Table and context parsing method and system for earth science field

Abstract

The present application relates to the field of geological information mining, and provides a table and context analysis method and system for the field of earth science, comprising: obtaining a geological data table from a geological database; preprocessing the geological data table to obtain a to-be-analyzed geological data table; extracting cells from the to-be-analyzed geological data table to obtain position information of the cells, and extracting content information of the cells in the cells corresponding to the position information through an OCR algorithm; converting the position information, belonging row and column information and content information of each cell into triple form data, aligning the triple form data and triple information obtained by analyzing table context content to obtain a geological knowledge graph. The improved MaskRCNN model selected by the present application finally achieves optimal evaluation indexes; and a question and answer platform constructed based on the geological knowledge graph provides intelligent knowledge query and retrieval services for users.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geological information mining, and in particular to a table and context analysis method and system for the field of earth science. BACKGROUND

[0002] Table analysis is an important task in the field of natural language processing, which aims to convert unstructured information in the table into structured information to form geological knowledge, which can effectively obtain information from geological report files and improve the cognitive intelligent analysis level in the geological investigation process. Current table analysis methods are mainly divided into two categories: traditional heuristic methods and deep learning-based methods.

[0003] The traditional heuristic method extracts tables from text by character-specific distance, table-specific title, line features, and table-specific alignment format. The deep learning-based method treats the table as an image and uses image segmentation and target extraction to analyze the table using a deep learning model.

[0004] At present, the research on table analysis only provides partial solutions. Many existing table analysis methods are for general-purpose tables and financial tables, and are not suitable for table analysis scenarios in specific professional fields. First, in cell extraction, the number of merged cells in the surface is large, resulting in large differences in cell size, and a large number of small target cells cannot be identified and extracted. Second, in table structure analysis, geological tables contain a large number of diagonal split table headers and merged cells, making structure analysis difficult. Finally, the lack of integration of table analysis with context causes knowledge to be single and limited. Therefore, the geological big data composed of a large amount of unstructured data such as geological investigation reports and work records has not been fully utilized and mined.

[0005] The above content is only used to assist in understanding the technical solutions of the present application and does not represent an acknowledgement of the above content as prior art. SUMMARY

[0006] To solve the above technical problems, the present application provides a table and context analysis method for the field of earth science, comprising:

[0007] S1: obtaining a geological data table from a geological database;

[0008] S2: preprocessing the geological data table to obtain a to-be-analyzed geological data table;

[0009] S3: extracting cells from the to-be-analyzed geological data table to obtain position information of the cells, and extracting content information of the cells in the cells corresponding to the position information by an OCR algorithm;

[0010] S4: Obtain the row and column information of a cell using table parsing methods based on the cell's position information;

[0011] S5: Convert the position information, row and column information, and content information of each cell into triplet data, align the triplet data with the triplet information obtained from parsing the table context, and obtain the geological knowledge graph.

[0012] Preferably, step S2 specifically includes:

[0013] Labels are used to annotate geological data tables, including both table labels and cell labels.

[0014] Preferably, step S3 specifically includes:

[0015] S31: The geological data tables to be parsed are divided into ordinary tables and tables with borders. The Mask R-CNN model is made to converge by using an improved loss function to obtain an improved Mask R-CNN model. The position information of the cells in the ordinary tables is extracted by the improved Mask R-CNN model.

[0016] S32: Use the OpenCV framework to parse the position of the dividing lines in a table with borders, and obtain the position information of the cells in the table with borders.

[0017] S33: Locate cells using location information and extract content information from cells in regular tables and cells in tables with borders using an OCR algorithm.

[0018] The preferred, improved expression for the loss function is:

[0019]

[0020]

[0021]

[0022]

[0023] L align =L a +L b +L c +L d

[0024] Loss = L class +L bbox +L mask +L align

[0025] Among them, L class L represents the classification loss.bbox represents the regression box loss, L mask represents the mask loss, SR represents the starting row, SC represents the starting column, ER represents the ending row, EC represents the ending column, C i represents two cells in a specific row i, C j represents two cells in a specific row j, (x1 Ci , y1 Ci ) and (x2 Ci , y2 Ci ) represent the bounding box coordinates of the cells, i.e. the top-left corner coordinates and the bottom-right corner coordinates.

[0026] Preferably, step S4 is specifically:

[0027] S41: In the cell containing the division line, convert its image to a binary image through the Opencv model, identify the line segment endpoint coordinates of the binary image, and regard the remaining cells without the division line as a rectangle with length and width;

[0028] S42: Set the shortest third of the long side of the rectangle of each cell as the horizontal coordinate tolerance threshold k x , and set the shortest third of the wide side of the rectangle of each cell as the vertical coordinate tolerance threshold k y , the expression is:

[0029]

[0030]

[0031] wherein (x1, y1) is the top-left corner coordinate of the cell, and (x2, y2) is the bottom-right corner coordinate of the cell;

[0032] S43: Obtain the corrected cell coordinates through the horizontal coordinate tolerance threshold, the vertical coordinate tolerance threshold, and the row-column vertical relationship of the table, and obtain the row information and column information of each cell through the corrected cell coordinates;

[0033] S44: In the cell containing the division line, locate the endpoint coordinates of the division line segment through the Opencv model, and when the cell (a x , b y ) belongs to the mth row and the nth column, the coordinate relationship satisfies the following formula:

[0034]

[0035]

[0036] wherein a m represents the minimum horizontal coordinate value of the mth row cell, b n represents the minimum vertical coordinate value of the nth column cell, am+1 represents the maximum horizontal coordinate value of the mth row cell, b n+1 represents the maximum vertical coordinate value of the nth column cell, start_row and end_row respectively represent the start and end of the row information of the cell, start_col and end_col respectively represent the start and end of the column information of the cell,

[0037] S45: The number, top-left corner coordinate and bottom-right corner coordinate of each cell are taken as the position information of each cell, and the row information and column information are taken as the row and column information to which the cell belongs.

[0038] Preferably, step S5 is specifically:

[0039] S51: The position information, the belonging row and column information and the content information of each cell are converted into triple form data by an ALBERT model, and the triple form data is represented as (a, r, b), wherein a and b represent entities, and r represents the relationship between entities a and b;

[0040] S52: The triple form data is aligned with the parsed geological data table, including entity alignment and embedding fusion of the table-based knowledge graph and the context-based knowledge graph, to obtain a geological knowledge graph.

[0041] A table and context analysis system for the field of earth science, comprising:

[0042] A table acquisition module, configured to acquire a geological data table from a geological database;

[0043] A preprocessing module, configured to preprocess the geological data table to obtain a to-be-analyzed geological data table;

[0044] An extraction module, configured to extract cells from the to-be-analyzed geological data table to obtain position information of the cells, and extract content information of the cells in the cells corresponding to the position information by an OCR algorithm;

[0045] A row and column information acquisition module, configured to obtain belonging row and column information of the cells by using a table analysis method based on the position information of the cells;

[0046] A graph construction module, configured to convert the position information, the belonging row and column information and the content information of each cell into triple form data, and align the triple form data with triple information obtained from context content of the parsed table to obtain a geological knowledge graph.

[0047] The present application has the following beneficial effects:

[0048] 1. The Mask RCNN model is improved, and compared with other baseline models, the improved Mask RCNN model selected by the application finally reaches the optimal evaluation index;

[0049] 2. The question and answer platform constructed by the geological knowledge graph provides a simpler channel for users to obtain geological information, which is beneficial for geologists to quickly obtain information and for non-geologists to easily understand geological information. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The method flowchart of the embodiment of the application is shown in the figure;

[0051] Figure 2 The structure diagram of the Albert-BiGRU-Attenetion model is shown in the figure;

[0052] Figure 3 The interface diagram of the geological knowledge graph is shown in the figure;

[0053] The implementation of the object, functional characteristics and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0054] It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0055] Referring to Figure 1 , the application provides a table and context analysis method for the field of earth science, comprising:

[0056] S1: obtaining a geological data table from a geological database;

[0057] Specifically, the geological database used includes basic geology, mineral geology, hydrogeology, engineering geology, environmental geology, etc., and the application collects tables from these geological databases;

[0058] S2: preprocessing the geological data table to obtain a to-be-analyzed geological data table;

[0059] S3: extracting a cell from the to-be-analyzed geological data table to obtain position information of the cell, and extracting content information of the cell in the cell corresponding to each position information by using an OCR algorithm;

[0060] S4: obtaining row and column information of the cell by using a table analysis method through the position information of the cell;

[0061] S5: Convert the position information, row and column information, and content information of each cell into triplet data, align the triplet data with the triplet information obtained from parsing the table context, and obtain the geological knowledge graph.

[0062] In this embodiment, step S2 specifically includes:

[0063] Labels are used to annotate geological data tables, including both table labels and cell labels.

[0064] In this embodiment, step S3 specifically includes:

[0065] S31: The geological data tables to be parsed are divided into ordinary tables and tables with borders. The Mask R-CNN model is made to converge by using an improved loss function to obtain an improved Mask R-CNN model. The position information of the cells in the ordinary tables is extracted by the improved Mask R-CNN model.

[0066] S32: Use the OpenCV framework to parse the position of the dividing lines in a table with borders, and obtain the position information of the cells in the table with borders.

[0067] S33: Locate cells using location information and extract content information from cells in regular tables and cells in tables with borders using an OCR algorithm.

[0068] In this embodiment, to address the issue of large error in the loss function of the original Mask RCNN model, the present invention redefines an improved loss function.

[0069] The improved loss function is expressed as follows:

[0070]

[0071]

[0072]

[0073]

[0074] L align =L a +L b +L c +L d

[0075] Loss = L class +L bbox +L mask +L align

[0076] Among them, Lclass L represents the classification loss. bbox L represents the regression box loss. mask Indicates mask loss, SR represents the start row, SC represents the start column, ER represents the end row, EC represents the end column, and C represents the end column. i C represents two cells in a specific row i. j This refers to two cells in a specific row j, (x1) Ci y1 Ci ) and (x2 Ci y2 Ci The numbers represent the cell's border coordinates, namely the top-left corner and bottom-right corner.

[0077] Specifically, compared with other baseline models, the table parsing model we designed achieved the best results in all evaluation metrics. The experimental results are shown in Table 1 below:

[0078] Table 1 Evaluation results compared with other table parsing models

[0079]

[0080] In this embodiment, step S4 specifically includes:

[0081] S41: In the cell containing the dividing line, convert its image into a binary image using the OpenCV model, identify the coordinates of the endpoints of the line segments in the binary image, and treat the remaining cells without dividing lines as a rectangle with length and width.

[0082] S42: Set the shortest third of the longest side of the rectangle in each cell as the horizontal coordinate tolerance threshold k. x Set the shortest third of the rectangle in each cell as the fault tolerance threshold k on the y-axis. y The expression is:

[0083]

[0084]

[0085] Where (x1, y1) is the coordinate of the top left corner of the cell, and (x2, y2) is the coordinate of the bottom right corner of the cell;

[0086] S43: Obtain the corrected cell coordinates by using the horizontal axis fault tolerance threshold, the vertical axis fault tolerance threshold, and the vertical relationship between rows and columns of the table; obtain the row and column information of each cell by using the corrected cell coordinates.

[0087] S44: In a cell containing a dividing line, locate the endpoint coordinates of the dividing line segment using the OpenCV model. When cell (a x b yWhen a given element is in the m-th row and n-th column, the coordinate relationship satisfies the following formula:

[0088]

[0089]

[0090] Among them, a m b represents the minimum x-coordinate value of the cell in row m. n a represents the minimum y-coordinate value of the cell in the nth column. m+1 b represents the maximum x-coordinate value of the cell in row m. n+1 This represents the maximum y-coordinate value of the cell in column n. `start_row` and `end_row` represent the start and end points of the row information for the cell, respectively. `start_col` and `end_col` represent the start and end points of the column information for the cell, respectively.

[0091] S45: Use the cell number, top-left corner coordinates, and bottom-right corner coordinates of each cell as the cell's position information, and the row and column information as the cell's row and column information.

[0092] In this embodiment, step S5 specifically includes:

[0093] S51: The ALBERT model is used to convert the position information, row and column information and content information of each cell into triplet data. The triplet data is represented as (a,r,b), where a and b represent entities and r represents the relationship between entities a and b.

[0094] Specifically:

[0095] (1) The text paragraphs in the geological data tables were structured using the Albert-BiGRU-Attenetion model structure; the serialized text was used as the input to the proposed model and the output was the corresponding annotation sequence. The output sequence system was the "BIO" annotation system.

[0096] The structure of the Albert-BiGRU-Attenetion model is as follows: Figure 2 As shown, based on practical experience, some important parameters involved in the ALBERT-BiGRU-Attention relation extraction model were determined through debugging. While ensuring the extraction effect, the training efficiency of the model was further improved and the overall complexity of the model was reduced, achieving better results compared with other models. The experimental results are shown in Table 2 below:

[0097] Table 2 Evaluation results compared with other text information extraction models

[0098] Model P R F1 T / s ALBERT-BiGRU 0.701 0.712 0.71 607.33 Word2vec-BiGRU-Attention 0.723 0.729 0.73 6.55 ELMo-BiGRU-Attention 0.756 0.766 0.76 8907.1 BERT-BiGRU-Attention 0.801 0.815 0.81 3775.6 Ours 0.811 0.833 0.82 611.8

[0099] (2) Each character input is converted into a vector form by the ALBERT model and used as the input of BiGRU to extract contextual features;

[0100] (3) Use the output feature vector as the input of the CRF layer. The CRF layer normalizes the input, learns the constraint relationship between labels, and outputs the predicted label sequence.

[0101] (4) Structured information is generated through the information extraction model and output in the form of triples;

[0102] S52: Align the triplet-form data with the parsed geological data table, including entity alignment and embedding and fusion of table-based knowledge graph and context-based knowledge graph to obtain a geological knowledge graph;

[0103] Specifically:

[0104] (1) First, the string is converted into a vector using the TF-IDF algorithm (Cohen W et al., 2002). The similarity between two vectors can be measured by the angle between them. The cosine similarity of entity vectors a and b is calculated using the following formula:

[0105]

[0106] By setting a similarity threshold k=0.95, entities that cannot be aligned are filtered out. Entities with k>0.95 are considered to have the same meaning and are renamed to the same name.

[0107] (2) Embed the table-based knowledge graph G2 as the target knowledge graph G2 into the text-based knowledge graph G1, where Q represents the embedding of all entities and relations in the two knowledge graphs. Define the objective function of this process as:

[0108] P(G1,G2|Q)=P(G1|Q)P(G2|G1,Q)

[0109] First, the embedding formula for entities and relations in the source knowledge graph G1 is:

[0110]

[0111] Where (a,r,b) are triples in the source knowledge graph G1;

[0112] (3) Use the TransE model to embed entities and relations in the source knowledge graph; the TransE model-based method performs separate representation learning on the knowledge graph to obtain the corresponding entity and relation vectors, and projects them into a unified vector space by judging entity alignment.

[0113] (4) Then, by integrating the information from the source knowledge graph, entities and relations in the target knowledge graph G2 are embedded. Some entities in G2 are linked to the source knowledge graph G1, denoted as M = {(k,l)|k}, where k is the entity corresponding to G1 and l is the entity corresponding to G2, as follows:

[0114]

[0115] Where (a, r, b) is a triple in the target knowledge graph G2, a and b represent entities, and r represents the relationship between entities a and b. This achieves entity alignment and the embedding and fusion of table-based and context-based knowledge graphs to obtain a geological knowledge graph. The interface of the geological knowledge graph is as follows: Figure 3 As shown;

[0116] The question-and-answer platform built upon the final geological knowledge map provides users with a simpler way to access geological information, enabling geological researchers to quickly obtain information and non-geological researchers to easily understand it.

[0117] This invention provides a table and context parsing system for the field of earth sciences, comprising:

[0118] The table retrieval module is used to retrieve geological data tables from a geological database.

[0119] The preprocessing module is used to preprocess the geological data table to obtain the geological data table to be parsed;

[0120] The extraction module is used to extract cells from the geological data table to be parsed, obtain the cell location information, and extract the cell content information from the cells corresponding to each location information using an OCR algorithm.

[0121] The row and column information acquisition module is used to obtain the row and column information of a cell by using the cell's position information and table parsing methods;

[0122] The knowledge graph construction module is used to convert the location information, row and column information, and content information of each cell into triplet data, and align the triplet data with the triplet information obtained by parsing the table context to obtain a geological knowledge graph.

[0123] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system 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 system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0124] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0125] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A table and context parsing method for the field of Earth sciences, characterized in that, include: S1: Obtain geological data tables from the geological database; S2: Preprocess the geological data table to obtain the geological data table to be parsed; S3: Extract cells from the geological data table to be analyzed, obtain cell location information, and extract cell content information from the cells corresponding to each location information using an OCR algorithm; S4: Obtain the row and column information of a cell using its position information and a table parsing method; specifically: S41: In the cell containing the dividing line, convert its image into a binary image using the OpenCV model, identify the coordinates of the endpoints of the line segments in the binary image, and treat the remaining cells without dividing lines as a rectangle with length and width. S42: Set the shortest third of the longest side of the rectangle in each cell as the horizontal coordinate tolerance threshold k. x Set the shortest third of the rectangle in each cell as the fault tolerance threshold k on the y-axis. y The expression is: Where (x1, y1) is the coordinate of the top left corner of the cell, and (x2, y2) is the coordinate of the bottom right corner of the cell; S43: Obtain the corrected cell coordinates by using the horizontal axis fault tolerance threshold, the vertical axis fault tolerance threshold, and the vertical relationship between rows and columns of the table; obtain the row and column information of each cell by using the corrected cell coordinates. S44: In a cell containing a dividing line, locate the endpoint coordinates of the dividing line segment using the OpenCV model. When cell (a x b y When a given element is in the m-th row and n-th column, the coordinate relationship satisfies the following formula: Among them, a m b represents the minimum x-coordinate value of the cell in row m. n a represents the minimum y-coordinate value of the cell in the nth column. m+1 b represents the maximum x-coordinate value of the cell in row m. n+1 This represents the maximum y-coordinate value of the cell in column n. `start_row` and `end_row` represent the start and end points of the row information for the cell, respectively. `start_col` and `end_col` represent the start and end points of the column information for the cell, respectively. S45: Use the cell number, top-left corner coordinates, and bottom-right corner coordinates of each cell as the cell's position information, and the row and column information as the cell's row and column information; S5: Convert the position information, row and column information, and content information of each cell into triplet data, align the triplet data with the triplet information obtained from parsing the table context, and obtain the geological knowledge map; specifically: S51: The ALBERT model is used to convert the position information, row and column information and content information of each cell into triplet data. The triplet data is represented as (a,r,b), where a and b represent entities and r represents the relationship between entities a and b. S52: Align the triplet-form data with the parsed geological data table, including entity alignment and embedding and fusion of table-based knowledge graph and context-based knowledge graph to obtain a geological knowledge graph.

2. The table and context parsing method for the Earth science field according to claim 1, characterized in that, Step S2 is as follows: Labels are used to annotate geological data tables, including both table labels and cell labels.

3. The table and context parsing method for the Earth science field according to claim 1, characterized in that, Step S3 is as follows: S31: The geological data tables to be parsed are divided into ordinary tables and tables with borders. The Mask R-CNN model is made to converge by using an improved loss function to obtain an improved Mask R-CNN model. The position information of the cells in the ordinary tables is extracted by the improved Mask R-CNN model. S32: Use the OpenCV framework to parse the position of the dividing lines in a table with borders, and obtain the position information of the cells in the table with borders. S33: Locate cells using location information and extract content information from cells in regular tables and cells in tables with borders using an OCR algorithm.

4. The table and context parsing method for the Earth science field according to claim 3, characterized in that, The improved loss function is expressed as follows: L align L a +L b +L c +L d Loss=L class +L bbox +L mask +L align Among them, L class L represents the classification loss. bbox L represents the regression box loss. mask Indicates mask loss, SR represents the start row, SC represents the start column, ER represents the end row, EC represents the end column, and C represents the end column. i C represents two cells in a specific row i. j This refers to two cells in a specific row j, (x1) Ci y1 Ci ) and (x2 Ci y2 Ci The numbers represent the cell's border coordinates, namely the top-left corner and bottom-right corner.

5. A table and context parsing system for the field of Earth sciences, characterized in that, include: The table retrieval module is used to retrieve geological data tables from a geological database. The preprocessing module is used to preprocess the geological data table to obtain the geological data table to be parsed; The extraction module is used to extract cells from the geological data table to be parsed, obtain the cell location information, and extract the cell content information from the cells corresponding to each location information using an OCR algorithm. The row and column information acquisition module is used to obtain the row and column information of a cell by using the cell's position information and table parsing methods; Specifically: Within the cells containing dividing lines, the image is converted into a binary image using the OpenCV model. The coordinates of the endpoints of the line segments in the binary image are identified, and the remaining cells without dividing lines are treated as rectangles with length and width. Set the shortest third of the longest side of the rectangle in each cell as the horizontal axis fault tolerance threshold k. x Set the shortest third of the rectangle in each cell as the fault tolerance threshold k on the y-axis. y The expression is: Where (x1, y1) is the coordinate of the top left corner of the cell, and (x2, y2) is the coordinate of the bottom right corner of the cell; The corrected cell coordinates are obtained by using the horizontal axis fault tolerance threshold, the vertical axis fault tolerance threshold, and the vertical relationship between rows and columns of the table. The row and column information of each cell are then obtained by using the corrected cell coordinates. In a cell containing a dividing line, locate the endpoint coordinates of the dividing line segment using an OpenCV model. When cell (a x b y When a given element is in the m-th row and n-th column, the coordinate relationship satisfies the following formula: Among them, a m b represents the minimum x-coordinate value of the cell in row m. n a represents the minimum y-coordinate value of the cell in the nth column. m+1 b represents the maximum x-coordinate value of the cell in row m. n+1 This represents the maximum y-coordinate value of the cell in column n. `start_row` and `end_row` represent the start and end points of the row information for the cell, respectively. `start_col` and `end_col` represent the start and end points of the column information for the cell, respectively. Use the cell number, top-left corner coordinates, and bottom-right corner coordinates as the cell's position information, and the row and column information as the cell's row and column information; The knowledge graph construction module converts the location, row, and column information of each cell into triplet data, aligns this triplet data with the triplet information obtained from parsing the table context, and generates a geological knowledge graph; specifically: The ALBERT model is used to transform the location information, row and column information, and content information of each cell into triplet data. The triplet data is represented as (a, r, b), where a and b represent entities and r represents the relationship between entities a and b. Aligning triplet-form data with parsed geological data tables involves entity alignment and embedding and fusing table-based knowledge graphs with context-based knowledge graphs to obtain a geological knowledge graph.

Images