A geological mineralization cause and effect knowledge extraction method, storage medium, device and product

Abstract

The application discloses a geological mineralization cause-and-effect knowledge extraction method, a storage medium, equipment and products, relates to the geological mineralization field, and comprises the following steps: constructing a cause-and-effect hypothesis of a cause-and-effect role from an unstructured geological text; constructing a cause-and-effect data set based on the cause-and-effect hypothesis, and training a natural language reasoning model by using the cause-and-effect data set; wherein, an input vector of the cause-and-effect data set is fused with a vector of the cause-and-effect role to form an input vector of the model; a bias matrix of the cause-and-effect role is introduced to calculate self-attention weights; gate weights of the cause-and-effect role and a global context are obtained based on [CLS] embedding, the embedding of the cause-and-effect role and the [CLS] embedding are weighted and summed by using the gate weights, and finally, a fused representation is obtained; the trained model is used to identify a high-confidence cause-and-effect chain from the geological text, and a typical cause-and-effect mode related to a specific mineralization is obtained. The application can realize chain identification of a complex mineralization mode.

Description

Technical Field

[0001] This invention relates to the field of geological mineralization, and in particular to a method, storage medium, equipment, and product for extracting causal knowledge of geological mineralization. Background Technology

[0002] As geological exploration activities deepen, a large number of geological survey reports, scientific research papers, and mineral data accumulate in the form of unstructured text. These texts often contain rich knowledge of metallogenic regularities, especially causal chain information in the form of "cause-effect". However, how to automatically identify and extract these causal chains from unstructured text is a key issue restricting the intelligentization of mineral exploration and the construction of knowledge graphs.

[0003] Existing technologies for extracting causal information from geological texts mainly focus on a few categories of methods. Early research relied heavily on Named Entity Recognition (NER) and relation extraction techniques. These methods can identify geological entities and their relationships in text, but are often limited to simple "point-to-point" associations, making it difficult to express causal chains spanning multiple stages and lacking the ability to reason about complex mineralization patterns. Subsequent rule- or template-driven methods extract causal information through manually designed rules and pattern matching. While they can achieve certain results in specific scenarios, their over-reliance on human experience leads to poor generality and difficulty in adapting to complex geological descriptions of different mineral types and regions. In recent years, with the widespread application of pre-trained language models such as BERT and RoBERTa, semantic understanding capabilities have been significantly improved. However, directly applying general natural language reasoning (NLI) models still has shortcomings: on the one hand, these models are not adapted to the causal structure of the geological field, and the reasoning results often contain a lot of noise; on the other hand, they lack explicit perception of the "cause-mechanism-effect" ternary structure, making it difficult to achieve chain recognition of complex mineralization patterns. Summary of the Invention

[0004] The purpose of this invention is to address the lack of causal structure adaptation in existing geological mineralization technologies by proposing a method for extracting causal knowledge of geological mineralization, comprising the following steps:

[0005] S1. Obtain unstructured geological text, and construct causal hypotheses from the unstructured geological text, wherein the causal hypotheses involve causal roles;

[0006] S2. Construct a causal dataset based on the causal hypothesis, and use the causal dataset to train a natural language inference model;

[0007] Specifically, the input vector of the causal dataset is fused with the independent embedding vectors of the causal roles to form the input vector of the model;

[0008] When calculating the self-attention weights, a bias matrix for the causal role is introduced;

[0009] Based on the [CLS] embedding, the gating weights of the causal role and the global context are obtained. The gating weights are then used to perform a weighted summation of the causal role embedding and the [CLS] embedding to obtain the final fused representation.

[0010] S3. Use the trained model to identify high-confidence causal chains from geological texts to obtain typical causal patterns related to specific mineralization processes.

[0011] Furthermore, the structure of the causal hypothesis is expressed as: "X affects Z through Y". X, Y and Z are three types of causal roles. X represents the ore-forming element or the main body of geological action, Y represents the action mechanism or process, and Z represents the ore-forming result or geological response.

[0012] Furthermore, the input vector of the model is represented as:

[0013]

[0014] in, For the model's input vector, For word embedding vectors, For position embedding vectors, Embed vectors for causal roles, which include X, Y, Z, and O, where O represents other causal roles.

[0015] Furthermore, the causal dataset labels include three categories: support-mineralization, neutral, and opposition-non-mineralization.

[0016] Furthermore, when calculating the self-attention weights, a bias matrix for the causal role is introduced, and the new formula for calculating the attention weights is:

[0017]

[0018] in, This indicates a new focus. express function, Let A represent the attention score matrix, B represent the bias matrix of the causal role, and V represent the value vector.

[0019] Furthermore, the bias matrix of the causal role is the same dimension as the attention score matrix. If there is an interaction between X and Y or X and Z, a positive bias value is set; otherwise, a negative bias value is set.

[0020] Furthermore, the final fusion characterization is as follows:

[0021]

[0022]

[0023] in, , Let X, Y, and Z represent the gating weights for the causal roles and the global context, respectively. This represents the final fusion representation. , and These represent the embeddings of causal roles X, Y, and Z, respectively. Indicates [CLS] embedding, This represents the Sigmoid activation function. Represents the weight parameters. This represents the bias parameter.

[0024] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for extracting causal knowledge of geological mineralization.

[0025] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described method for extracting geological mineralization causal knowledge.

[0026] The present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described method for extracting causal knowledge of geological mineralization.

[0027] The beneficial effects of the technical solution provided by this invention are:

[0028] This invention constructs causal hypotheses of causal roles, builds a causal dataset based on these hypotheses, and trains a natural language inference model using this dataset. Semantic role embeddings are added to the input layer, and prior knowledge of a "cause-mechanism-effect" structure is explicitly injected into the input layer. A causal bias matrix is ​​introduced into the attention weight calculation, guiding the model to focus on the deep representations between causal components at the feature interaction level, reducing logical noise. Gated weights are used to weightedly fuse the vectors of each causal role with the global vector, generating a fused representation for final classification. This significantly enhances the interpretability of the model's decisions and addresses the lack of adaptation to causal structures in existing geological mineralization technologies. This invention can achieve chain-like recognition of complex mineralization patterns. Attached Figure Description

[0029] Figure 1 This is a flowchart of the geological mineralization causal knowledge extraction method according to an embodiment of the present invention;

[0030] Figure 2 This is a block diagram of an electronic device according to an exemplary embodiment of Embodiment 1 of the present invention;

[0031] Figure 3 This is a UMAP clustering visualization diagram of high-confidence support class causal chains in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0033] The flowchart of the geological mineralization causal knowledge extraction method of this invention is as follows: Figure 1 Specifically, it includes the following steps:

[0034] S1. Obtain unstructured geological text and utilize a Large Language Model (LLM) to generate candidate hypothesis sentences conforming to the structure "X affects Z through Y" based on the premises. Construct causal hypotheses from the unstructured geological text, involving causal roles X, Y, and Z. The structure of the causal hypothesis is represented as: "X affects Z through Y," where X represents the ore-forming element or geological agent, Y represents the mechanism or process, and Z represents the ore-forming result or geological response. This representation method fully preserves the core scientific logic in the causal chain, overcomes the defect of traditional "cause-effect" binary representations that result in the loss of mechanism information, and provides a clear structured target for machine understanding and computation.

[0035] S2. Construct a causal dataset based on causal assumptions. Use LLM to assign labels to each "premise-hypothesis" pair. The labels include three categories: support-affected, neutral, and opposition-non-affected, forming a structured causal dataset containing a large number of high-quality samples, providing supervision signals for model training.

[0036] This invention trains a natural language inference model using a causal dataset. Based on the standard RoBERTa-wwm-ext model, it makes the following three improvements:

[0037] (1) In the model input layer, semantic role labels are assigned to each lexical unit in the premise text according to the structure of the hypothesis sentence, and these labels are embedded into vectors, which are then added to the word embeddings and position embeddings, thereby explicitly injecting the "XYZ" structure prior knowledge into the input layer. For a given premise-hypothesis pair, we assign a semantic role label to each token in the premise text based on the explicit structure of the hypothesis sentence "X affects Z through Y". Then, a learnable embedding matrix is ​​used to map the discrete role labels to continuous role embedding vectors. The input representation of each token is formed by adding its word embedding, position embedding and role embedding, and its formal representation is as follows:

[0038]

[0039]

[0040] in, For the model's input vector, For word embedding vectors, For position embedding vectors, Embed vectors for causal roles, It is the embedding vector of character X. It is the embedding vector of character Y. It is the embedding vector of character Z. It is the embedding vector of other characters.

[0041] (2) When calculating the self-attention weights, a bias matrix of the causal role is introduced, and a bias matrix with the same dimension as the attention score matrix is ​​predefined. Its design principle is as follows: on a predefined causal chain (X→Y→Z), when attention occurs between positive roles in the causal chain (such as X→Y, Y→Z), a positive bias (+B) is applied at the corresponding position of the bias matrix B to enhance its attention weight and encourage the model to capture such causal relationships; when a causal role (X) interacts with an irrelevant role (O), or when the attention direction is opposite to the causal chain direction (such as Y→X, Z→Y), a negative bias (-B) is applied to suppress these invalid or erroneous associations.

[0042] The bias matrix B is superimposed onto the standard attention score matrix, and then softmax normalization is performed. The new attention weight calculation formula is:

[0043]

[0044] in, This indicates a new focus. express function, Let A represent the attention score matrix, B represent the bias matrix of the causal role, and V represent the value vector.

[0045] This operation applies structural constraints directly at the level of attention probability distribution without changing the model parameters.

[0046] (3) Let the hidden state output by the final layer of the model be H. Extract the global semantic vector corresponding to the [CLS] label from H. Simultaneously, mean pooling is calculated for the hidden states of all tokens labeled X, Y, and Z to obtain the corresponding role representation vectors. , , Based on the [CLS] embedding, the gating weights of the causal role and the global context are obtained. The gating weights are then used to perform a weighted summation of the causal role embedding and the [CLS] embedding to obtain the final fused representation.

[0047] global vector A gated weight vector g is generated through a linear transformation layer and a sigmoid activation function:

[0048]

[0049] The fused representation is obtained by using gating weights to sum the features.

[0050]

[0051] in, , Let X, Y, and Z represent the gating weights for the causal roles and the global context, respectively. This represents the final fusion representation. , and These represent the embeddings of causal roles X, Y, and Z, respectively. Indicates [CLS] embedding, This represents the Sigmoid activation function. Represents the weight parameters. This represents the bias parameter.

[0052] The constructed dataset was input into the Geo-CausalNLI model for training. During training, a class-weighted strategy and Focal Loss were employed to mitigate training bias caused by uneven class distribution. Simultaneously, an Early Stopping mechanism and a dynamic learning rate adjustment strategy (ReduceLROnPlateau) were introduced to ensure rapid convergence of the model under limited data conditions and to avoid overfitting. This optimized design enables the model to maintain high accuracy while possessing stronger generalization performance.

[0053] S3. Use the trained model to identify high-confidence causal chains from geological texts to obtain typical causal patterns related to specific mineralization processes.

[0054] After model training, the model is used to infer from unlabeled geological texts, automatically identifying potential causal chains. Specifically, the input is a "premise-hypothesis" pair, and the output is the prediction results of three types of labels and their confidence scores. By setting a reasonable confidence threshold (e.g., 0.8), high-confidence "support-mineralization" causal chains are selected. This process effectively reduces the interference of logical noise on the mining of mineralization patterns and ensures the reliability of subsequent pattern induction. Based on the causal chain identification, the KMeans clustering algorithm is used to aggregate high-confidence causal chains. To enhance the interpretability of the clustering results, this step introduces a "central chain extraction" and "keyword compression and induction" mechanism: First, by calculating the cosine similarity between the cluster center and each causal chain, the chain that best represents the cluster is selected as the "central chain"; then, high-frequency keywords in the same cluster are compressed and summarized to form a generalized causal pattern description. The resulting model can reveal typical action paths in the mineralization system, such as "fractures control ore body occurrence through fluid activity" or "Mn / Cr ratio affects manganese ore enrichment by reflecting reduction conditions".

[0055] The method proposed in this invention can not only automatically identify ore-forming causal chains in geological texts, but also obtain typical causal patterns related to specific mineral types through clustering and induction. This method can be widely applied to the understanding of patterns and prediction of deep mineral exploration for various mineral types, such as manganese and gold. On the one hand, researchers can use this method to construct a geological causal knowledge base and summarize regional metallogenic patterns; on the other hand, combined with existing geological survey data and model prediction results, it can assist in the prediction of deep mineral exploration target areas, providing more scientific and efficient decision support for mineral resource exploration and evaluation.

[0056] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the above-described method for extracting causal knowledge of geological mineralization.

[0057] Please see Figure 2 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0058] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the above-mentioned method for extracting geological mineralization causal knowledge.

[0059] In one exemplary embodiment, a computer program product is proposed, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method for extracting causal knowledge of geological mineralization.

[0060] In this invention, the causal dataset constructed in this embodiment contains a total of 8148 samples, including 5704 training samples, 1222 validation samples, and 1222 test samples. The dataset is divided into training, validation, and test sets in a ratio of 7:1.5:1.5, and the specific data distribution is shown in Table 1.

[0061] Table 1

[0062]

[0063] For each input "premise-assumption" pair, based on the explicit assumption in the hypothesis statement... <x> , <y> , <z>The tags assign the same semantic role label to the corresponding lexical units in the context text. Unlabeled lexical units are uniformly assigned the label "O" (Other).

[0064] To evaluate model performance, this invention employs an evaluation system comprised of overall metrics and categorical metrics. Overall metrics include accuracy and Macro-F1, where Macro-F1 is the arithmetic mean of the F1 scores for each class, used to measure the model's overall prediction accuracy and balanced performance under multi-class conditions. Categorical metrics include Entailment-F1, Neutral-F1, and Contradiction-F1, used to reflect the model's ability to identify "support / mineralization," "neutrality," and "opposition / non-mineralization" inference categories, respectively.

[0065] Two-dimensional visualization of clustering results is performed using the UMAP algorithm, as referenced. Figure 3 Different colors represent different categories of typical mineralization causal patterns. Clustering results show that the causal chains are divided into five patterns: Cluster 0 (tectonic-lithological control), Cluster 1 (geochemical-reduction), Cluster 2 (fluid-reduction), Cluster 3 (sedimentation-environment), and Cluster 4 (tectonic-stress). This result intuitively demonstrates the effectiveness and interpretability of the method of this invention in pattern induction.

[0066] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.< / z> < / y> < / x>

Claims

1. A method for extracting causal knowledge of geological mineralization, characterized in that, Includes the following steps: S1. Obtain unstructured geological text and construct causal hypotheses of causal roles from the unstructured geological text; S2. Construct a causal dataset based on the causal hypothesis, and use the causal dataset to train a natural language inference model; Specifically, the input vector of the causal dataset is fused with the independent embedding vectors of the causal roles to form the input vector of the model; When calculating self-attention weights, a bias matrix for the causal role is introduced; the new formula for calculating attention weights is: in, This indicates a new focus. express function, Let B represent the attention score matrix, B represent the bias matrix of causal roles, and V represent the value vector. The bias matrix B of the causal role has the same dimension as the attention score matrix. Based on the predefined causal chain, when attention occurs between positive roles in the causal chain, a positive bias is applied at the corresponding position of the bias matrix B; when the causal role in the causal chain interacts with a role unrelated to the causal chain, or when the attention direction is opposite to the causal chain direction, a negative bias is applied. Based on the [CLS] embedding, the gating weights of the causal role and the global context are obtained. The gating weights are then used to perform a weighted summation of the causal role embedding and the [CLS] embedding to obtain the final fused representation. S3. Use the trained model to identify high-confidence causal chains from geological texts to obtain typical causal patterns related to specific mineralization processes.

2. The method for extracting geological mineralization causal knowledge according to claim 1, characterized in that, The structure of a causal hypothesis is expressed as: "X affects Z through Y". X, Y and Z are three types of causal roles. X represents the mineralization element or the main body of geological action, Y represents the action mechanism or process, and Z represents the mineralization result or geological response.

3. The method for extracting geological mineralization causal knowledge according to claim 2, characterized in that, The input vector of the model is represented as: in, For the model's input vector, For word embedding vectors, For position embedding vectors, Embed vectors for causal roles, which include X, Y, Z, and O, where O represents other causal roles.

4. The method for extracting geological mineralization causal knowledge according to claim 1, characterized in that, The causal dataset labels include three categories: support-mineralization, neutral, and opposition-non-mineralization.

5. The method for extracting geological mineralization causal knowledge according to claim 2, characterized in that, The bias matrix for causal roles is the same dimension as the attention score matrix. If there is an interaction between X and Y or X and Z, a positive bias value is set; otherwise, a negative bias value is set.

6. The method for extracting geological mineralization causal knowledge according to claim 2, characterized in that, The final fusion characterization is as follows: in, , Let X, Y, and Z represent the gating weights for the causal roles and the global context, respectively. This represents the final fusion representation. , and These represent the embeddings of causal roles X, Y, and Z, respectively. Indicates [CLS] embedding, This represents the Sigmoid activation function. Represents the weight parameters. This represents the bias parameter.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

8. An electronic device, characterized in that, The device includes a processor and a memory, the processor being interconnected with the memory, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to perform the method as described in any one of claims 1-6.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-6.

Images