Method and device for retrieving mineral resource information based on text vector serialization

Abstract

The application discloses a kind of based on text vector serialization's mineral resources intelligence information retrieval method and device, belong to information retrieval technical field, solve the problem of unable to give consideration to cost and efficiency in prior art.Therein, based on text vector serialization's mineral resources intelligence information retrieval method, text is quantified by general vector model, the similarity of any text block and keyword is obtained, then keyword similarity is projected into text, the preliminary relationship between keyword and text is constructed, let two form a whole, improve matching degree.On this basis, by clustering, the elements of text structure are extracted, then the clustering is serialized, so that the text structure obtains digital expression, then this structure unit is image processing, the spatial position relationship between structure elements is captured, the accuracy of mineral resources intelligence retrieval is improved, and then the latest mineral resources dynamics can be understood more quickly, and coping strategies are formulated.

Description

Technical Field

[0001] This invention belongs to the field of information retrieval technology, and specifically relates to a method and apparatus for retrieving mineral resource intelligence information based on text vector serialization. Background Technology

[0002] Mineral resources refer to the general term for useful elements, compounds, minerals, and mineral aggregates that are naturally occurring in the near-crust or on the surface of the Earth, formed by geological processes, and possess economic value. They are natural resources existing in solid, liquid, and gaseous states. Mineral resources have natural, economic, and environmental attributes. They form the foundation of a country's manufacturing industry. To ensure mineral resource security, it is necessary to obtain key intelligence on mineral resources from multiple channels. Currently, mining publicly available online resources to obtain information on the dynamics of mineral resources in various countries and to understand the behavior of competitors in order to formulate countermeasures is the lowest-cost and most information-rich channel.

[0003] However, the internet is a vast and complex collection of information, filled with articles of all types. For strategic intelligence related to mineral resources, existing retrieval methods often include knowledge graph construction, keyword retrieval, and natural language processing using neural networks. However, these general-purpose search engines, even those powered by AI, cannot accurately understand the deeper meaning behind news reports and cannot meet the strategic analysis needs of mineral resources. Some key information initially has low visibility, appearing only in inconspicuous links on some websites, making it difficult to discover. Deliberate searching requires extensive reading, which is time-consuming and laborious, inefficient, and unlikely to yield useful information every time.

[0004] In summary, the inventors found that due to the massive amount of information on the internet and the presence of various noises, traditional retrieval methods cannot balance cost and efficiency. Summary of the Invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a low-cost, high-efficiency method and apparatus for retrieving mineral resource intelligence information based on text vector serialization, so as to solve the problem that the prior art cannot balance cost and efficiency.

[0006] The objective of this invention is achieved as follows: A method for retrieving mineral resource intelligence information based on text vector serialization includes: Obtain sample article data and keyword set; The sample article data is divided into multiple text blocks, and each text block and the keywords in the keyword set are vectorized to obtain text block vectors and keyword vectors. Obtain the similarity between any text block vector and each keyword vector, and based on each similarity, obtain the similarity feature information of any text block vector; Set the number of clusters, and cluster the text block vectors according to the number of clusters to obtain multiple clusters; Extract the set of position labels for each text block in each cluster, and assign label coefficients to each set of position labels; A one-dimensional structural sequence is generated based on the label coefficients, similarity feature information, and location label set. Convert a one-dimensional structural sequence into a template two-dimensional image; The system obtains candidate two-dimensional images of multiple candidate articles, and retrieves the target article from each candidate article based on the similarity between the template two-dimensional image and the candidate two-dimensional image.

[0007] In a preferred embodiment of the present invention, the step of segmenting template article data into multiple text blocks includes: Extract punctuation marks from template article data; Based on punctuation marks, the template article data is segmented into multiple text blocks.

[0008] In a preferred embodiment of the present invention, the step of obtaining similarity feature information of any text block vector based on various similarities includes: Sum the similarities of each text block vector to obtain the sum of similarities. Construct a summation vector based on the similarity sum of any text block vector; By normalizing the summed vector using the L2 norm, similarity feature information of any text block vector can be obtained.

[0009] In a preferred embodiment of the present invention, the step of setting the number of clusters is as follows: The number of key points in the sample article data; Set the number of clusters based on the number of key points.

[0010] A preferred embodiment of the present invention includes the following steps for generating a one-dimensional structural sequence based on label coefficients, similarity feature information, and a set of position labels: Based on the set of location labels, similarity feature information is processed to obtain a similarity set; Based on the set of position labels, the coefficients of each label are superimposed onto the similarity set to obtain a one-dimensional structure sequence.

[0011] In a preferred embodiment of the present invention, the reference coefficients are integers greater than or equal to 0.

[0012] In a preferred embodiment of the present invention, the step of converting a one-dimensional structural sequence into a template two-dimensional image includes: Convert a one-dimensional structure sequence into a two-dimensional sequence; The numerical values ​​in the two-dimensional sequence are converted into grayscale values, and a template two-dimensional image is generated based on the grayscale values.

[0013] In a preferred embodiment of the present invention, the step of converting a one-dimensional structural sequence into a two-dimensional sequence includes: The preset values ​​are added to the one-dimensional structure sequence to obtain the current structure sequence, so that the number of elements in the one-dimensional structure sequence can be squared. Divide the current structure sequence into k equal-length subsequences; where k is a positive integer; Based on the segmentation order, each equal-length subsequence is transformed into a two-dimensional matrix.

[0014] In a preferred embodiment of the present invention, the step of obtaining candidate two-dimensional images of multiple candidate articles includes: When the resolution of the candidate 2D image is greater than that of the template 2D image, primitive downsampling processing is performed on the candidate 2D image to obtain the downsampled image; When the resolution of the candidate 2D image is less than that of the template 2D image, primitive upsampling processing is performed on the candidate 2D image to obtain the upsampled image. The steps for retrieving the target article from the candidate articles include: Based on the similarity between the downsampled image and the template 2D image, the target article is retrieved from each candidate article, or; The target article is retrieved from each candidate article based on the similarity between the upsampled image and the template 2D image.

[0015] A mineral resource intelligence information retrieval device based on text vector serialization includes: The acquisition module is used to retrieve template article data and keyword sets; The vectorization module is used to divide the template article data into multiple text blocks, and to vectorize each text block and the keywords in the keyword set to obtain text block vectors and keyword vectors. The similarity calculation module is used to obtain the similarity between any text block vector and each keyword vector, and based on each similarity, to obtain the similarity feature information of any text block vector; The clustering module is used to set the number of clusters and to cluster the text block vectors according to the number of clusters to obtain multiple clusters; The label coefficient allocation module is used to extract the set of position labels for each text block in each cluster and assign label coefficients to each set of position labels. The sequence generation module is used to generate one-dimensional structure sequences based on label coefficients, similarity feature information, and position label sets; The conversion module is used to convert one-dimensional structural sequences into template two-dimensional images; The retrieval module is used to obtain candidate two-dimensional images of multiple candidate articles, and to retrieve the target article from each candidate article based on the similarity between the template two-dimensional image and the candidate two-dimensional image.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: The aforementioned mineral resource intelligence retrieval method based on text vector serialization quantifies the text using a general vector model, obtaining the similarity between any text block and keywords. Then, the keyword similarity is projected onto the text to construct a preliminary relationship between keywords and text, forming a unified whole and improving matching accuracy. Building upon this, clustering extracts structural elements from the text, and the clusters are serialized to digitally represent the text structure. This structural unit is then visualized to capture the spatial relationships between structural elements. Ultimately, without requiring massive amounts of labeled training data or pre-training, and avoiding reliance on enormous computing power, this method achieves control over both the macro-level textual structure and micro-level semantic information of articles, selecting the most relevant intelligence documents and improving the accuracy of mineral resource intelligence retrieval. This allows for faster understanding of the latest mineral resource dynamics and the development of appropriate strategies.

[0017] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings. Figure 1 A schematic flowchart illustrating the mineral resource intelligence information retrieval method based on text vector serialization provided by the present invention; Figure 2 A schematic flowchart illustrating the steps of dividing template article data into multiple text blocks, as provided by the present invention; Figure 3 A schematic flowchart illustrating the steps for obtaining similarity feature information of any text block vector based on various similarities, as provided by this invention; Figure 4A schematic flowchart illustrating the steps of generating a one-dimensional structural sequence based on label coefficients, similar feature information, and a set of position labels provided by the present invention; Figure 5 A schematic flowchart illustrating the steps of converting a one-dimensional structural sequence into a template two-dimensional image provided by the present invention; Figure 6 This is a diagram of the graphical text retrieval unit provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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, 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.

[0020] To facilitate understanding of the embodiments of this application, further explanation and description will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of this application. In the drawings, the dimensions and relative dimensions of components may be exaggerated for clarity and / or descriptive purposes. When exemplary embodiments can be implemented differently, a specific process sequence may be performed in a different order than that described. For example, two consecutively described processes may be performed substantially simultaneously or in the reverse order of their description. Furthermore, the same reference numerals denote the same components.

[0021] Currently, the following information retrieval methods are commonly used, each with its own problems, as detailed below: 1. Building a knowledge graph: This method typically requires manpower, involving manual selection of articles, designing article interpretation methods and tags, and building a knowledge graph. Future articles are then filtered according to the rules of the knowledge graph. However, the graph construction and iteration process relies heavily on manpower, resulting in high initial investment and subsequent maintenance costs.

[0022] 2. Rule-based keyword retrieval: This method pre-selects keywords and uses regular expressions to capture specific expressions, thereby extracting text that conforms to specific rules. However, this method lacks semantic understanding capabilities and relies on rule matching.

[0023] 3. Neural Network-Based Natural Language Processing: This method vectorizes text and extracts the top-ranked entries based on the similarity between vectors, thus completing the filtering. A typical example is using an embedding model to convert text into vectors of 768, 1024, or even higher dimensions, comparing the similarity between the sample and the target object (e.g., cosine similarity) to obtain a similarity ranking, and finally extracting the text with the closest semantics. However, this approach may not perform well in vertical domains; fine-tuning or retraining requires significant investment of data and computing power, resulting in high deployment costs.

[0024] 4. Extraction based on a reranker model: This method typically trains a model with long text understanding capabilities, similar to a large language model. Candidate texts are presented as an array, and the model ranks the candidate texts based on their semantic similarity to the target text. However, this method still requires optimization and suffers from issues such as the need for data and computing power investment required for fine-tuning or retraining natural language processing methods, resulting in high deployment costs.

[0025] 5. Large-scale model-based extraction: This method guides a large model to understand the target text from a specific perspective using prompts. It evaluates candidate texts based on the target text's semantics, expression, and even thought process, selecting the closest match. However, this method requires significant computational power, has a slow response time, and requires fine-tuning or retraining for vertical domain adaptation, resulting in high costs and deployment difficulties.

[0026] Therefore, in a specific embodiment of the present invention, such as... Figure 1 As shown, a method for retrieving mineral resource intelligence information based on text vector serialization is provided, including: S110, obtain template article data and keyword set; The keyword set refers to the core terms used to pinpoint the search target (e.g., mineral resources, rare earth, reserves, mineralization patterns, etc.). The template article data refers to standard template articles that meet the user's analytical expectations in terms of semantic content and text structure, serving as a reference benchmark for subsequent search comparisons.

[0027] Specifically, the keyword set is obtained by extracting keywords from a specific domain. Taking the mineral resources domain as an example, the set M represents the keywords in the mineral resources domain, and the set W contains the keywords. These two sets satisfy the following relationship: Where iw is the number of keywords.

[0028] Furthermore, the sample article data can be viewed as another set, denoted by A, where, .

[0029] The sample article should contain keywords. The relationship between the mineral resources field, keywords, and sample articles is as follows: .

[0030] S120, the template article data is divided into multiple text blocks, and each text block and the keywords in the keyword set are vectorized to obtain text block vectors and keyword vectors. Specifically, due to the length of the article and its multiple semantic layers, the sample article is divided into several text blocks with relatively complete semantics. Semantic segmentation can be performed using methods from the field of natural language processing, or it can be done using punctuation marks. This approach can capture deep semantic relationships within the text, overcome the limitations of literal keyword matching, and effectively handle complex linguistic phenomena such as polysemy and homonymy, thereby improving generalization and semantic accuracy.

[0031] Furthermore, vectorization can be performed using the Embedding model in natural language processing to transform text blocks and keywords into high-dimensional vectors. A text block can also be viewed as a set, denoted by the symbol Ab.

[0032] Let index set Indicates the positional order of text blocks Assuming template article data Depend on It consists of several text blocks, denoted as ,in It is the first A text block.

[0033] The vectorization process can be represented as: Therefore, after the text blocks of the template article are vectorized, the set of text blocks is represented as AbV.

[0034] .

[0035] The vectorized keywords are expressed as a set of keyword vectors, KWV.

[0036] S130, obtain the similarity between any text block vector and each keyword vector, and based on each similarity, obtain the similarity feature information of any text block vector; Specifically, select any text block vector and calculate the similarity between that selected text block vector and all keyword vectors. In other words, for any given text block vector, calculate its similarity to each keyword. Cosine similarity can be used for calculation. In scenarios with well-defined rules, this approach achieves high-accuracy text recognition and extraction, ensuring rapid identification of target information within structured or semi-structured content, and providing a stable and reliable foundation for the overall retrieval process.

[0037] Furthermore, after determining the similarity between any text block vector and each keyword vector, the similarities can be summed for any text block vector. Then, the sum of similarities for each text block vector is obtained, and multiple similarity sums are normalized to obtain the similarity feature information for any text block vector. The similarity feature information is a decimal. Normalization is used to reduce the influence of the number of keywords. In other words, for any text block vector, the cosine similarity between it and each keyword vector is calculated, and these similarities are summed. The formula for cosine similarity is as follows: ; The summed values ​​are then normalized using the L2 norm. After the above calculations, each text block obtains a scalar value, representing the similarity feature information (Am). The formula is shown below: Where, idxa∈IdA; S140, Set the number of clusters, and cluster the text block vectors according to the number of clusters to obtain multiple clusters; Specifically, an article typically consists of multiple core points (such as background, current situation, and strategies). Based on the number of points, an integer greater than or equal to 3 can be set as the number of clusters (K value). Using text block vectors as data points, the K-means clustering algorithm is employed, minimizing the sum of squared errors within each cluster using Euclidean distance. After clustering, semantically similar text blocks are grouped into the same cluster. Each cluster represents a "core point category" within the article, thus achieving a digital deconstruction of the article's deep logical components.

[0038] Furthermore, the clustering process is as follows: 1) The calculation of minimizing the sum of squared errors within a cluster is shown in the following formula: ;in, Where K is the number of clusters; It is the point set of the ic-th cluster; It is a subset of AbV, that is, the vectorized text block; It is the centroid of the ic-th cluster. 2) Assign each text block data point to the nearest center. 3) Update the center of each cluster ;in, ; 4) After feature enhancement is completed, the key point set (i.e., the clusters mentioned above) is obtained. In the above formula, That is, the text block vectors assigned to different clusters.

[0039] S150, extract the set of position labels for each text block in each cluster, and assign label coefficients to each set of position labels; Specifically, while clustering identifies key categories, it disrupts the original text order. To restore the article's structure, the original sequence numbers of text blocks within each cluster are extracted from the original article, forming a set of positional labels. The positional label set Im for each cluster is then extracted and labeled, completing the construction of the structural sequence.

[0040] Then, for each set of labels, a label coefficient is assigned as a second type of label. The introduction of the label coefficient reduces the dimensionality of the originally multidimensional semantic categories and transforms them into pure numerical identifiers, greatly reducing the computational cost of subsequent calculations. The coefficient formula is as follows: S160 generates a one-dimensional structure sequence based on label coefficients, similarity feature information, and position label set; Specifically, using the set of position labels, Am is marked, becoming Am2, which is the aforementioned one-dimensional structure sequence. In other words, based on the original order recorded in the set of position labels, the label coefficients of the corresponding categories obtained in S150 are superimposed on the similarity feature information calculated in S130. In a specific example, the order of elements in Am corresponds to the order of text blocks in the article.

[0041] Where idxa∈IdA; S170, converts a one-dimensional structural sequence into a template two-dimensional image; Specifically, obtain the number of elements N in the one-dimensional structure sequence, and calculate whether it is greater than or equal to... The smallest integer is used as the matrix side length. The one-dimensional structure sequence is divided into equal-length segments and folded into square matrices in row-major order (if the length is insufficient, a preset value is added at the end to fill it). Based on the values ​​in the matrix, it is mapped to the image grayscale value domain, thus converting the template article into a template two-dimensional image. It should be noted that the two-dimensional image can also be rectangular. It is divided into rectangles according to the number of elements (if the length is insufficient, a preset value is added at the end to fill it), forming a length and width segment, and folded into a rectangular matrix in row-major order. The preset value can be the mean of the last N elements in the one-dimensional structure sequence. Up to this step, by obtaining the similarity between any text block and the keyword, and further obtaining the similarity feature information of any text block vector, the text blocks of different clusters are assigned label coefficients, so that the one-dimensional structure sequence simultaneously reflects the text logic and the matching logic of the keywords, transforming the natural language logic into an image matrix, avoiding the deployment of a large language model and greatly reducing the computing power cost. Intermediate images and image vectors are generated as a visualization of the reasoning process. This not only improves operational efficiency but also provides intuitive and interpretable intermediate layer support for process control and result traceability, enhancing the transparency and controllability of the overall methodology.

[0042] S180: Obtain candidate two-dimensional images of multiple candidate articles, and retrieve the target article from each candidate article based on the similarity between the template two-dimensional image and the candidate two-dimensional image.

[0043] Specifically, candidate 2D images can be generated using the same processing steps as S110-S170. Image similarity algorithms (such as calculating the cosine similarity of the image feature matrices) are then used to compare the two. Candidate articles with extremely high similarity not only indicate that their keyword semantics are similar to the template, but also demonstrate a high degree of consistency in their writing style and content structure. Based on the similarity ranking, target articles that meet the needs of intelligence analysis are finally selected.

[0044] Furthermore, the step of obtaining candidate two-dimensional images of multiple candidate articles can be achieved through the following steps: Using steps S110-S160, a one-dimensional structure sequence corresponding to the candidate articles is generated. This involves dividing the candidate article data into multiple candidate text blocks, and vectorizing each candidate text block and the keywords in the keyword set to obtain candidate text block vectors and keyword vectors. The similarity between any candidate text block vector and each keyword vector is obtained, and based on these similarities, candidate similarity feature information for any candidate text block vector is obtained. The number of clusters is set, and based on the number of clusters, the candidate text block vectors are clustered to obtain multiple candidate clusters. The candidate position label set for each candidate text block in each candidate cluster is extracted, and candidate label coefficients are assigned to each candidate position label set. Based on the candidate label coefficients, candidate similarity feature information, and candidate position label sets, a candidate one-dimensional structure sequence is generated. Then, based on the resolution of the template 2D image, preset values ​​are filled into the candidate 1D structure sequence to obtain the target 1D structure sequence. The preset values ​​can be the average of the last N elements in the candidate 1D structure sequence. This preset value is filled cyclically to ensure that the number of elements in the candidate 1D structure sequence matches the resolution of the template 2D image. That is, when the resolution of the range 2D image is 10x10, the number of elements in the candidate 1D structure sequence is 100.

[0045] Finally, the target one-dimensional structure sequence is converted into a candidate two-dimensional image. The conversion method can refer to the method described above for converting a one-dimensional structure sequence into a template two-dimensional image. This method avoids disrupting the original structural logic of the article due to scaling during the two-dimensional image stage. For example, in a 4x4 matrix, sentence 4 is at the end of the first row and sentence 5 is at the beginning of the second row. In a two-dimensional image, sentences 1 and 5 become adjacent. Subsequent image scaling using bilinear interpolation or region averaging would forcibly mix the features of sentences 1 and 5. The above method effectively avoids this problem, while using preset values ​​makes the image transition smoother and avoids computational distortion.

[0046] The aforementioned information retrieval method quantifies text using a general vector model to obtain the similarity between any text block and keywords. Then, it projects the keyword similarity onto the text, constructing a preliminary relationship between keywords and text, forming a unified whole and improving matching accuracy. Based on this, clustering extracts structural elements from the text, and then serializes the clusters to digitally represent the text structure. This structural unit is then visualized to capture the spatial relationships between structural elements. Ultimately, without requiring massive amounts of labeled training data or pre-training, and avoiding reliance on huge computational resources, this method achieves control over the macro-level textual structure and micro-level semantic information of articles, selecting the most relevant intelligence documents, improving the accuracy of mineral resource intelligence retrieval, and enabling faster understanding of the latest mineral resource dynamics and the formulation of response strategies.

[0047] In one embodiment, such as Figure 2 As shown, the steps for segmenting the template article data into multiple text blocks include: S210, Extract punctuation marks from template article data; S220, based on punctuation marks, divides the template article data into multiple text blocks.

[0048] Specifically, punctuation marks, including natural pauses such as commas, periods, and semicolons, are used as segmentation boundaries. This approach ensures that each segmented text block contains relatively independent and complete semantic information.

[0049] In one embodiment, such as Figure 3 As shown, the steps for obtaining similarity feature information for any text block vector based on various similarity scores include: S310, sum the similarities of each similarity to obtain the sum of similarities for any text block vector; S320: Construct a summation vector based on the similarity sum of any text block vector; S330 uses L2 norm normalization to process the summation vector, obtaining similarity feature information for any text block vector.

[0050] Specifically, when calculating the relevance of a text block to multiple keywords, simply using summation can lead to an infinitely large final score when the number of target keywords is large. This embodiment first calculates the sum of similarities of a single text block to all keywords (i.e., the aforementioned similarity sum); then, it combines the similarity sums of all text blocks in the article to construct a global summation vector; finally, it uniformly performs L2 norm normalization on this summation vector, so that the similarity feature information of all text blocks is compressed and mapped to the interval [0, 1] (i.e., a decimal), which preserves the relative strength of the relevance between different short sentences and prevents numerical overflow, thus preparing for subsequent feature overlay.

[0051] In one embodiment, such as Figure 4 As shown, the steps for generating a one-dimensional structural sequence based on label coefficients, similarity feature information, and a set of position labels include: S410, based on the set of location labels, processes similar feature information to obtain a similarity set; S420, based on the set of position labels, superimposes the coefficients of each label onto the similarity set to obtain a one-dimensional structure sequence.

[0052] The coefficients are integers greater than or equal to 0.

[0053] Specifically, based on the set of positional labels (i.e., the order of the original text), the normalized similarity feature information (in decimal form) of each text block is arranged in order to form a similarity set. Since the pre-set label coefficients are limited to integers greater than or equal to 0 (e.g., 0, 1, 2), when the system superimposes the label coefficients of integers greater than or equal to 0 onto the decimal similarity set, each value in the generated one-dimensional structural sequence will have two independent layers of information: the integer part represents the textual logic and structural cluster to which the sentence belongs, and the decimal part represents the semantic similarity between the sentence and the target keyword.

[0054] In one embodiment, such as Figure 5 As shown, the steps for converting a one-dimensional structure sequence into a template two-dimensional image include: S510 converts a one-dimensional structure sequence into a two-dimensional sequence; Specifically, preset values ​​are added to the one-dimensional structure sequence to obtain the current structure sequence, so that the number of elements in the one-dimensional structure sequence can be square rooted; the current structure sequence is divided into k equal-length subsequences; where k is a positive integer; based on the division order, each equal-length subsequence is transformed into a two-dimensional matrix.

[0055] S520 converts the numerical values ​​in the two-dimensional sequence into grayscale values ​​and generates a template two-dimensional image based on the grayscale values.

[0056] Specifically, the floating-point values ​​within the matrix are globally mapped and converted into a standard image grayscale range (e.g., 0-255). The mapped two-dimensional sequence is visually presented as an image composed of blocks of different grayscale values. The distribution of light and dark and the texture features of this image intuitively and comprehensively express the writing structure and semantic focus of the sample article.

[0057] In one embodiment, the step of obtaining candidate two-dimensional images of multiple candidate articles includes: When the resolution of the candidate 2D image is greater than that of the template 2D image, primitive downsampling processing is performed on the candidate 2D image to obtain the downsampled image; When the resolution of the candidate 2D image is less than that of the template 2D image, primitive upsampling processing is performed on the candidate 2D image to obtain the upsampled image. The steps for retrieving the target article from the candidate articles include: Based on the similarity between the downsampled image and the template 2D image, the target article is retrieved from each candidate article, or; The target article is retrieved from each candidate article based on the similarity between the upsampled image and the template 2D image.

[0058] Specifically, due to the varying lengths of candidate articles, the resolution (size) of the generated candidate 2D images does not match that of the template articles, making direct image comparison impossible. Therefore, this embodiment introduces an adaptive scale alignment mechanism. When the image generated from a candidate long article is too large, primitive downsampling (such as the region mean method) is used to reduce its size, compressing redundant information without losing the overall textual structure. When the image generated from a candidate short article is too small, primitive upsampling (such as bilinear interpolation) is used for smooth enlargement.

[0059] After unifying the resolution, the feature vector similarity between the downsampled / upsampled image and the template 2D image is calculated. The candidate article with the highest similarity is the target article.

[0060] To further illustrate the technical solution of this application, a specific scenario is provided below: 1. Obtain a keyword set. The keywords are selected as follows: Mineral resources, light rare earth, medium and heavy rare earth, rare earth industry chain, global reserves distribution, geopolitical pattern 2. Obtaining template article data: Rare earth elements, as a strategic mineral comprising 17 elements, exhibit a highly concentrated global reserve distribution with significant geographical attributes. As of 2025, China ranked first with 44 million tons of reserves, accounting for 40%-48.9% of the global total. Light rare earth elements are concentrated in Bayan Obo, Inner Mongolia, while medium and heavy rare earth elements are concentrated in Ganzhou, Jiangxi. Mongolia ranked second with 31-32 million tons, while Vietnam and Brazil, each with reserves exceeding 20 million tons, ranked third and fourth respectively. Russia has 10-20 million tons of reserves concentrated in the Siberian permafrost zone. Under this distribution pattern, China not only leads in reserves but also controls a complete mining and processing industrial chain. While the United States has 1.4-1.9 million tons of reserves, it lacks core separation technology, requiring 80% of its raw materials to be shipped to China for processing. Russia's separation technology remains stuck in the Soviet era, with annual production capacity only meeting 30% of domestic demand. In recent years, many countries have attempted to break their dependence through the exploration of new mines or international cooperation, but technological bottlenecks and high mining costs make it difficult to shake China's dominant position in the global rare earth geopolitical landscape in the short term.

[0061] 3. The similarity feature information after vectorization, similarity calculation, and L2 norm normalization is shown in the table below: 4. The clustering results are shown in the table below: 5. The generated one-dimensional structure sequence is shown in the table below: 6. Text blocks with equal-length marker values ​​are stacked, with padding at the ends to form squares. 7. The stacked text block marker values ​​are image-processed and converted into squares with corresponding grayscale values. The image-based text retrieval unit is as follows: Figure 6 As shown.

[0062] This embodiment focuses on rare earth elements and rare earth permanent magnets, preparing six texts with similar structures, word counts, and meanings; additionally, it prepared 12 texts with similar meanings but different structures, and 12 texts with different meanings, themes, and structures, and conducted a comparison test on the mean similarity. The table below is the mean image similarity table.

[0063] The test results show that vectorized, similarity-based, feature-enhanced, and image-based vectors can distinguish differences in the text's logical flow. Therefore, the method presented in this paper can be used to filter texts with similar meanings and structures.

[0064] In one embodiment, a mineral resource intelligence information retrieval device based on text vector serialization is also provided, comprising: The acquisition module is used to retrieve template article data and keyword sets; The vectorization module is used to divide the template article data into multiple text blocks, and to vectorize each text block and the keywords in the keyword set to obtain text block vectors and keyword vectors. The similarity calculation module is used to obtain the similarity between any text block vector and each keyword vector, and based on each similarity, to obtain the similarity feature information of any text block vector; The clustering module is used to set the number of clusters and to cluster the text block vectors according to the number of clusters to obtain multiple clusters; The label coefficient allocation module is used to extract the set of position labels for each text block in each cluster and assign label coefficients to each set of position labels. The sequence generation module is used to generate one-dimensional structure sequences based on label coefficients, similarity feature information, and position label sets; The conversion module is used to convert one-dimensional structural sequences into template two-dimensional images; The retrieval module is used to obtain candidate two-dimensional images of multiple candidate articles, and to retrieve the target article from each candidate article based on the similarity between the template two-dimensional image and the candidate two-dimensional image.

[0065] Specific limitations regarding the mineral resource information retrieval device based on text vector serialization can be found in the limitations of the mineral resource information retrieval method based on text vector serialization mentioned above, and will not be repeated here. Each module in the aforementioned mineral resource information retrieval device based on text vector serialization can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. It should be noted that the module division in this embodiment is illustrative and only represents a logical functional division; other division methods may be used in actual implementation.

[0066] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: Obtain sample article data and keyword set; The sample article data is divided into multiple text blocks, and each text block and the keywords in the keyword set are vectorized to obtain text block vectors and keyword vectors. Obtain the similarity between any text block vector and each keyword vector, and based on each similarity, obtain the similarity feature information of any text block vector; Set the number of clusters, and cluster the text block vectors according to the number of clusters to obtain multiple clusters; Extract the set of position labels for each text block in each cluster, and assign label coefficients to each set of position labels; A one-dimensional structural sequence is generated based on the label coefficients, similarity feature information, and location label set. Convert a one-dimensional structural sequence into a template two-dimensional image; The system obtains candidate two-dimensional images of multiple candidate articles, and retrieves the target article from each candidate article based on the similarity between the template two-dimensional image and the candidate two-dimensional image.

[0067] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, performs the following steps: Obtain sample article data and keyword set; The sample article data is divided into multiple text blocks, and each text block and the keywords in the keyword set are vectorized to obtain text block vectors and keyword vectors. Obtain the similarity between any text block vector and each keyword vector, and based on each similarity, obtain the similarity feature information of any text block vector; Set the number of clusters, and cluster the text block vectors according to the number of clusters to obtain multiple clusters; Extract the set of position labels for each text block in each cluster, and assign label coefficients to each set of position labels; A one-dimensional structural sequence is generated based on the label coefficients, similarity feature information, and location label set. Convert a one-dimensional structural sequence into a template two-dimensional image; The system obtains candidate two-dimensional images of multiple candidate articles, and retrieves the target article from each candidate article based on the similarity between the template two-dimensional image and the candidate two-dimensional image.

[0068] In specific implementation, the embodiments of this application can be referred to the above embodiments and have corresponding technical effects.

[0069] It is understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described herein, or combinations thereof.

[0070] For software implementation, the techniques described herein can be implemented by units that perform the functions described herein. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or external to the processor.

[0071] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0072] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0073] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0074] If the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks. It should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0075] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for retrieving mineral resource intelligence information based on text vector serialization, characterized in that, include: Obtain template article data and keyword set; wherein, the keyword set refers to the core vocabulary used to lock in the search target; the template article data refers to standard template articles that meet the user's analysis expectations in terms of semantic content and text structure; The template article data is divided into multiple text blocks, and each text block and the keywords in the keyword set are vectorized to obtain text block vectors and keyword vectors. Obtain the similarity between any of the text block vectors and each of the keyword vectors, and based on each of the similarities, obtain the similarity feature information of any of the text block vectors; Set the number of clusters, and cluster the text block vectors according to the number of clusters to obtain multiple clusters; Extract the set of position labels for each text block in each cluster, and assign label coefficients to each set of position labels; A one-dimensional structure sequence is generated based on the label coefficients, the similarity feature information, and the location label set; wherein, the step of generating the one-dimensional structure sequence based on the label coefficients, the similarity feature information, and the location label set includes: arranging the normalized similarity feature information of each text block in order according to the location label set to form a similarity set; and superimposing the label coefficients onto the similarity set to generate the one-dimensional structure sequence; The one-dimensional structure sequence is converted into a template two-dimensional image; A candidate two-dimensional image of multiple candidate articles is obtained, and the target article is retrieved from each candidate article based on the similarity between the template two-dimensional image and the candidate two-dimensional image.

2. The mineral resource intelligence information retrieval method based on text vector serialization according to claim 1, characterized in that, The step of segmenting the template article data into multiple text blocks includes: Extract punctuation marks from the template article data; Based on the punctuation marks, the template article data is divided into multiple text blocks.

3. The mineral resource intelligence information retrieval method based on text vector serialization according to claim 1, characterized in that, The step of obtaining similarity feature information for any of the text block vectors based on the aforementioned similarities includes: The similarities of each text block vector are summed to obtain the sum of similarities. Based on the similarity sum of the text block vectors, a summation vector is constructed; The summation vector is normalized using the L2 norm to obtain similarity feature information for any of the text block vectors.

4. The mineral resource intelligence information retrieval method based on text vector serialization according to claim 1, characterized in that, Steps to set the number of clusters: Obtain the number of key points in the template article data; The number of clusters is set based on the number of key points.

5. The mineral resource intelligence information retrieval method based on text vector serialization according to claim 1, characterized in that, The step of generating a one-dimensional structure sequence based on the label coefficients, the similarity feature information, and the position label set includes: Based on the set of location labels, the similarity feature information is processed to obtain a similarity set; Based on the set of position labels, the coefficients of each label are superimposed on the similarity set to obtain the one-dimensional structure sequence.

6. The mineral resource intelligence information retrieval method based on text vector serialization according to claim 5, characterized in that, The label coefficients are integers greater than or equal to 0.

7. The mineral resource intelligence information retrieval method based on text vector serialization according to claim 1, characterized in that, The step of converting the one-dimensional structure sequence into a template two-dimensional image includes: Convert the one-dimensional structure sequence into a two-dimensional sequence; The numerical values ​​in the two-dimensional sequence are converted into grayscale values, and the template two-dimensional image is generated based on the grayscale values.

8. The mineral resource intelligence information retrieval method based on text vector serialization according to claim 7, characterized in that, The step of converting the one-dimensional structure sequence into a two-dimensional sequence includes: The preset values ​​are added to the one-dimensional structure sequence to obtain the current structure sequence, so that the number of elements in the one-dimensional structure sequence can be squared. The current structure sequence is divided into k equal-length subsequences; where k is a positive integer; Based on the segmentation order, each of the equal-length subsequences is transformed into the two-dimensional sequence.

9. The mineral resource intelligence information retrieval method based on text vector serialization according to claim 1, characterized in that, The steps for obtaining candidate two-dimensional images of multiple candidate articles include: If the resolution of the candidate two-dimensional image is greater than the resolution of the template two-dimensional image, the candidate two-dimensional image is subjected to primitive downsampling processing to obtain a downsampled image. If the resolution of the candidate two-dimensional image is less than the resolution of the template two-dimensional image, the candidate two-dimensional image is subjected to primitive upsampling processing to obtain an upsampled image. The steps for retrieving the target article from each of the candidate articles include: Based on the similarity between the downsampled image and the template 2D image, the target article is retrieved from each of the candidate articles, or; Based on the similarity between the upsampled image and the template 2D image, the target article is retrieved from each of the candidate articles.

10. A mineral resource information retrieval device based on text vector serialization, characterized in that, include: The acquisition module is used to acquire template article data and keyword sets; wherein, the keyword set refers to the core vocabulary used to lock in the search target; the template article data refers to standard template articles that meet the user's analysis expectations in terms of semantic content and text structure; The vectorization module is used to divide the template article data into multiple text blocks, and to vectorize each text block and the keywords in the keyword set to obtain text block vectors and keyword vectors. The similarity calculation module is used to obtain the similarity between any of the text block vectors and each of the keyword vectors, and based on each of the similarities, to obtain the similarity feature information of any of the text block vectors; The clustering module is used to set the number of clusters and to cluster the text block vectors according to the number of clusters to obtain multiple clusters; The label coefficient allocation module is used to extract the set of position labels of each text block in each cluster and to allocate label coefficients to each set of position labels; The sequence generation module is used to generate a one-dimensional structure sequence based on the label coefficients, the similarity feature information, and the position label set; it is also used to arrange the normalized similarity feature information of each text block in order according to the position label set to form a similarity set; and to superimpose the label coefficients onto the similarity set to generate the one-dimensional structure sequence. A conversion module is used to convert the one-dimensional structure sequence into a template two-dimensional image; The retrieval module is used to obtain candidate two-dimensional images of multiple candidate articles, and retrieve the target article from each candidate article based on the similarity between the template two-dimensional image and the candidate two-dimensional image.

Images