features is performed. The current task to be extracted, the fused feature vector, and the reliability label of multi-source data are defined as the state. The action space is defined by adjusting the weights of each modality feature and the given weight value range. The weights are adjusted by a given reward function, which includes positive reward and negative penalty. The positive reward is the matching of the extraction result with the geological knowledge graph, and the negative penalty is the entity type error.

[0053] Relational Reasoning and Information Extraction Layer: BiLSTM-CRF is used to identify entity boundaries, and a geological entity dictionary is combined for entity boundary detection; attribute values ​​are bound to entities through an entity-attribute attention matrix to obtain attribute value associations; entity pair relation probabilities are calculated based on a knowledge graph embedding model, and relation reasoning is performed based on entity pair relation probabilities; the entity set, attribute set, and relation are output as extraction results;

[0054] Training data is constructed, and multi-task joint training is performed using a main task and an auxiliary task. The main task includes entity recognition and relation extraction, while the auxiliary task is multi-source data weight prediction. Geological constraints are added during the training process to correct prediction results that violate the rules through penalty terms.

[0055] Furthermore, a method for human-machine collaborative optimization of the geological literature key information extraction model based on result deviations includes:

[0056] Set confidence thresholds for entity recognition and relation extraction, and mark low-confidence samples as samples to be verified; call the geological knowledge graph inference engine to identify conflicting data that contradicts the domain rules, and the results include entity type conflicts, relational logic contradictions, and attribute value anomalies;

[0057] Compare the extraction results of text, images, and tables: if the same entity has conflicting attributes in different data sources, mark it as a cross-modal conflict sample; classify and statistically analyze high-frequency error types by entity, relationship, and attribute, and push them to obtain an error sample pool;

[0058] The system performs entity-level, relation-level, and attribute-level corrections on the erroneous sample pool. When experts correct entities, it automatically recommends standard attributes of similar entities in the knowledge graph and displays historical amendment examples of this type of error to assist experts in making quick decisions and obtaining corrected samples. Entity-level correction supports entity type adjustment and boundary calibration; relation-level correction provides a drop-down selection of relation types and supports relation direction adjustment; attribute-level correction supports unit calibration and error range supplementation for numerical attributes, and allows free editing of text attributes.

[0059] The format of the corrected samples is standardized, and model parameters are extracted based on the corrected samples and key information from geological literature for domain adaptation fine-tuning and small sample learning optimization.

[0060] The beneficial effects of this invention are:

[0061] This invention is a method for extracting key information from geological literature based on large-scale models. Compared with existing technologies, this invention has the following technical advantages:

[0062] This invention addresses the challenge of heterogeneous data extraction by employing preprocessing, joint information extraction, corresponding hierarchical relationships, setting initial weights, dynamically adjusting initial weights, heterogeneous knowledge graph fusion, model construction, and model optimization steps. It adapts to multimodal data, simultaneously processing text, image, and tabular data. Furthermore, it improves the accuracy of entity recognition and relationship extraction through multi-granular semantic alignment and dynamic weight optimization. A human-machine collaborative optimization mechanism continuously corrects model biases, enhancing robustness. Automated extraction replaces manual work, significantly improving efficiency and adapting to various geological scenarios, demonstrating strong universality. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the steps of the method for extracting key information from geological documents based on a large model, as described in this invention. Detailed Implementation

[0064] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0065] The method for extracting key information from geological literature based on a large model, as described in this invention, includes the following steps:

[0066] like Figure 1 As shown, this embodiment includes the following steps:

[0067] Collect geological literature data from a preset area and preprocess the geological literature data; the geological literature data includes text data, map data, and tabular data.

[0068] In the actual assessment, the geological literature on Mesozoic magmatic rocks in a certain region was used as the extraction object, covering 300 text documents, 80 geological maps and 50 borehole table data;

[0069] The data collected included geological survey reports, journal articles, and other literature data for the region. The text data included descriptions of igneous rock types and isotopic ages, the map data included geological profiles and lithological distribution maps, and the tabular data mainly consisted of borehole core analysis tables. During the preprocessing stage, 23 duplicate documents were removed, 17 blurry areas in the maps were corrected, and the units of numerical values ​​in the tables were standardized.

[0070] Literature information data is obtained by jointly extracting information from the geological multi-source data using a geological knowledge graph, and semantic alignment relationship is obtained by performing corresponding hierarchical relationship on the literature information data using a multi-granularity semantic alignment mechanism.

[0071] The initial weights of the literature information data are set according to the geological rule base and the semantic alignment relationship. The initial weights are dynamically adjusted based on the decision routing of reinforcement learning to obtain optimized weights. The heterogeneous knowledge graph is fused according to the literature information data and the optimized weights to obtain fused data.

[0072] A key information extraction model for geological documents is constructed based on the fused data. The key information extraction model for geological documents is then optimized by human-computer collaboration based on the result deviation. The data to be extracted is input into the key information extraction model for geological documents, and the extraction results are output.

[0073] In this embodiment, the method for obtaining literature information data by jointly extracting information from the multi-source geological data using a geological knowledge graph includes:

[0074] For the text data: pre-trained word vectors based on geological word vectors, and the word vector space that is easily confused is corrected by a geological terminology dictionary; character-level BPE segmentation is used for obscure terms, and after the input text is encoded by BERT, the contextual dependencies are captured by a bidirectional long short-term memory network, and the entity label sequence is obtained by combining the conditional random layer with the geological entity transition probability matrix; the entity disambiguation algorithm is used to match geological knowledge graph nodes, and for entities with the same name, subject labels are assigned based on contextual similarity to obtain the text entity set;

[0075] For map data: DBNet algorithm is used to detect text regions in the map, CRNN model is used to identify text content, spatial coordinates of text detection boxes are extracted, and combined with map scale to convert to actual geographic coordinates or relative positions, establish spatial association between text and entities, and output map entity set;

[0076] For tabular data: Based on text density and table line features, the table is divided into high-density data area and low-density annotation area. For borehole tables, the three-dimensional data of depth, lithology and fossils are automatically extracted and arranged. The numerical data is standardized by unit, the qualitative description is mapped to quantitative attribute values, matched with the knowledge graph attribute dictionary, and the output is a table entity set.

[0077] The method combines rule base and statistical learning to extract entity pair relationships, calls the geological knowledge graph reasoning engine to verify the relationships, automatically marks contradictory relationships as pending confirmation, and outputs non-contradictory relationships as a set of verified relationships; rule-driven: when an entity is followed by ± value, an entity plus value attribute relationship is automatically established; statistical-driven: relationships are identified by combining co-occurrence frequency with BERT relationship classifier.

[0078] Output the text entity set, figure entity set, table entity set, and verification relation set as document information data;

[0079] In the actual assessment, the text data processing involved: pre-training word vectors based on geological word vectors, correcting the easily confused word vector space of granite and granodiorite using the "Geological Terminology" dictionary; character-level BPE segmentation was used for rare zircon U-Pb dating terms, and after BERT encoding, entity label sequences were obtained by combining BiLSTM with the geological entity transition probability matrix; an entity disambiguation algorithm was used to match the geological knowledge graph, and for quartz entities with the same name, subject labels of igneous rock minerals and sedimentary rock minerals were assigned according to the context, ultimately obtaining more than 1,200 text entities;

[0080] Map data processing: The DBNet algorithm is used to detect text regions in the map, and the CRNN model is used to identify the text content of granite bodies and fault zones. The spatial coordinates of the text detection boxes are extracted, and combined with the 1:50000 scale of the map, they are converted into actual geographic coordinates, spatial associations are established, and more than 850 map entities are output.

[0081] Table data processing: Based on text density and table line features, the table region is split, and the three-dimensional data of depth, lithology and fossils are automatically extracted from the borehole table. The qualitative description of dense and hard is mapped to the quantitative attribute value of 0.85, matched with the knowledge graph attribute dictionary, and more than 600 table entities are output.

[0082] The system identifies the relationship between the ±20 Ma entity and numerical attributes of granite through rule-driven identification, and identifies intrusive contact and phase transition relationships through statistical-driven identification combined with the BERT relationship classifier. The system calls the geological knowledge graph inference engine for verification, marks 32 contradictory relationships as pending confirmation, and outputs a set of more than 1,500 verification relationships.

[0083] In this embodiment, the method for obtaining semantic alignment relationships by performing corresponding hierarchical relationships on the document information data using a multi-granularity semantic alignment mechanism includes:

[0084] The entities in the literature information data are labeled with granularity, and the rationality of the granularity is verified by the geological ontology database. Entities with mixed granularity are split or merged, and the entities in the literature information data are layered according to a 5-level granularity system, where the 5 levels are domain level, discipline level, type level, instance level, and attribute level.

[0085] Semantic association of entities at different granularities is performed based on dual verification of word vector similarity and domain rules;

[0086] Word vector similarity verification: The cosine similarity between entities is calculated using optimized geological word vectors. The similarity thresholds between domain-level and subject-level entities, and between subject-level and type-level entities are greater than or equal to 0.85. The similarity thresholds between type-level and instance-level entities, and between instance-level and attribute-level entities are greater than or equal to 0.78. Entities with hierarchical relationships but below the similarity threshold are subject to rule verification.

[0087] Domain rule validation: If the definition of entity A contains the core features of entity B, then establish a hierarchical relationship between A and B; if entity B inherits all the core attributes of entity A, then confirm the hierarchical relationship; entities with semantic conflicts are marked as parallel relationships.

[0088] Complete hierarchical relationship chains of entities at different granularities are constructed through hierarchical transmission and conflict resolution, and the output is a structured semantic alignment relationship table; hierarchical transmission automatically completes the relationship chain based on transitivity rules; conflict resolution: when an entity has multiple potential superior concepts, subject tags are selected first, and intermediate nodes are automatically inserted for relationships that skip intermediate levels.

[0089] In the actual evaluation, each level corresponds to the hierarchical structure of the geological knowledge graph, and entities are layered according to a five-level granularity system: domain level, discipline level, type level, instance level, and attribute level. For example, the geological domain (domain level) - igneous petrology (discipline level) - intrusive rocks (type level) - granite bodies (instance level) - age (attribute level). Cosine similarity is calculated using optimized geological word vectors. The similarity between discipline level and type level entities is ≥0.85, and the similarity between type level and instance level entities is ≥0.78. For entities of basalt and basic volcanic rocks that are below the threshold but have a hierarchical relationship, alignment relationships are established through domain rule verification, and finally a structured semantic alignment relationship table is formed.

[0090] In this embodiment, the method for setting the initial weight of the document information data based on the geological rule base and the semantic alignment relationship includes:

[0091] The rule base includes entity type rules, data source rules, and attribute importance rules, and the weight benchmark values ​​are determined through expert consensus.

[0092] The baseline weights are corrected by multi-granular semantic alignment. For hierarchical relationships, the subordinate entity inherits the weight of the superior entity and increases the weight. For synonym relationships, the weight of authoritative domain terms is given priority.

[0093] Based on the rule base, entity type matching and data source matching are performed on entities in the literature information data to obtain entity type value and data source weight. Semantic alignment correction value is obtained by correcting the semantic alignment baseline value. The entity type value and data source weight are weighted and summed, and the semantic alignment correction value is added to obtain the entity weight.

[0094] Attribute type matching is performed on the attributes associated with entities, and attribute weights are obtained by correcting the reliability of attribute values ​​based on quantitative attributes, qualitative data, precise numerical values, and fuzzy descriptions.

[0095] The relationship between entities is matched by relationship type, and the relationship weight is obtained by correcting the evidence strength based on multi-source evidence, single-source evidence, knowledge graph verification passed, and relationship to be verified.

[0096] Output the entity weight, attribute weight, and relationship weight as the initial weights;

[0097] In actual assessments, the following rules apply: Entity type: igneous rocks > sedimentary rocks > metamorphic rocks; tectonic entities > mineral entities, with a weight range of 0.6-0.9; Data source: laboratory analysis data > field measurement data > literature citation data, with a weight range of 0.5-0.95; Attribute importance: spatiotemporal attributes > physical attributes > descriptive attributes, with a weight range of 0.4-0.85.

[0098] Entity types include igneous rocks, sedimentary rocks, and metamorphic rocks; attribute types include spatiotemporal attributes, physical attributes, and descriptive attributes, with physical attributes including tectonic attributes and structural attributes; relationship types include intrusive contact and phase transition.

[0099] The baseline weight for igneous rocks is 0.85, the baseline weight for isotopic age is 0.85, the baseline weight for tectonic attributes is 0.7, the baseline weight for structural attributes is 0.5, the baseline weight for intrusive contact relationships in the rule base is 0.8, and the baseline weight for phase transition relationships is 0.75.

[0100] Based on the rule base, the baseline weight of igneous rock in the entity type rule is 0.85, the weight of laboratory analysis data in the data source rule is 0.95, and the weight of isotopic age in the attribute importance rule is 0.85. Combining the semantic alignment relationship, the lower entity granite inherits the weight of the upper entity intrusive rock and increases it by 0.05 to obtain the initial weight set.

[0101] In this embodiment, the method for dynamically adjusting the initial weights to obtain optimized weights based on reinforcement learning-based decision routing includes:

[0102] The weight adjustment task is modeled as a Markov decision process, and dynamic decision-making is carried out through state, action, and reward loops. The geological entity type and literature topic to be extracted are the task context, the identified entity set and attributes are the extracted information, and the credibility score of each data source is the multi-source data reliability label.

[0103] The adjustment of the weights of multi-source data includes three basic actions and magnitude control. The three basic actions are to enhance text weight, increase map weight, and integrate multi-source weights. The weight adjustment range for enhancing text weight is +0.1 to +0.3, the weight adjustment range for increasing map weight is +0.1 to +0.3, and the weight adjustment range for integrating multi-source weights is ±0.05 for each source weight.

[0104] The reward value is calculated based on the consistency between the extraction results and the geological knowledge map, data reliability, and task objectives. The expression is:

[0105] ;

[0106] in The degree of matching between entity attributes and geological knowledge graphs. To ensure the weight allocation aligns with the credibility of the data source, To ensure the completeness of entity attribute extraction under the current weight, For matching degree weight, As a weight for fit, Integrity weight;

[0107] An evaluation network and a target network strategy are introduced to optimize the deep Q-network in reinforcement learning. The deep Q-network structure consists of three fully connected layers and an activation function. The evaluation network outputs the Q-value of each action in real time, with the current state S as input and the Q-value of each action in the action space as output. The target network freezes and copies parameters from the evaluation network periodically to calculate the target Q-value. The three fully connected layers are the input layer, hidden layer, and output layer.

[0108] Using initial weights based on a geological rule base as initial actions, the network is evaluated by inputting the current state S and outputting the Q-value of each action. The greedy strategy selects actions, adjusts weights, extracts entity attributes, calculates reward values, observes new states, and updates the Q-value. The expression is:

[0109] ;

[0110] in This is the current state. For the new state, As a discount factor, For the current action, For new actions, The evaluation network value under the new state and new action. The target network value under the current state and current action. To maximize the value of all possible new actions in the action space;

[0111] Every 100 training steps, the parameters of the evaluation network are copied to the target network. The temporal difference error is calculated based on the absolute value of the difference between the evaluation network and the target network. Sample priorities are assigned according to the importance of the samples, and the sample sampling probability is calculated. Importance sampling weights are introduced to correct the loss function. The process is iterated until the policy network converges. The optimized weight set for the current task is then output as the optimized weights.

[0112] In actual assessment, Greedy strategy selects actions: There is a 10% chance of random exploration and a 90% chance of choosing the action with the highest Q value.

[0113] The sample sampling probability is calculated by dividing the sample priority by the sum of sample priorities. Weight adjustment is modeled as a Markov decision process, using the extraction of magmatic rock age attributes as the task context. - Greedy strategy ( The selected action increases the weight of the text data by 0.2; the R value is calculated based on the reward function, where the matching degree weight is 0.4, the fit degree weight is 0.3, and the integrity weight is 0.3. After 1000 steps of training, the policy network converges and outputs the optimized weight set.

[0114] In this embodiment, the method for obtaining fused data by fusing heterogeneous knowledge graphs based on the literature information data and the optimized weights includes:

[0115] Based on the entities and relationships within the document information data, text subgraphs, image subgraphs, structured subgraphs, and external data subgraphs are obtained. A multimodal subgraph is constructed based on these subgraphs, and fusion priorities are assigned to each subgraph based on optimized weights. The text subgraph contains document text entities and attribute relationships, with weights initialized according to entity extraction confidence. The image subgraph contains spatial entities and coordinate attributes, with weights dynamically adjusted based on the recognition accuracy of the deep bounding box network and convolutional recurrent neural network. Structured subgraphs and external data subgraphs are also included.

[0116] The BERT-base-cased model is used to encode the entity attribute text. A 128-dimensional feature vector is obtained through max pooling. The embedding space is optimized by fusing it with geological word vectors. The spatial topological relationship of map entities is learned by using hyperbolic graph convolutional network. The entity features are trained by adaptive gradient algorithm, and the depth, lithology and fossil triples are transformed into structured embedding vectors.

[0117] By comparing the similarity of the embedding vectors of entities in different subgraphs, the geological knowledge graph reasoning engine is invoked to verify the rationality of the alignment, and subject labels are assigned to entities with the same name by combining optimized weights and contextual keywords.

[0118] The confidence level of multi-source evidence for the same relationship is calculated by weighting it with optimized weights. A relationship confidence threshold of 0.6 is set, and relationships with confidence levels below the threshold are marked as requiring manual verification.

[0119] Using verified relationships in the knowledge graph to infer missing relationships, and combining geophysical data subgraphs to complete relationships between entities across subgraphs;

[0120] Output the entity ID, name, type, multi-source attributes, weight value, relationship ID, head / tail entity ID, relationship type, confidence level, and weight source as fused data;

[0121] In the actual evaluation, the optimization weights are ranked as follows: text subgraph weights > structured subgraphs > map subgraphs > external data subgraphs; the spatial topological relationships of map entities are learned using hyperbolic graph convolutional networks, with an embedding dimension of 128.

[0122] A multimodal subgraph is constructed from text, maps, structured data, and external geophysical data. Text subgraphs are assigned the highest priority based on optimized weights. An BERT-base-cased model is used to encode entity attribute text, obtaining 128-dimensional feature vectors. A hyperbolic graph convolutional network is used to learn the spatial topological relationships of map entities. The similarity of subgraph entity embedding vectors is compared, and subject labels are assigned to entities with the same fault name. Relationship confidence is calculated using optimized weights, and 28 relationships with confidence scores below 0.6 are marked for manual verification. 35 missing relationships across subgraphs are completed, and the fused data is output.

[0123] In this embodiment, the method for constructing a key information extraction model from geological literature based on the fused data includes:

[0124] The TransE algorithm is used to map knowledge graph entities to a 128-dimensional vector space, so that the vector distance between semantically similar entities is less than the similarity threshold, and the distance between dissimilar entities is greater than the dissimilarity threshold.

[0125] For multi-source entities, feature vectors are fused by weighted averaging, with weights dynamically allocated proportionally to vectorized entity features, and geological relationship path features are introduced to enhance relationship features.

[0126] A framework for a key information extraction model from geological documents is constructed based on a multimodal feature encoder, a dynamic weight fusion layer, and a relational reasoning and information extraction layer.

[0127] Multimodal feature encoder: A geological-domain fine-tuned BERT input text fragment is used to output a contextual feature vector, obtaining text feature encoding; based on the spatial coordinates of map entities and text content, visual features are extracted through a convolutional neural network, concatenated with text features, and then weighted through an attention mechanism to obtain map feature encoding; for structured tabular data, Transformer-XL is used to capture long sequence dependencies, outputting a temporal feature vector to obtain table feature encoding; the weights of text feature encoding, map feature encoding, and table feature encoding are dynamically allocated through a multimodal attention layer;

[0128] Dynamic Weight Fusion Layer: Based on reinforcement learning decision routing, dynamic weight adjustment of multi-source features is performed. The current task to be extracted, the fused feature vector, and the reliability label of multi-source data are defined as the state. The action space is defined by adjusting the weights of each modality feature and the given weight value range. The weights are adjusted by a given reward function, which includes positive reward and negative penalty. The positive reward is the matching of the extraction result with the geological knowledge graph, and the negative penalty is the entity type error.

[0129] Relational Reasoning and Information Extraction Layer: BiLSTM-CRF is used to identify entity boundaries, and a geological entity dictionary is combined for entity boundary detection; attribute values ​​are bound to entities through an entity-attribute attention matrix to obtain attribute value associations; entity pair relation probabilities are calculated based on a knowledge graph embedding model, and relation reasoning is performed based on entity pair relation probabilities; the entity set, attribute set, and relation are output as extraction results;

[0130] Training data is constructed, and multi-task joint training is performed using a main task and an auxiliary task. The main task includes entity recognition and relation extraction, while the auxiliary task is multi-source data weight prediction. Geological constraints are added during the training process to correct prediction results that violate the rules through penalty terms.

[0131] In actual evaluation, entity attributes are standardized, relationships with confidence scores less than 0.5 are filtered out, and core relationship pairs are retained; the similarity threshold is 0.3, the heterogeneity threshold is 0.7, and the vector space is 128-dimensional.

[0132] Training data: 5,000 geological documents were selected, and 120,000+ entities and 80,000+ relationships were manually annotated to form a geological annotation corpus; pseudo-label data was automatically generated from the "Geological Dictionary" and "Regional Geological Records" to expand the training sample to 300,000+.

[0133] The TransE algorithm is used to map knowledge graph entities to a 128-dimensional vector space, constructing a framework consisting of a multimodal feature encoder, a dynamic weight fusion layer, and a relation reasoning and information extraction layer. The multimodal feature encoder uses a BERT fine-tuned approach from the geological domain to extract text features, a convolutional neural network to extract map features, and Transformer-XL to extract table features. The dynamic weight fusion layer adjusts modal weights based on reinforcement learning. The relation reasoning layer uses BiLSTM-CRF to identify entity boundaries and combines this with a knowledge graph embedding model to reason about relationships.

[0134] In this embodiment, the method for human-machine collaborative optimization of the geological literature key information extraction model based on the result deviation includes:

[0135] Set confidence thresholds for entity recognition and relation extraction, and mark low-confidence samples as samples to be verified; call the geological knowledge graph inference engine to identify conflicting data that contradicts the domain rules, and the results include entity type conflicts, relational logic contradictions, and attribute value anomalies;

[0136] Compare the extraction results of text, images, and tables: if the same entity has conflicting attributes in different data sources, mark it as a cross-modal conflict sample; classify and statistically analyze high-frequency error types by entity, relationship, and attribute, and push them to obtain an error sample pool;

[0137] The system performs entity-level, relation-level, and attribute-level corrections on the erroneous sample pool. When experts correct entities, it automatically recommends standard attributes of similar entities in the knowledge graph and displays historical amendment examples of this type of error to assist experts in making quick decisions and obtaining corrected samples. Entity-level correction supports entity type adjustment and boundary calibration; relation-level correction provides a drop-down selection of relation types and supports relation direction adjustment; attribute-level correction supports unit calibration and error range supplementation for numerical attributes, and allows free editing of text attributes.

[0138] The format of the corrected samples is standardized, and model parameters are extracted based on the corrected samples and key information from geological literature for domain adaptation fine-tuning and small sample learning optimization.

[0139] In actual evaluation, the confidence threshold for entity recognition was 0.75, and the confidence threshold for relation extraction was 0.65.

[0140] Domain adaptation fine-tuning: The hierarchical parameter update freezes the first 8 layers of the BERT pre-trained layer, and only fine-tunes the last 4 layers and the BiLSTM-CRF layer; a domain penalty term is added to the loss function to add extra loss for predictions that violate geological rules;

[0141] Minimal learning optimization: SMOTE oversampling is used for low-frequency error types, expanding the sample size to 500+; the model converges after 30 iterations using a cosine annealing learning rate.

[0142] 120 low-confidence samples were labeled, identifying 25 sets of entity type conflicts and 18 sets of contradictory relational logic. Experts corrected these conflicts by adjusting entity types, resolving relational directions, and calibrating numerical units. Based on the corrected samples, the model underwent domain-adaptation fine-tuning and few-sample learning optimization, converging after 30 iterations.

[0143] The 100 new geological documents to be extracted were input into the optimized model, and more than 2,800 entities, more than 1,500 attributes, and more than 1,200 sets of relationships were successfully extracted. The entity recognition accuracy rate reached 92.5%, and the relationship extraction accuracy rate reached 89.3%. Key information such as the type, distribution, age, and contact relationship with the surrounding rocks of the Mesozoic magmatic rocks in the region were effectively extracted.

[0144] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting key information from geological literature based on a large model, characterized in that, Includes the following steps: Collect geological literature data from a preset area and preprocess the geological literature data; the geological literature data includes text data, map data, and tabular data. Geological knowledge graphs are used to extract literature information data from multi-source geological data, and a multi-granularity semantic alignment mechanism is used to obtain semantic alignment relationships by establishing corresponding hierarchical relationships in the literature information data. Initial weights are set for the literature information data based on the geological rule base and the semantic alignment relationship. Optimized weights are obtained by dynamically adjusting the initial weights based on reinforcement learning-based decision routing. Finally, heterogeneous knowledge graphs are fused based on the literature information data and the optimized weights to obtain fused data. This includes: The weight adjustment task is modeled as a Markov decision process, and dynamic decision-making is carried out through state, action, and reward loops. The geological entity type and literature topic to be extracted are the task context, the identified entity set and attributes are the extracted information, and the credibility score of each data source is the multi-source data reliability label. The adjustment of the weights of multi-source data includes three basic actions and magnitude control. The three basic actions are to enhance text weight, increase map weight, and integrate multi-source weights. The weight adjustment range for enhancing text weight is +0.1 to +0.3, the weight adjustment range for increasing map weight is +0.1 to +0.3, and the weight adjustment range for integrating multi-source weights is ±0.05 for each source weight. The reward value is calculated based on the consistency between the extraction results and the geological knowledge map, data reliability, and task objectives. The expression is: ； in The degree of matching between entity attributes and geological knowledge graphs. To ensure the weight allocation aligns with the credibility of the data source, To ensure the completeness of entity attribute extraction under the current weight, For matching degree weight, As a weight for fit, Integrity weight; An evaluation network and a target network strategy are introduced to optimize the deep Q-network in reinforcement learning. The deep Q-network structure consists of three fully connected layers and an activation function. The evaluation network outputs the Q-value of each action in real time, with the current state S as input and the Q-value of each action in the action space as output. The target network freezes and copies parameters from the evaluation network periodically to calculate the target Q-value. The three fully connected layers are the input layer, hidden layer, and output layer. Using initial weights based on a geological rule base as initial actions, the network is evaluated by inputting the current state S and outputting the Q-value of each action. The greedy strategy selects actions, adjusts weights, extracts entity attributes, calculates reward values, observes new states, and updates the Q-value. The expression is: ； in This is the current state. For the new state, As a discount factor, For the current action, For new actions, The evaluation network value under the new state and new action. The target network value under the current state and current action. To maximize the value of all possible new actions in the action space; Every 100 training steps, the parameters of the evaluation network are copied to the target network. The temporal difference error is calculated based on the absolute value of the difference between the evaluation network and the target network. Sample priorities are assigned according to the importance of the samples, and the sample sampling probability is calculated. Importance sampling weights are introduced to correct the loss function. The process is iterated until the policy network converges. The optimized weight set for the current task is then output as the optimized weights. A key information extraction model for geological documents is constructed based on the fused data. The key information extraction model for geological documents is then optimized by human-computer collaboration based on the result deviation. The data to be extracted is input into the key information extraction model for geological documents, and the extraction results are output.

2. The method for extracting key information from geological documents based on a large model according to claim 1, characterized in that, Methods for obtaining literature information data by jointly extracting information from multi-source geological data using geological knowledge graphs include: For the text data: pre-trained word vectors based on geological word vectors, and the word vector space that is easily confused is corrected by a geological terminology dictionary; character-level BPE segmentation is used for obscure terms, and after the input text is encoded by BERT, the contextual dependencies are captured by a bidirectional long short-term memory network, and the entity label sequence is obtained by combining the conditional random layer with the geological entity transition probability matrix; the entity disambiguation algorithm is used to match geological knowledge graph nodes, and for entities with the same name, subject labels are assigned based on contextual similarity to obtain the text entity set; For map data: DBNet algorithm is used to detect text regions in the map, CRNN model is used to identify text content, spatial coordinates of text detection boxes are extracted, and combined with map scale to convert to actual geographic coordinates or relative positions, establish spatial association between text and entities, and output map entity set; For tabular data: Based on text density and table line features, the table is divided into high-density data area and low-density annotation area. For borehole tables, the three-dimensional data of depth, lithology and fossils are automatically extracted and arranged. The numerical data is standardized by unit, the qualitative description is mapped to quantitative attribute values, matched with the knowledge graph attribute dictionary, and the output is a table entity set. The method combines rule base and statistical learning to extract entity pair relationships, calls the geological knowledge graph reasoning engine to verify the relationships, automatically marks contradictory relationships as pending confirmation, and outputs non-contradictory relationships as a set of verified relationships; rule-driven: when an entity is followed by ± value, an entity plus value attribute relationship is automatically established; statistical-driven: relationships are identified by combining co-occurrence frequency with BERT relationship classifier. Output the text entity set, figure entity set, table entity set, and verification relation set as document information data.

3. The method for extracting key information from geological documents based on a large model according to claim 1, characterized in that, A method for obtaining semantic alignment relationships by performing corresponding hierarchical relationships on the document information data using a multi-granularity semantic alignment mechanism includes: The entities in the literature information data are labeled with granularity, and the rationality of the granularity is verified by the geological ontology database. Entities with mixed granularity are split or merged, and the entities in the literature information data are layered according to a 5-level granularity system, where the 5 levels are domain level, discipline level, type level, instance level, and attribute level. Semantic association of entities at different granularities is performed based on dual verification of word vector similarity and domain rules; Word vector similarity verification: The cosine similarity between entities is calculated using optimized geological word vectors. The similarity thresholds between domain-level and subject-level entities, and between subject-level and type-level entities are greater than or equal to 0.

85. The similarity thresholds between type-level and instance-level entities, and between instance-level and attribute-level entities are greater than or equal to 0.

78. Entities with hierarchical relationships but below the similarity threshold are subject to rule verification. Domain rule validation: If the definition of entity A contains the core features of entity B, then establish a hierarchical relationship between A and B; if entity B inherits all the core attributes of entity A, then confirm the hierarchical relationship; entities with semantic conflicts are marked as parallel relationships. By constructing complete hierarchical relationship chains for entities of different granularities through hierarchical transmission and conflict resolution, the structured semantically aligned relationship table is output. Hierarchical transmission automatically completes the relationship chain based on transitivity rules. Conflict resolution: when an entity has multiple potential superior concepts, subject tags are selected first, and intermediate nodes are automatically inserted for relationships that skip intermediate levels.

4. The method for extracting key information from geological documents based on a large model according to claim 1, characterized in that, A method for setting the initial weights of the document information data based on the geological rule base and the semantic alignment relationship includes: The rule base includes entity type rules, data source rules, and attribute importance rules, and the weight benchmark values ​​are determined through expert consensus. The baseline weights are corrected by multi-granular semantic alignment. For hierarchical relationships, the subordinate entity inherits the weight of the superior entity and increases the weight. For synonym relationships, the weight of authoritative domain terms is given priority. Based on the rule base, entity type matching and data source matching are performed on entities in the literature information data to obtain entity type value and data source weight. Semantic alignment correction value is obtained by correcting the semantic alignment baseline value. The entity type value and data source weight are weighted and summed, and the semantic alignment correction value is added to obtain the entity weight. Attribute type matching is performed on the attributes associated with entities, and attribute weights are obtained by correcting the reliability of attribute values ​​based on quantitative attributes, qualitative data, precise numerical values, and fuzzy descriptions. The system performs relationship type matching between entities, and obtains relationship weights by adjusting the evidence strength based on multi-source evidence, single-source evidence, knowledge graph verification passed, and relationships to be verified; the entity weights, attribute weights, and relationship weights are output as the initial weights.

5. The method for extracting key information from geological documents based on a large model according to claim 1, characterized in that, A method for obtaining fused data by fusing heterogeneous knowledge graphs based on the literature information data and the optimized weights includes: Based on the entities and relationships within the document information data, text subgraphs, image subgraphs, structured subgraphs, and external data subgraphs are obtained. A multimodal subgraph is constructed based on these subgraphs, and fusion priorities are assigned to each subgraph based on optimized weights. The text subgraph contains document text entities and attribute relationships, with weights initialized according to entity extraction confidence. The image subgraph contains spatial entities and coordinate attributes, with weights dynamically adjusted based on the recognition accuracy of the deep bounding box network and convolutional recurrent neural network. Structured subgraphs and external data subgraphs are also included. The BERT-base-cased model is used to encode the entity attribute text. A 128-dimensional feature vector is obtained through max pooling. The embedding space is optimized by fusing it with geological word vectors. The spatial topological relationship of map entities is learned by using hyperbolic graph convolutional network. The entity features are trained by adaptive gradient algorithm, and the depth, lithology and fossil triples are transformed into structured embedding vectors. By comparing the similarity of the embedding vectors of entities in different subgraphs, the geological knowledge graph reasoning engine is invoked to verify the rationality of the alignment, and subject labels are assigned to entities with the same name by combining optimized weights and contextual keywords. The confidence level of multi-source evidence for the same relationship is calculated by weighting it with optimized weights. A relationship confidence threshold of 0.6 is set, and relationships with confidence levels below the threshold are marked as requiring manual verification. Using verified relationships in the knowledge graph to infer missing relationships, and combining geophysical data subgraphs to complete relationships between entities across subgraphs; Output the entity ID, name, type, multi-source attributes, weight value, relationship ID, head / tail entity ID, relationship type, confidence level, and weight source as fused data.

6. The method for extracting key information from geological literature based on a large model according to claim 1, characterized in that, The method for constructing a key information extraction model from geological literature based on the fused data includes: The TransE algorithm is used to map knowledge graph entities in the fused data to a vector space, so that the vector distance between semantically similar entities is less than the similarity threshold, and the distance between dissimilar entities is greater than the dissimilarity threshold. For multi-source entities in the fused data, the feature vector is fused by weighted average. The weights are dynamically allocated proportionally to the vectorized entity features, and geological relationship path features are introduced to enhance the relationship features. A framework for a key information extraction model from geological documents is constructed based on a multimodal feature encoder, a dynamic weight fusion layer, and a relational reasoning and information extraction layer. Multimodal feature encoder: A geological-domain fine-tuned BERT input text fragment is used to output a contextual feature vector, obtaining text feature encoding; based on the spatial coordinates of map entities and text content, visual features are extracted through a convolutional neural network, concatenated with text features, and then weighted through an attention mechanism to obtain map feature encoding; for structured tabular data, Transformer-XL is used to capture long sequence dependencies, outputting a temporal feature vector to obtain table feature encoding; the weights of text feature encoding, map feature encoding, and table feature encoding are dynamically allocated through a multimodal attention layer; Dynamic Weight Fusion Layer: Based on reinforcement learning decision routing, dynamic weight adjustment of multi-source features is performed. The current task to be extracted, the fused feature vector, and the reliability label of multi-source data are defined as the state. The action space is defined by adjusting the weights of each modality feature and the given weight value range. The weights are adjusted by a given reward function, which includes positive reward and negative penalty. The positive reward is the matching of the extraction result with the geological knowledge graph, and the negative penalty is the entity type error. Relational Reasoning and Information Extraction Layer: BiLSTM-CRF is used to identify entity boundaries, and a geological entity dictionary is combined for entity boundary detection; attribute values ​​are bound to entities through an entity-attribute attention matrix to obtain attribute value associations; entity pair relation probabilities are calculated based on a knowledge graph embedding model, and relation reasoning is performed based on entity pair relation probabilities; the entity set, attribute set, and relation are output as extraction results; Training data is constructed, and multi-task joint training is carried out using a main task and auxiliary tasks. During the training process, geological constraints are added to correct the prediction results that violate the rules through penalty terms.

7. The method for extracting key information from geological documents based on a large model according to claim 1, characterized in that, A method for human-machine collaborative optimization of the geological literature key information extraction model based on the result deviation includes: Set confidence thresholds for entity recognition and relation extraction, and mark low-confidence samples as samples to be verified; call the geological knowledge graph inference engine to identify conflicting data that contradicts the domain rules, and the results include entity type conflicts, relational logic contradictions, and attribute value anomalies; Compare the extraction results of text, images, and tables: if the same entity has conflicting attributes in different data sources, mark it as a cross-modal conflict sample; classify and statistically analyze high-frequency error types by entity, relationship, and attribute, and push them to obtain an error sample pool; The system performs entity-level, relation-level, and attribute-level corrections on the erroneous sample pool. When experts correct entities, it automatically recommends standard attributes of similar entities in the knowledge graph and displays historical amendment examples of this type of error to assist experts in making quick decisions and obtaining corrected samples. Entity-level correction supports entity type adjustment and boundary calibration; relation-level correction provides a drop-down selection of relation types and supports relation direction adjustment; attribute-level correction supports unit calibration and error range supplementation for numerical attributes, and allows free editing of text attributes. The format of the corrected samples is standardized, and model parameters are extracted based on the corrected samples and key information from geological literature for domain adaptation fine-tuning and small sample learning optimization.

Images