A geoscience big data knowledge completion method, system, device and medium
By sampling neighborhood subgraphs and processing semantic mosaics in geoscience knowledge graphs, and combining this with a pre-built parameter-frozen LLM, a lightweight classifier is trained, solving the problems of incompleteness and sparsity in geoscience big data knowledge graphs and achieving efficient and reliable knowledge completion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INSTITUTE OF GEOLOGY AND GEOPHYSICS CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2026-05-18
- Publication Date
- 2026-07-31
AI Technical Summary
Geoscience big data knowledge graphs suffer from incompleteness and sparsity. When large language models are in a frozen state, they are affected by entity ambiguity and domain terminology, leading to discriminative distortion. Furthermore, fine-tuning is costly and difficult to afford due to limited computing resources.
By sampling neighborhood subgraphs from geoscience knowledge graph data, performing semantic jigsaw puzzle block division and linearization, binary and multi-class sample sets are constructed. The hidden states of intermediate layers are extracted using pre-built parameter-frozen LLM, and a lightweight classifier is trained for knowledge completion.
It achieves efficient and reliable knowledge completion, reduces computing power and engineering costs, improves entity disambiguation and semantic alignment capabilities, and reduces the risk of generative illusions.
Smart Images
Figure CN122198086B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of geoscience knowledge graph technology, and in particular to a method, system, device and medium for geoscience big data knowledge completion. Background Technology
[0002] Geoscience big data comes from a wide range of sources, is heterogeneous, massive in scale, and continues to grow. To achieve cross-source fusion and inference analysis in resource exploration, geological hazard assessment, environmental geology, and land space governance, it is typically necessary to construct a geoscience knowledge graph (KG). However, geoscience knowledge graphs generally suffer from incompleteness and sparsity, requiring knowledge graph completion (KGC) methods to predict potential relationships and discover missing facts. Currently, KGC based on large language models exists, but it suffers from the following problems: Large Language Models (LLMs) possess strong semantic modeling capabilities and rich implicit world knowledge. However, when reasoning through direct cues in a frozen state, they often suffer from entity ambiguity and domain terminology, leading to illusions or inconsistencies. Furthermore, entity names in geoscientific knowledge graphs may be numbers, abbreviations, or technical terms, and many KGs lack high-quality entity descriptions. A frozen LLM struggles to correctly link entity names to their internal semantic concepts, resulting in distorted KGC judgments. While fine-tuning large language models can achieve better task alignment and effectively reduce illusions through parameter optimization, it requires significant computational and memory resources and relies on ample domain-labeled data. However, in geoscientific big data scenarios, high-quality labeled samples are scarce, and computational resources are limited, making these costs often unbearable in reality. Summary of the Invention
[0003] The purpose of this application is to provide a method, system, device, and medium for knowledge completion in geoscience big data scenarios, which can achieve high-efficiency, high-reliability, and high-accuracy knowledge completion.
[0004] To achieve the above objectives, this application provides the following solution.
[0005] Firstly, this application provides a method for knowledge completion using geoscientific big data, including: Acquire geoscience knowledge graph data; For any entity in the aforementioned geoscience knowledge graph data, sample a neighborhood subgraph; For the neighborhood subgraph, semantic puzzle pieces are divided, linearized, and assembled according to semantic facets to obtain entity descriptions; Based on the aforementioned geoscience knowledge graph data, a binary classification sample set and a multi-classification sample set for relationship prediction are constructed. For any sample in the binary classification sample set and the relation prediction multi-class sample set, different stimulus prompts are constructed in combination with the entity description for different types of geoscience knowledge graph knowledge completion tasks. The stimulus cues are input into a pre-built LLM with frozen parameters, and the hidden states of the intermediate layers are extracted as task-driven features. Based on the binary classification sample set and the relation prediction multi-class sample set, a lightweight classifier is trained using the task-driven features corresponding to each sample as input to obtain a knowledge completion model. Obtain the triples to be completed and construct the corresponding stimulus cues to be used. Input the stimulus cues to be used into the pre-constructed parameter-frozen LLM to obtain the task-driven features to be used. Input the task-driven features to be used into the knowledge completion model to obtain the completed triples.
[0006] Secondly, this application provides a geoscience big data knowledge completion system, including: The knowledge graph acquisition module is used to acquire geoscience knowledge graph data; The subgraph sampling module is used to sample a neighborhood subgraph for any entity in the geoscience knowledge graph data; The jigsaw puzzle block partitioning module is used to perform semantic jigsaw puzzle block partitioning, semantic jigsaw puzzle block linearization processing and assembly processing on the neighborhood subgraph according to semantic facets to obtain entity descriptions; The sample set construction module is used to construct a binary classification sample set and a multi-classification sample set for relationship prediction based on the geoscience knowledge graph data. The stimulus prompt construction module is used to construct different stimulus prompts in combination with the entity description for any sample in the binary classification sample set and the relation prediction multi-class sample set, corresponding to different types of geoscience knowledge graph knowledge completion tasks. The hidden state extraction module is used to input the stimulus cue into the pre-constructed parameter-frozen LLM and extract the hidden states of the intermediate layer as task-driven features; The model training module is used to train a lightweight classifier based on the binary classification sample set and the relation prediction multi-class sample set, using the task-driven features corresponding to each sample as input, to obtain a knowledge completion model. The knowledge completion module is used to obtain the triples to be completed and construct the corresponding stimulus cues to be used. The stimulus cues to be used are input into the pre-constructed parameter-frozen LLM to obtain the task-driven features to be used. The task-driven features to be used are input into the knowledge completion model to obtain the completed triples.
[0007] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a geoscience big data knowledge completion method.
[0008] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for knowledge completion of geoscientific big data.
[0009] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application samples a neighborhood subgraph of any entity in the geoscientific knowledge graph data, performs semantic jigsaw puzzle block division, semantic jigsaw puzzle block linearization processing, and assembly processing to obtain an entity description. Specifically, the processing of semantic jigsaw puzzle blocks bridges the symbol-semantic gap, improving the disambiguation and semantic alignment capabilities of geoscientific entities. Simultaneously, it enables accurate semantic disambiguation for the subsequent use of pre-built parameter-frozen LLMs, thereby ensuring the accuracy and reliability of subsequent knowledge completion. In particular, constructing multi-faceted semantic jigsaw puzzle blocks from the local subgraph of the target entity and arranging them descriptively into a model-friendly entity description provides a high-information-density semantic context under structural constraints for frozen LLMs, significantly improving entity semantic alignment capabilities and reducing misjudgments caused by relying solely on entity names.
[0010] This application directly uses a pre-built parameter-frozen LLM, eliminating the need for fine-tuning large model parameters, significantly reducing computational power and engineering costs, and improving efficiency. Only the intermediate hidden state representations of the frozen LLM are read, and a lightweight classifier is trained to complete the knowledge completion, avoiding the large-scale GPU memory and long training time required for efficient full-parameter or parameter fine-tuning of the LLM. This achieves orders-of-magnitude GPU memory savings and training acceleration while maintaining near-fine-tuning performance. Stimulus cues are input into the pre-built parameter-frozen LLM, and the intermediate hidden states are extracted as task-driven features. The intermediate representations are used to detect and reduce the risk of generative hallucinations, improving the reliability of the completion.
[0011] This application utilizes the strong semantic priors of the hidden representations in the intermediate layers of an LLM with pre-constructed parameters and the structural constraints of the semantic puzzle of the subgraph to train an effective classifier with fewer training samples. This enables efficient and reliable knowledge completion of geoscientific knowledge graphs, making it suitable for small-sample or incremental learning. Attached Figure Description
[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0013] Figure 1 This is a diagram illustrating the application environment of the geoscience big data knowledge completion method in one embodiment of this application.
[0014] Figure 2 This is a flowchart illustrating a geoscience big data knowledge completion method in one embodiment of this application.
[0015] Figure 3 This is a schematic diagram of the algorithm architecture of the geoscience big data knowledge completion method in one embodiment of this application.
[0016] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0018] This application avoids fine-tuning costs by directly reading the hidden state of the intermediate layer of the LLM without updating the LLM parameters. On the other hand, it utilizes local subgraph facts to construct narrative entity descriptions under structural constraints, reducing entity ambiguity and improving semantic alignment, thereby achieving highly efficient and reliable knowledge completion in geoscience big data scenarios.
[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0020] The geoscientific big data knowledge completion method provided in this application can be applied to, for example... Figure 1 The application environment shown is illustrated. Terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be set up independently, integrated into server 102, or placed in the cloud or on another server.
[0021] Terminal 101 can send geoscientific knowledge graph data to server 102. Upon receiving the data, server 102 samples the neighborhood subgraph for any entity, then performs semantic jigsaw puzzle partitioning, semantic jigsaw puzzle linearization, and assembly based on semantic facets to obtain the entity description. Based on the geoscientific knowledge graph data, it constructs a binary classification sample set and a multi-classification sample set for relation prediction. For any sample, corresponding to different types of geoscientific knowledge graph knowledge completion tasks, it constructs different stimulus cues based on the entity description. The stimulus cues are input into a pre-constructed LLM with frozen parameters, and the hidden states of the intermediate layers are extracted as task-driven features. A lightweight classifier is trained based on this to obtain the knowledge completion model. Server 102 can then feed back the obtained knowledge completion model to terminal 101.
[0022] Terminal 101 can also send the triples to be completed to server 102. After receiving the triples, server 102 constructs the corresponding stimulus cues and inputs them into the pre-constructed parameter-frozen LLM to obtain the task-driven features. Then, the task-driven features are input into the knowledge completion model to obtain the completed triples. Server 102 can feed back the obtained completed triples to terminal 101.
[0023] Furthermore, in some embodiments, the geoscientific big data knowledge completion method can also be implemented independently by server 102 or terminal 101. Terminal 101 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Server 102 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.
[0024] In one exemplary embodiment, such as Figure 2 As shown, a method for knowledge completion in geoscientific big data is provided. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is applied to... Figure 1 Taking server 102 as an example, the explanation includes the following steps 201 to 208.
[0025] Step 201: Obtain the geoscience knowledge graph data G. Construct or integrate entity-relationship triples from multi-source heterogeneous geoscience data (geological maps, remote sensing interpretation, borehole columnar sections, seismic / geochemical observations, survey reports, etc.) to obtain the geoscience knowledge graph data G.
[0026] The geoscience knowledge graph data G includes: triples; structured attributes such as entity type, ontological constraints, spatial location (latitude / longitude / projected coordinates), temporal range, attribute values and their units; and raw text fragments from geological reports, literature, remote sensing interpretation descriptions, etc. The entity set is E, the relation set is R, and the triple set is T={(h,r,t)}, where h,t... E, h are the head entity, t are the tail entity, r is the relation, r R, e E. The entity name is denoted as EN(e).
[0027] Before proceeding with the next steps, it is necessary to unify the spatial coordinate reference system, time format, and units of measurement in the geoscience knowledge graph data G, and to constrain and label the entity types for subsequent candidate filtering and semantic mosaic faceting.
[0028] Step 202: For any entity in the geoscience knowledge graph data, sample the neighborhood subgraph.
[0029] In a specific application, for any target entity e∈E, its k-hop neighborhood subgraph Subgraph_k(e) is sampled from the geoscience knowledge graph data. Here, k=1 (one-hop subgraph), which can also be selectively extended to k=2 according to the relation type to cover the geological genetic semantic information carried by multi-hop relation paths such as "tectonic unit-stratigraphic unit-lithology-mineralization alteration".
[0030] In practical applications, when sampling k-hop neighborhood subgraphs from geoscientific knowledge graph data, a breadth-first search (BFS) approach can be used. The steps are as follows: Starting from entity e, expand along any relation type, collecting all directly adjacent triples within k steps (k=1 or 2). When k=1, the sampling result is the set of all triples satisfying the form (e,r,t) or (h,r,e); when k=2, it further includes triples directly adjacent to 1-hop neighbors. To control the subgraph size, a maximum sampling limit can be set for entities with excessively high degree (e.g., a maximum of 50 triples per hop), and uniform random sampling or sampling based on relation type weights can be used for truncation to obtain the neighborhood subgraph.
[0031] Step 203: For the neighborhood subgraph, perform semantic puzzle block division, semantic puzzle block linearization processing and assembly processing based on semantic facets to obtain entity descriptions.
[0032] In a specific application, step 203 includes the following steps (31)-(34).
[0033] (31) Based on the preset semantic facet set, the neighborhood subgraph is divided into multiple semantic puzzle blocks; specifically, the facts in Subgraph_k(e) obtained in the previous step 202 are divided into several semantic puzzle blocks P_f(e), where f is the semantic facet index.
[0034] The preset semantic facets set includes: type / definition facets, spatial facets, temporal facets, attribute facets, process / cause facets, and observation / evidence facets.
[0035] Type / Definition Facet: Geological category of e (fault, fold, stratigraphic unit, mineral deposit, remote sensing feature, etc.).
[0036] Spatial facets: e is located in / passes through / adjacent to administrative regions, tectonic units, watersheds, tectonic plates, etc.
[0037] Time-based aspects: formation era, observation time window, geological period, etc.
[0038] Attributes are divided into: lithology, composition, physical properties, scale, strike and dip, etc.
[0039] Process / genesis aspects: sedimentary environment, tectonic evolution, mineralization, etc.
[0040] Observation / Evidence Division: Data Sources (stations, remote sensing images, boreholes, profiles) and Evidence Chain.
[0041] (32) Perform linear colloquialization on the triples (hi,ri,ti) in the semantic puzzle block to obtain the short sentence V(EN(hi),EN(ri),EN(ti)). In a specific application, obtain a predefined colloquialization template for geoscientific relations (such as "EN(h) is located in EN(t)", "EN(h) was formed during the EN(t) period", etc.); fit the triples in the semantic puzzle block into the predefined colloquialization template for geoscientific relations to generate the short sentence.
[0042] (33) All the short sentences are spliced together in a preset facet order to obtain the structured short sentence description D1(e).
[0043] The preset facet order can be "Type-Space-Time-Attribute-Cause-Data Source", and short sentences can be concatenated into natural language paragraphs using line breaks or semicolons. Example: D1(e)=CONCAT({V(EN(hi),EN(ri),EN(ti))|(hi,ri,ti)∈Subgraph_k(e)}); CONCAT(·) indicates concatenation by delimiter; V(·) indicates that a preset geoscientific relation template library can be used, or a rule-based synonym normalization method can be used to generate it.
[0044] (34) Call the generation model, using the structured short sentence description as a constraint, assemble natural language paragraphs through narrative logic, and mark them as entity descriptions corresponding to the entity.
[0045] Since the structured short sentence sequence and the LLM pre-training corpus have different distributions, this application uses D1(e) as the "constraint set" and assembles each faceted puzzle piece into a natural language paragraph according to the narrative logic of "definition → space → time → attribute → process → data source" to generate a model-friendly narrative description, that is, entity description D(e).
[0046] In a specific application, based on the structured short sentence description, factual constraints, length constraints, and terminology constraints are determined; the LLM narrative generation model (which is an independent model, or can be a small model, rule engine, etc.) and preset narrative logic are invoked; based on the factual constraints, the length constraints, and the terminology constraints, the LLM narrative generation model is used to transform the structured short sentence description into natural language paragraphs according to the preset narrative logic.
[0047] The constraints include: Factual constraint: D(e) should cover the facts corresponding to the selected key constraint phrases in D1(e). Length constraint: The length of D(e) should not exceed the preset token budget B to fit the context window of subsequent stimulus prompts. Terminology constraint: Standardized writing should be used for geoscientific names, units, geological ages, etc., to avoid ambiguity. In other words, the generated result must meet the above factual constraint, length constraint, and terminology standardization constraint.
[0048] The following prompt template can be used in this application: "User: Given an entity name 'EN(e)', and the entity satisfies the following constraint: D1(e). Please generate an entity description that satisfies the constraint for geoscientific knowledge graph reasoning."
[0049] In addition, a consistency check can be performed after generation: D(e) is reverse-parsed into a set of constraints and compared with D1(e). If the coverage is lower than the threshold τ, a rewrite or supplement is triggered.
[0050] For static or slowly changing geoscientific knowledge graph data, this application caches D(e) in a description library. When incremental updates (addition / deletion of edges) occur in the geoscientific knowledge graph data, only the local subgraphs of the affected entities are resampled and the corresponding D(e) is updated to reduce overall overhead.
[0051] Step 204: Based on the geoscience knowledge graph data, construct a binary classification sample set and a multi-classification sample set for relationship prediction.
[0052] In a specific application, step 204 includes the following steps (41)-(43).
[0053] (41) Extract triplet samples from the geoscience knowledge graph data to construct a positive sample set.
[0054] (42) For the triplet samples in the positive sample set, with the same relation r fixed, the head entity or tail entity is randomly replaced to generate negative samples, thus constructing a negative sample set; the positive sample set and the negative sample set constitute a binary classification sample set. Among them, after the negative samples are generated, obviously invalid samples can be filtered out by constraints (type consistency, spatial / temporal legality, etc.) to improve the training quality.
[0055] The binary classification sample set obtained through the above steps (41) and (42) can achieve classification based on rationality (True / False binary classification).
[0056] (43) Extract triplet samples with different relation types r1, r2, ..., rn under the same head entity h from the geoscience knowledge graph data, and construct a relation prediction multi-classification sample set. The relation prediction multi-classification sample set obtained in this step can realize classification by relation type (multi-classification).
[0057] Step 205: For any sample in the binary classification sample set and the relation prediction multi-class sample set, construct different stimulus cues based on the entity description for different types of geoscientific knowledge graph knowledge completion tasks.
[0058] In a specific application, the geoscientific knowledge graph knowledge completion task includes at least three categories: triplet discrimination, relation prediction, and entity prediction. Different stimulus cues (Prompt(s)) are constructed for different task categories.
[0059] (1) Triple discrimination (binary classification), corresponding to the binary classification sample set, the stimulus cues are as follows: PT1: Is it really true that `Verbalize(EN(h),EN(r),EN(t))`? PT2: Background: 1. {D(h)} 2. {D(t)} Question: Is it true that Verbalize(EN(h),EN(r),EN(t))? (2) Relationship prediction (multi-class classification), corresponding to the multi-class sample set for relationship prediction, with the following stimuli: PT3: What is the relationship between EN(h) and EN(t)? Please choose your answer from r1|r2|...|rn.
[0060] PT4: Background: 1. {D(h)} 2. {D(t)} Question: What is the relationship between EN(h) and EN(t)? Please choose your answer from r1|r2|...|rn.
[0061] Where Verbalize(·) is the verbalization function for triples; {D(h)} and {D(t)} are the narrative subgraph semantic puzzle descriptions of the head entity h and the tail entity t, respectively.
[0062] (3) Entity prediction task: (h,r,?) or (?,r,t) can be fixed. Candidate entity names are filled into Verbalize one by one and sorted by classifier scores. The corresponding scores are used as cues. In this process, the subgraph semantic puzzle description D(e) is injected as background information to assist the classifier in accurate scoring. The specific implementation method is as follows: For the tail entity position (h, r, ?) to be predicted, the hint is constructed as: "Background: 1. {D(h)}. Question: Select the entity connected to EN(h) through relation EN(r) from the candidate list [e1, e2, ..., en]." Alternatively, construct triplet candidate hints (h, r, ei) one by one, inject the head entity description D(h) as background, and have the classifier score and sort each candidate triplet, selecting the Top-K candidate entities with the highest confidence.
[0063] Step 206: Input the stimulus cue into the pre-built parameter-frozen LLM and extract the hidden states of the intermediate layer as task-driven features.
[0064] Existing methods for direct fine-tuning or parameter-efficient fine-tuning of large language models in KGC improve task alignment by updating LLM parameters, but these methods suffer from high training costs, large memory consumption, and slow iteration. In geoscientific big data scenarios, frequent updates are difficult, and the models are more sensitive to data leakage and overfitting. Using only non-fine-tuned LLM generative completion for engineering prompts is computationally efficient but prone to illusions and severely affected by entity name ambiguity. For specialized geoscientific terms or strata / tectonic units with the same name, the error rate is high, making it difficult to meet engineering reliability requirements. A hybrid approach of "structural embedding (TransE / GNN, etc.) + LLM" utilizes structural embeddings to provide candidates or as external features, but requires additional training of large-scale graph embedding models and maintenance of embedding updates. When geoscientific KGs are continuously incrementally updated, embedding retraining or alignment costs are high, and there are still difficulties in integrating external embeddings with the LLM semantic space. Using only the final layer representation of LLM or only entity names without subgraph semantic jigsaw puzzles is simple to implement, but in many tasks, the final layer tends to generate output rather than discriminative information; the lack of subgraph structural constraints also leads to insufficient semantic alignment. This invention significantly improves discriminability and stability through intermediate layer readout, SMI layer selection, and subgraph semantic puzzle description generation.
[0065] In response to the above situation, this application uses a pre-built parameter-frozen LLM based on the use of puzzle pieces described above.
[0066] In a specific application, step 206 includes the following steps (61)-(64).
[0067] (61) Call the pre-trained autoregressive large language model as the pre-built parameter frozen LLM.
[0068] The LLM used for intermediate layer representation readout in this application can be directly called without additional training or specific fine-tuning, and the model parameters remain frozen throughout the entire process. Specifically, the models used in this application include, but are not limited to: 1) Generative large language models with autoregressive structures, such as the LLaMA series (Meta), Mistral series, GPT-2 / GPT-J, and other open-source models. This application can use open-source autoregressive large language models (such as LLaMA-2-7B, LLaMA-2-13B, Mistral-7B, etc.) that support obtaining the hidden state output of intermediate layers, deployed locally or on a private server, and calling the hidden state tensors output by each layer of its Transformer without calling its generation and decoding functions. This approach does not require modification of the model structure, is fully compatible with mainstream deep learning frameworks (PyTorch / HuggingFace Transformers), and has mature and reproducible engineering implementations.
[0069] Specifically, this application employs a large language model with an autoregressive structure (such as a pre-trained model based on the Transformer Decoder-Only architecture), which can leverage the characteristic that its last token gathers the contextual information of the entire sequence under the self-attention mechanism, thereby enabling Rep_l(s) to fully express the overall semantics of the stimulus cue.
[0070] 2) Encoder-decoder structure models, such as T5, Flan-T5, etc.
[0071] 3) Models accessible via commercial APIs (such as GPT-3.5) can obtain the hidden state of a specified layer through the API (intermediate layer output interfaces must be supported).
[0072] (62) Measure the correlation between the hidden states of different intermediate layers in the LLM with the pre-constructed parameters frozen and the knowledge completion task of the geoscience knowledge graph, and determine the optimal number of intermediate layers or the intermediate layer weight set. In order to automatically select the intermediate layer most suitable for the knowledge completion task, this application uses slice mutual information (SMI) to measure the statistical dependence of representation X and label Y.
[0073] Slice Mutual Information (SMI) is defined as follows: For the representation vector X (dimension d) and the task label Y, after mapping the high-dimensional vector to a one-dimensional scalar through a large number of random one-dimensional projection directions (uniformly sampled from a d-dimensional unit sphere), the (empirical) mutual information between the one-dimensional projection result and the label Y is calculated, and then the average of all projection directions is taken to obtain the final metric value. The specific implementation steps are as follows.
[0074] 1) Uniformly and randomly sample multiple projection directions from a d-dimensional unit sphere.
[0075] 2) For any of the projection directions theta_i, project the representation vectors of multiple samples into scalars to obtain a one-dimensional projection result sequence; wherein, the representation vector refers to any intermediate layer hidden state in the LLM with pre-constructed parameters frozen.
[0076] 3) Calculate the mutual information between the one-dimensional projection result sequence and the corresponding label sequence; wherein, the label sequence refers to the label sequence of any task type in the geoscience knowledge graph knowledge completion task.
[0077] 4) The mutual information of all projection directions is averaged to obtain the SMI value. This metric reduces the dimensionality of high-dimensional mutual information estimation to the average of multiple one-dimensional mutual information estimates through random slicing. It is robust in engineering implementation and suitable for high-dimensional representation evaluation with limited samples.
[0078] 5) Determine the optimal readout layer based on the SMI values corresponding to different intermediate layers in the LLM frozen by the pre-constructed parameters. This is then marked as the optimal number of intermediate layers. Among these, the optimal readout layer... Refers to a specified layer With the specified token location.
[0079] 6) Determine the intermediate layer weight set based on the SMI values corresponding to different intermediate layers in the LLM frozen by the pre-constructed parameters. Specifically, based on each candidate layer... The SMI value is used as the weight basis, and the weight vector (w_1, w_2, ..., w_L) is obtained after Softmax normalization of the SMI value of each layer.
[0080] (63) When the optimal number of intermediate layers is obtained, the stimulus cue is input into the pre-constructed parameter frozen LLM, and the hidden state of the intermediate layer corresponding to the optimal number of intermediate layers is read as the task-driven feature.
[0081] In a specific application, let the pre-built parameter-frozen LLM be M, and its total number of layers be L. For any sample s ∈ positive sample set S+ ∪ negative sample set S-, input the stimulus cue Prompt(s) into M, and from the th... Extract the vector corresponding to the j-th token from the hidden state sequence output by the layer as the intermediate layer representation of the sample: Rep_{l,j}(s)=M_{1: }(Prompt(s))[j], where 0 < <L,0<j<N。
[0082] Where N is the number of tokens.
[0083] The last token representation (i.e., j=N) of the pre-built parameter-frozen LLM with a Decoder-Only structure is used as the sample representation Rep_ (s) to fully aggregate contextual information.
[0084] (64) Upon obtaining the intermediate layer weight set, the stimulus cue is input into the pre-constructed parameter-frozen LLM, and all intermediate layer hidden features are read. These features are then weighted and fused with the intermediate layer weight set to obtain the final intermediate layer hidden state, which serves as the task-driven feature. Similarly to step (63), multiple candidate layers can be extracted in parallel to form a candidate representation set. Then, the representation vectors Rep_ of each candidate layer are... (s) The multi-layer representations are weighted and summed according to their respective weights to obtain the fused representation. Alternatively, a concatenation method can be used (the multi-layer representations are concatenated and then input into a larger-dimensional classifier), or a trainable attention mechanism can be used to dynamically weight the multi-layer representations.
[0085] This application ultimately selects the layer that maximizes SMI. This serves as the optimal readout layer; or, based on SMI, multiple layers are weighted and fused to obtain the final representation, i.e., the task-driven feature Rep. (s).
[0086] Step 207: Based on the binary classification sample set and the relation prediction multi-classification sample set, a lightweight classifier is trained using the task-driven features corresponding to each sample as input to obtain a knowledge completion model; wherein, the lightweight classifier includes a binary classifier and a multi-classifier, and the knowledge completion model includes a triplet true / false discrimination sub-model, a relation prediction sub-model, and an entity prediction sub-model.
[0087] For the true / false discrimination task (binary classification): using Rep (s) is the input, the sub-model is a binary classifier, and the output is the probability p that the triplet is true, p = Sigmoid(C(Rep)). (s))). During training, the labels are "positive sample = 1, negative sample = 0", and the labels are derived from the positive sample set S+ generated by the known graph triples and the negative sample set S- generated by negative sampling.
[0088] For relation prediction tasks (multi-class classification): using Rep (s) is the input, the sub-model is a multi-classifier, and the output is the probability distribution of each relation category. During training, the label is the integer index or one-hot encoding of the relation category, which is derived from the relation r of known triples in the geoscience knowledge graph data.
[0089] For entity prediction tasks: using Rep (s) is the input, and the sub-model is a multi-classifier. For each candidate entity e_c, candidate triples are formed and p_c is calculated. The Top-K completion results are obtained by sorting p_c. During training, the label is whether the candidate entity is a correct tail entity (binary), which is also derived from the known triples.
[0090] The labels used in the three tasks above are all derived from known triples in existing knowledge graphs, requiring no additional manual annotation. The overall training is conducted under weak supervision / self-supervised settings.
[0091] Binary or multi-class classifiers can employ two-layer MLPs, logistic regression, or support vector machines, etc. The above-mentioned use of three independent classifiers ensures that the optimization objectives of each task do not interfere with each other, achieving more robust training. In practical applications, a multi-task classifier can also be used, sharing Rep... The input is processed into binary and multi-class classifications through different output heads. Due to LLM freezing, only the classifier and a small number of cue parameters (if soft cueing is used) need to be trained, thus significantly reducing GPU memory and training time.
[0092] Steps 201-207 above describe the training process, which can be found in [reference needed]. Figure 3 The process can be summarized as follows: Geoscience KG access and normalization → Subgraph semantic puzzle generation and entity description generation → Task stimulus prompt construction → SMI layer selection → Freeze LLM intermediate layer representation readout → Lightweight classifier training / inference → Output completion results and confidence.
[0093] Step 208: Obtain the triples to be completed and construct the corresponding stimulus cues to be used. Input the stimulus cues to be used into the pre-constructed parameter-frozen LLM to obtain the task-driven features to be used. Input the task-driven features to be used into the knowledge completion model to obtain the completed triples.
[0094] In online knowledge completion scenarios, this application generates stimulus cues for the triples to be completed (such as (h,r,?), (?,r,t), or (h,?,t)), and then uses a frozen LLM as a representation encoder to extract the corresponding task-driven features (layers identical to those in the training phase). Alternatively, the frozen LLM still needs to be called during the online reasoning stage, but only forward propagation to the specified layer is performed, without performing generation and decoding, and the computational cost is much lower than that of the complete generation reasoning). The trained classifier (i.e. the knowledge completion model) obtains the confidence and ranking results, and finally outputs: (1) Top-K candidates for completed entities / relationships; (2) Confidence of each candidate; (3) Optional explanatory information (evidence fragments from the semantic puzzle of the subgraph).
[0095] In summary, this application relates to the field of intelligent processing of geoscientific big data and knowledge graphs. It uses a parameter-frozen LLM as a semantic / knowledge representation probe, and reads context-related hidden states from its designated intermediate layers as task-driven features. Simultaneously, it proposes an entity description generation mechanism for subgraph semantic mosaics. By sampling, faceting, compressing, and narrating the local subgraphs of the target entity, it forms an entity background description that combines structural constraints and natural language readability, bridging the gap between the symbolic space of geoscientific KG and the semantic space of LLM. Finally, it trains only lightweight classifiers (such as multilayer perceptron MLP, logistic regression, etc.) to complete the knowledge completion task. Because the parameters of the large language model are frozen throughout the process, this invention can significantly reduce GPU memory usage, computational overhead, and the requirement for labeled training samples while maintaining comparable accuracy to the fine-tuned model.
[0096] This application belongs to the fields of computer data processing, natural language processing and knowledge graph reasoning, and the executing entity is a geoscience big data knowledge completion system deployed on a server, workstation or cloud computing platform.
[0097] Based on the same inventive concept, this application also provides a system. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more system embodiments provided below can be found in the limitations of the method above, and will not be repeated here.
[0098] In one exemplary embodiment, a geoscience big data knowledge completion system is provided, comprising the following modules.
[0099] The knowledge graph acquisition module is used to acquire geoscience knowledge graph data.
[0100] The subgraph sampling module is used to sample neighborhood subgraphs for any entity in the geoscience knowledge graph data.
[0101] The jigsaw puzzle block partitioning module is used to perform semantic jigsaw puzzle block partitioning, semantic jigsaw puzzle block linearization processing, and assembly processing on the neighborhood subgraph based on semantic facets to obtain entity descriptions.
[0102] The sample set construction module is used to construct a binary classification sample set and a multi-classification sample set for relationship prediction based on the geoscience knowledge graph data.
[0103] The stimulus prompt construction module is used to construct different stimulus prompts for any sample in the binary classification sample set and the relation prediction multi-class sample set, corresponding to different types of geoscientific knowledge graph knowledge completion tasks, in combination with the entity description.
[0104] The hidden state extraction module is used to input the stimulus cues into the pre-built parameter-frozen LLM and extract the hidden states of the intermediate layer as task-driven features.
[0105] The model training module is used to train a lightweight classifier based on the binary classification sample set and the relation prediction multi-class sample set, using the task-driven features corresponding to each sample as input, to obtain a knowledge completion model.
[0106] The knowledge completion module is used to obtain the triples to be completed and construct the corresponding stimulus cues to be used. The stimulus cues to be used are input into the pre-constructed parameter-frozen LLM to obtain the task-driven features to be used. The task-driven features to be used are input into the knowledge completion model to obtain the completed triples.
[0107] The above modules can be deployed on the same computing node or in a distributed manner, and the knowledge completion module can provide service interfaces.
[0108] Compared to existing technologies, this application achieves knowledge completion of geoscientific knowledge graphs by utilizing frozen LLM intermediate layer context-related representations and a subgraph semantic jigsaw puzzle mechanism for geoscientific entities, without requiring fine-tuning of large language model parameters. This offers the following advantages: (1) No need to fine-tune large model parameters, significantly reducing computing power and engineering costs.
[0109] This application uses an LLM with frozen stimulus-cue input parameters, and reads the hidden state vector of a specified layer and a specified token position j as a representation. In an extended implementation, it also supports multi-layer candidate joint evaluation and weighted fusion of inter-layer representations to further improve the task adaptability of the representation. Then, a lightweight classifier (binary / multi-class) or scorer is trained using the above representation to achieve triplet true / false discrimination, relation prediction, and entity prediction; and can be combined with geoscientific ontology / spatiotemporal constraints for candidate filtering.
[0110] The above approach avoids the massive memory usage and long training time required for fully parametric or parameter-efficient fine-tuning of LLMs (such as LoRA). Under similar settings, this approach achieves orders-of-magnitude memory savings and training acceleration while maintaining near-fine-tuning performance.
[0111] (2) By using “subgraph semantic puzzle” to bridge the symbol-semantic gap, we can improve the ability of geoscientific entities to disambiguate and semantically align.
[0112] This application samples a one-hop / multi-hop subgraph for entity e; splits the facts of the subgraph into several semantic puzzle pieces (including type blocks, spatial blocks, temporal blocks, causal / process blocks, attribute blocks, observation / evidence blocks, etc.); assembles the puzzle pieces in a structured, colloquial, and narrative manner to generate entity descriptions that satisfy structural constraints and closely resemble the LLM pre-trained distribution; and provides description consistency verification and cache reuse.
[0113] Geoscientific entities commonly suffer from issues such as homonyms (e.g., the same stratigraphic group or tectonic zone name appearing repeatedly in different regions), multilingual / multi-version naming, and highly specialized abbreviations. Freezing LLM struggles to achieve accurate semantic disambiguation when relying solely on entity names as input. By constructing faceted semantic puzzle pieces from the local subgraphs of the target entity and narrating and assembling them into a model-friendly entity description D(e), freezing LLM is provided with a high-information-density semantic context under structural constraints. This significantly improves entity semantic alignment capabilities and reduces misjudgments caused by relying solely on entity names.
[0114] (3) Reduce the risk of generative illusions and improve the reliability of completion by using intermediate layer representation detection.
[0115] Existing LLM-based generative decoding methods are prone to illusions in knowledge graph completion tasks, where the model outputs relationships or entities inconsistent with the facts in the knowledge graph. This makes it difficult to meet the high standards of factual reliability required for geoscientific big data applications. This application transforms KGC into a representation reading + discrimination / ranking problem, avoiding the direct generation of facts and reducing the risk of incorrect completion due to illusions from a mechanism perspective. This approach is more suitable for scenarios with high reliability requirements, such as geoscientific disaster early warning and resource exploration.
[0116] (4) Introduce SMI information theory metric to achieve “interpretable layer selection” and improve cross-model and cross-dataset generalization.
[0117] This application employs information theory metrics such as Sliced Mutual Information (SMI) to measure the relevance of different layer representations to KGC labels, automatically selecting the optimal layer or generation layer weights. This process provides a quantitative basis for selecting readout layers or layer fusion, interpretably revealing "which layer carries more information relevant to geoscientific relationship inferences," and supports migration between different LLM backbones, reducing empirical parameter tuning.
[0118] (5) High data efficiency, suitable for scenarios where geoscientific annotation is scarce.
[0119] The construction of geoscientific knowledge graphs relies on manual annotation and verification by experts from multiple fields such as geology, geophysics, and remote sensing. This annotation process is costly and time-consuming, resulting in a generally limited number of available training samples. This application addresses this issue by freezing the strong semantic priors of LLM representations and the structural constraints of subgraph semantic mosaics, enabling the training of an effective classifier with fewer training samples. This approach is suitable for small-sample or incremental learning.
[0120] (6) Easy to deploy and update online.
[0121] This application provides description caching and incremental updates for updating geoscience knowledge graphs. The knowledge completion model can output confidence scores and evidence fragments, which facilitates integration with geoscience knowledge service platforms, intelligent question answering, geoscience large model intelligent agents, and other systems to achieve continuous completion and traceable reasoning.
[0122] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 4 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a geoscientific big data knowledge completion method.
[0123] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0124] In one exemplary embodiment, a computer device is also provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the above-described method embodiments.
[0125] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0126] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0127] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0128] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0129] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0131] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for geoscience big data knowledge completion, characterized in that, The method includes: Acquire geoscience knowledge graph data; For any entity in the geoscience knowledge graph data, a neighborhood subgraph is sampled; a breadth-first traversal is used to sample a k-hop neighborhood subgraph from the geoscience knowledge graph data; when k=2, it includes triples that are directly adjacent to 1-hop neighbors; For the neighborhood subgraph, semantic puzzle pieces are divided, linearized, and assembled based on semantic facets to obtain entity descriptions; including: Based on a preset semantic facet set, the neighborhood subgraph is divided into multiple semantic puzzle pieces; the preset semantic facet set includes: type / definition facet, spatial facet, temporal facet, attribute facet, process / cause facet, and observation / evidence facet; The triples in the semantic puzzle block are linearized into colloquial sentences. All the short sentences are concatenated according to a preset facet order to obtain a structured short sentence description; The generative model is invoked, and the structured short sentence description is used as a constraint. Natural language paragraphs are generated by assembling them through narrative logic and marked as entity descriptions corresponding to the entities. Based on the structured short sentence description, factual constraints, length constraints, and terminology constraints are determined, and each faceted puzzle piece is assembled into a natural language paragraph according to the narrative logic of "definition → space → time → attribute → process → data source". Based on the aforementioned geoscience knowledge graph data, a binary classification sample set and a multi-classification sample set for relationship prediction are constructed. For any sample in the binary classification sample set and the relation prediction multi-class sample set, different stimulus prompts are constructed in combination with the entity description for different types of geoscience knowledge graph knowledge completion tasks. The stimulus cues are input into a pre-built parameter-frozen LLM, and the hidden states of intermediate layers are extracted as task-driven features. The correlation between different hidden states of intermediate layers in the pre-built parameter-frozen LLM and the geoscience knowledge graph knowledge completion task is measured, and the optimal number of intermediate layers or intermediate layer weight set is determined. Based on the binary classification sample set and the relation prediction multi-class sample set, a lightweight classifier is trained using the task-driven features corresponding to each sample as input to obtain a knowledge completion model. Obtain the triples to be completed and construct the corresponding stimulus cues to be used. Input the stimulus cues to be used into the pre-constructed parameter-frozen LLM to obtain the task-driven features to be used. Input the task-driven features to be used into the knowledge completion model to obtain the completed triples. 2.The geoscience big data knowledge completion method according to claim 1, characterized in that, The process of linearizing the triples in the semantic puzzle block to obtain short sentences includes: obtaining a predefined colloquial template for geoscientific relations; and fitting the triples in the semantic puzzle block into the predefined colloquial template for geoscientific relations to generate short sentences. The generative model is invoked, using the structured short sentence description as constraints, to assemble and generate natural language paragraphs through narrative logic, including: Based on the structured short sentence description, factual constraints, length constraints, and terminology constraints are determined; Invoke the LLM narrative generation model and preset narrative logic; Based on the factual constraints, length constraints, and terminology constraints, the LLM narrative generation model is used to transform the structured short sentence descriptions into natural language paragraphs according to the preset narrative logic. 3.The geoscience big data knowledge completion method of claim 1, wherein, Based on the aforementioned geoscience knowledge graph data, a binary classification sample set and a multi-classification sample set for relationship prediction are constructed, including: Triple samples are extracted from the aforementioned geoscience knowledge graph data to construct a positive sample set; For the triplet samples in the positive sample set, the head entity or the tail entity is randomly replaced to generate negative samples, and a negative sample set is constructed; the positive sample set and the negative sample set constitute a binary classification sample set; Triple samples with different relation types under the same entity are extracted from the geoscience knowledge graph data to construct a multi-classification sample set for relation prediction. 4.The geoscience big data knowledge completion method of claim 1, wherein, The stimulus cues are input into a pre-built LLM with frozen parameters, and the hidden states of the intermediate layers are extracted as task-driven features, including: Call the pre-trained autoregressive large language model as a pre-built parameter-frozen LLM; The correlation between the hidden states of different intermediate layers in the LLM with the pre-constructed parameters frozen and the knowledge completion task of the geoscience knowledge graph is measured, and the optimal number of intermediate layers or the intermediate layer weight set is determined. When the optimal number of intermediate layers is obtained, the stimulus cue is input into the pre-constructed parameter frozen LLM, and the hidden state of the intermediate layer corresponding to the optimal number of intermediate layers is read as the task-driven feature. When the intermediate layer weight set is obtained, the stimulus cue is input into the pre-constructed parameter frozen LLM, and all intermediate layer hidden features are read. The intermediate layer weight set is then combined with the weighted fusion to obtain the final intermediate layer hidden state, which serves as the task-driven feature. 5.The geoscience big data knowledge completion method of claim 1, wherein, The knowledge completion task of the geoscience knowledge graph includes at least three categories: triplet discrimination, relation prediction, and entity prediction; correspondingly, the lightweight classifier includes a binary classifier and a multi-classifier, and the knowledge completion model includes a triplet true / false discrimination sub-model, a relation prediction sub-model, and an entity prediction sub-model. 6.The geoscience big data knowledge completion method of claim 4, wherein, The correlation between the hidden states of different intermediate layers in the pre-built parameter-frozen LLM and the knowledge completion task of the geoscience knowledge graph is measured using the slice mutual information (SMI) method, with the following steps: Multiple projection directions are uniformly and randomly sampled from a d-dimensional unit sphere. For any of the projection directions, the representation vectors of multiple samples are projected as scalars to obtain a one-dimensional projection result sequence; wherein, the representation vector refers to any intermediate layer hidden state in the LLM with pre-constructed parameters frozen; Calculate the mutual information between the one-dimensional projection result sequence and the corresponding label sequence; wherein, the label sequence refers to the label sequence of any task type in the geoscience knowledge graph knowledge completion task; The SMI value is obtained by averaging the mutual information of all projection directions. Based on the SMI values corresponding to different intermediate layers in the LLM frozen by the pre-constructed parameters, the optimal readout layer is determined and marked as the optimal number of intermediate layers; Based on the SMI values corresponding to different intermediate layers in the LLM frozen by the pre-constructed parameters, the intermediate layer weight set is determined.
7. A geoscience big data knowledge completion system, characterized in that, The system employing the geoscientific big data knowledge completion method according to any one of claims 1-6 comprises: The knowledge graph acquisition module is used to acquire geoscience knowledge graph data; The subgraph sampling module is used to sample the neighborhood subgraph for any entity in the geoscience knowledge graph data; The jigsaw puzzle block partitioning module is used to perform semantic jigsaw puzzle block partitioning, semantic jigsaw puzzle block linearization processing and assembly processing on the neighborhood subgraph according to semantic facets to obtain entity descriptions; The sample set construction module is used to construct a binary classification sample set and a multi-classification sample set for relationship prediction based on the geoscience knowledge graph data. The stimulus prompt construction module is used to construct different stimulus prompts in combination with the entity description for any sample in the binary classification sample set and the relation prediction multi-class sample set, corresponding to different types of geoscience knowledge graph knowledge completion tasks. The hidden state extraction module is used to input the stimulus cue into the pre-constructed parameter-frozen LLM and extract the hidden states of the intermediate layer as task-driven features; The model training module is used to train a lightweight classifier based on the binary classification sample set and the relation prediction multi-class sample set, using the task-driven features corresponding to each sample as input, to obtain a knowledge completion model. The knowledge completion module is used to obtain the triples to be completed and construct the corresponding stimulus cues to be used. The stimulus cues to be used are input into the pre-constructed parameter-frozen LLM to obtain the task-driven features to be used. The task-driven features to be used are input into the knowledge completion model to obtain the completed triples.
8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the geoscientific big data knowledge completion method according to any one of claims 1-6.
9. A computer readable storage medium having stored thereon a computer program, characterized in that, When executed by a processor, the computer program implements the geoscience big data knowledge completion method as described in any one of claims 1-6.