A Method for Constructing a Semantic Understanding-Based Intelligent Annotation Platform for Nuclear Power Plants

Abstract

This invention discloses a method for constructing a nuclear power intelligent annotation platform based on semantic understanding, relating to the field of data processing technology. The method includes: determining a classification and recognition scheme for nuclear power file formats; manually annotating to form annotation sample data; establishing an annotation task recognition model, forming an annotation task decomposition mechanism, and forming an annotation precision setting mechanism; forming an annotation implementation mechanism, integrating the task recognition model, annotation task decomposition mechanism, and annotation precision setting mechanism into a nuclear power intelligent annotation platform; completing the construction of the nuclear power intelligent annotation platform after iterative training; and annotating different formats of content separately after format recognition. By forming a classification and recognition scheme, an annotation task recognition model, annotation task decomposition mechanism, annotation precision setting mechanism, and annotation implementation mechanism, different annotation methods are used to determine annotations based on different text formats, and different annotations are performed according to different annotation requirements.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method for constructing a nuclear power intelligent annotation platform based on semantic understanding. Background Technology

[0002] With the rapid development of artificial intelligence technologies, primarily machine learning, text annotation is becoming increasingly important in the field of AI. Text annotation is the process of labeling the features of text. Through labeled training data, machines can be taught how to recognize the intent contained in the text, enabling them to better understand natural language. In the nuclear power industry, a large amount of text needs to be autonomously identified, thus creating a significant demand for annotation.

[0003] However, text in the nuclear power field exists in various forms. Text such as images cannot be directly labeled using text recognition methods. At the same time, there are also various labeling requirements, which will have a certain impact on the accuracy of labeling and result in some room for improvement in the quality of various operations that rely on labeling. Summary of the Invention

[0004] To address the aforementioned technical issues, this technical solution provides a method for constructing a nuclear power intelligent annotation platform based on semantic understanding. The solution resolves the problems mentioned in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for constructing a semantic understanding-based intelligent annotation platform for nuclear power plants includes:

[0007] Obtain at least one nuclear power file format, which can be text, formula, table, or image.

[0008] Determine a classification and identification scheme for nuclear power plant document formats;

[0009] Manual annotation creates annotation sample data, which consists of the content included in the annotation range and annotation attributes;

[0010] Establish a labeling task identification model, form a labeling task decomposition mechanism, and form a labeling accuracy setting mechanism;

[0011] A labeling implementation mechanism is formed, and using this mechanism, the task identification model, labeling task decomposition mechanism, and labeling accuracy setting mechanism are integrated into a nuclear power intelligent labeling platform.

[0012] Based on the self-verification results of the annotation, the nuclear power intelligent annotation platform is iteratively trained, and after iterative training, the construction of the nuclear power intelligent annotation platform is completed.

[0013] The nuclear power intelligent annotation platform uses a nuclear power file format recognition scheme to identify the format of nuclear power files, and then annotates the content of different formats separately after format recognition.

[0014] Preferably, the classification and identification scheme for determining the nuclear power file format includes the following steps:

[0015] Obtain sample nuclear power files, and count the spacing between adjacent text, formulas, tables or images in the sample nuclear power files to obtain at least one sample spacing. The minimum value of the sample spacing is taken as the preset distance.

[0016] The sample nuclear power file is uniformly segmented to obtain at least one sample identification block, and the format of the sample identification block is manually determined.

[0017] The content of the nuclear power file format is uniformly divided to obtain at least one block to be identified. The block to be identified is a square and the side length of the block to be identified is a preset distance.

[0018] Obtain the pixel values ​​of the blank pixels as feature pixel values;

[0019] The average pixel value of each pixel in the sample recognition block is taken to obtain the sample mean;

[0020] The sample mean of the sample recognition block, which is a text, formula or table, is subtracted from the feature pixel value and the absolute value is taken to obtain at least one sample difference. The maximum value of the sample difference is taken as the critical value.

[0021] The average pixel values ​​of the pixels in the block to be identified are taken to obtain the mean value to be identified. The difference between the mean value to be identified and the feature pixel value is taken and the absolute value is obtained to obtain the difference to be identified. When the difference to be identified exceeds the threshold value, the format of the block to be identified corresponding to the difference to be identified is recognized as an image.

[0022] Aggregate adjacent blocks to be identified that are in the format of images to obtain at least one image region. Identify the shape of the image region. If it is not a rectangle, update the image region using the smallest rectangle that contains the image region. The format of the blocks to be identified within the image region is identified as images.

[0023] Blocks whose internal lines are of undefined format (horizontal, vertical, or L-shaped) are classified as unrestricted blocks.

[0024] Adjacent undefined blocks are aggregated to obtain at least one aggregated region. When the aggregated region is a closed box, the aggregated region and the area it encloses are taken as a table region. The format of the blocks to be identified within the table region is recognized as a table.

[0025] The maximum value of the average curvature of the lines in the sample recognition block formatted as text is taken as the first curvature;

[0026] Calculate the average curvature of the lines in the unformatted block to be recognized as the feature curvature. If the feature curvature is less than the first curvature, the format of the unformatted block to be recognized is determined as text; otherwise, the format of the unformatted block to be recognized is determined as formula.

[0027] Adjacent blocks to be recognized that are formatted as text are aggregated to form a text region, and adjacent blocks to be recognized that are formatted as formulas are aggregated to form a formula region.

[0028] Preferably, the manual annotation to form annotated sample data includes the following steps:

[0029] Obtain at least one sample nuclear power data. Based on the annotation requirements, use manual annotation to form annotated sample data. Determine the annotation range and annotation attributes of the content within the annotation range in the sample nuclear power data.

[0030] Preferably, the establishment of the annotation task recognition model includes the following steps:

[0031] At least one requirement for forming labeled sample data of sample nuclear power data is taken as a benchmark requirement. The benchmark requirement is split into word groups to obtain at least one benchmark word. The benchmark words are combined to form a benchmark vocabulary library.

[0032] Based on semantics, at least one set of similar words is formed, and the similar words in the set of similar words are all synonyms of the same base word;

[0033] An autonomous comparison and replacement mechanism is formed. Specifically, the acquired annotation task is split into at least one task word group, and the benchmark word corresponding to the set of similar words that are consistent with the task word group is obtained as the feature benchmark word. The corresponding task word group is replaced by the feature benchmark word.

[0034] When identifying annotation tasks, an autonomous comparison and replacement mechanism is used to replace the identified annotation task with a standard annotation task.

[0035] The annotation task input end and the standard annotation task output end are set up and combined with the autonomous comparison and replacement mechanism to form the annotation task recognition model.

[0036] Preferably, the method for forming the annotation task decomposition mechanism includes the following steps:

[0037] Summarize at least one baseline requirement into a baseline requirement set;

[0038] Obtain at least one part that overlaps with the baseline requirements and standard annotation tasks, and summarize it as a feature part;

[0039] The proportion of the calculated feature components in the baseline requirements is used to obtain the recognition ratio.

[0040] Baseline requirements with overlapping features are grouped into the same category to obtain a baseline requirement classification.

[0041] According to the recognition ratio, the baseline requirements in the baseline requirement category are sorted from largest to smallest. In the order of sorting, the baseline requirements in the baseline requirement category are deleted one by one, and the feature parts corresponding to the baseline requirements are deleted in the standard annotation task at the same time, until there are no feature parts corresponding to the baseline requirements in the baseline requirement category in the standard annotation task.

[0042] The baseline requirements that were deleted from the baseline requirement classification are respectively used as target baseline requirements, and the target baseline requirements are used as the output results of the annotation task decomposition.

[0043] Preferably, the mechanism for setting annotation precision includes the following steps:

[0044] Set the allowable error value for the annotation;

[0045] Feature extraction is performed on the contents contained in the labeled range of the labeled sample data to obtain the identified features;

[0046] Obtain the labeled sample data corresponding to the benchmark requirements. When performing recognition, if the proportion of the recognized content that is not included in the recognized feature is less than the allowable error value, then the recognized content is labeled as the labeled attribute in the labeled sample data corresponding to the recognized feature.

[0047] Preferably, the annotation formation mechanism includes the following steps:

[0048] A classification and recognition scheme is used to identify the format of nuclear power documents, resulting in text, formula, table, and image parts.

[0049] Extract the part that describes the formula from the text adjacent to the formula part, and use it as formula comments;

[0050] Extract the parts of the text adjacent to the table section that describe the table section, and use them as table notes;

[0051] Extract the descriptive text portion of the image from the text adjacent to the image portion, and use it as image captions;

[0052] Text, formula notes, table notes, and image notes are used as the content to be identified, and a labeling precision setting mechanism is used for identification and labeling.

[0053] Preferably, the method of using annotation implementation mechanism, which integrates task recognition model, annotation task decomposition mechanism and annotation accuracy setting mechanism into nuclear power intelligent annotation platform, includes the following steps:

[0054] Based on the annotation accuracy setting mechanism, the annotation implementation mechanism identifies the results of the task identification model and the annotation task decomposition mechanism. According to the identified topology, the annotation implementation mechanism, the task identification model, the annotation task decomposition mechanism and the annotation accuracy setting mechanism are integrated to form a nuclear power intelligent annotation platform.

[0055] Preferably, the iterative training of the nuclear power intelligent annotation platform based on the annotation self-verification results includes the following steps:

[0056] At least one requirement test set is formed by randomly combining the baseline requirements. The labeled sample data corresponding to the baseline requirements in the requirement test set are summarized to obtain the labeled sample data set.

[0057] Input the demand test set and sample nuclear power data into the nuclear power intelligent annotation platform to obtain the annotated test data set, and compare the annotated sample data set with the annotated test data set;

[0058] When the labeled sample data set and the labeled test data set do not overlap, the allowable error value is updated to half of the original value, and the updated labeled sample data set and the labeled test data set are compared. This step is repeated until the labeled sample data set and the labeled test data set overlap.

[0059] When the labeled sample data set and the labeled test data set overlap, the allowable error value is updated to twice the original value, and the updated labeled sample data set and the labeled test data set are compared. This step is repeated until the labeled sample data set and the labeled test data set no longer overlap, and then the allowable error value is updated to half of the current value.

[0060] The comparison is complete, and the iterative training is finished.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] By establishing a classification and recognition scheme, a labeling task recognition model, a labeling task decomposition mechanism, a labeling precision setting mechanism, and a labeling implementation mechanism, it is possible to accurately recognize text in various formats. Based on the recognition of different text formats, different methods are used to determine the labeling. At the same time, through the generation of various mechanisms and models, different labeling can be performed according to different labeling requirements. Furthermore, in order to control the workload of labeling, the precision in the labeling precision setting mechanism is controlled, thereby minimizing the computing power consumed by labeling while ensuring sufficient labeling accuracy, thus improving the efficiency of labeling. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the construction method of the nuclear power intelligent annotation platform based on semantic understanding according to the present invention.

[0064] Figure 2 This is a flowchart illustrating the classification and identification scheme for determining the format of nuclear power plant documents according to the present invention.

[0065] Figure 3 This is a schematic diagram of the process for establishing the annotation task recognition model of the present invention;

[0066] Figure 4 This is a flowchart illustrating the process of forming the annotation task decomposition mechanism of the present invention;

[0067] Figure 5 This is a flowchart illustrating the mechanism for setting annotation accuracy according to the present invention.

[0068] Figure 6 This is a flowchart illustrating the annotation formation mechanism of the present invention;

[0069] Figure 7 This is a schematic diagram illustrating the process of iteratively training the nuclear power intelligent annotation platform based on annotation self-verification results according to the present invention. Detailed Implementation

[0070] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0071] Reference Figure 1 As shown, the method for constructing a nuclear power intelligent annotation platform based on semantic understanding includes:

[0072] Obtain at least one nuclear power file format, which can be text, formula, table, or image.

[0073] Determine a classification and identification scheme for nuclear power plant document formats;

[0074] Manual annotation creates annotation sample data, which consists of the content included in the annotation range and annotation attributes;

[0075] Establish a labeling task identification model, form a labeling task decomposition mechanism, and form a labeling accuracy setting mechanism;

[0076] A labeling implementation mechanism is formed, and using this mechanism, the task identification model, labeling task decomposition mechanism, and labeling accuracy setting mechanism are integrated into a nuclear power intelligent labeling platform.

[0077] Based on the self-verification results of the annotation, the nuclear power intelligent annotation platform is iteratively trained, and after iterative training, the construction of the nuclear power intelligent annotation platform is completed.

[0078] The nuclear power intelligent annotation platform uses a nuclear power file format recognition scheme to identify the format of nuclear power files, and then annotates the content of different formats separately after format recognition.

[0079] Reference Figure 2 As shown, the classification and identification scheme for nuclear power file formats includes the following steps:

[0080] Obtain sample nuclear power files, and count the spacing between adjacent text, formulas, tables or images in the sample nuclear power files to obtain at least one sample spacing. The minimum value of the sample spacing is taken as the preset distance.

[0081] The sample nuclear power file is uniformly segmented to obtain at least one sample identification block, and the format of the sample identification block is manually determined.

[0082] The content of the nuclear power file format is uniformly divided to obtain at least one block to be identified. The block to be identified is a square and the side length of the block to be identified is a preset distance.

[0083] Obtain the pixel values ​​of the blank pixels as feature pixel values;

[0084] The average pixel value of each pixel in the sample recognition block is taken to obtain the sample mean;

[0085] The sample mean of the sample recognition block, which is a text, formula or table, is subtracted from the feature pixel value and the absolute value is taken to obtain at least one sample difference. The maximum value of the sample difference is taken as the critical value.

[0086] The average pixel values ​​of the pixels in the block to be identified are taken to obtain the mean value to be identified. The difference between the mean value to be identified and the feature pixel value is taken and the absolute value is obtained to obtain the difference to be identified. When the difference to be identified exceeds the threshold value, the format of the block to be identified corresponding to the difference to be identified is recognized as an image.

[0087] Aggregate adjacent blocks to be identified that are in the format of images to obtain at least one image region. Identify the shape of the image region. If it is not a rectangle, update the image region using the smallest rectangle that contains the image region. The format of the blocks to be identified within the image region is identified as images.

[0088] Blocks whose internal lines are of undefined format (horizontal, vertical, or L-shaped) are classified as unrestricted blocks.

[0089] Adjacent undefined blocks are aggregated to obtain at least one aggregated region. When the aggregated region is a closed box, the aggregated region and the area it encloses are taken as a table region. The format of the blocks to be identified within the table region is recognized as a table.

[0090] The maximum value of the average curvature of the lines in the sample recognition block formatted as text is taken as the first curvature;

[0091] Calculate the average curvature of the lines in the unformatted block to be recognized as the feature curvature. If the feature curvature is less than the first curvature, the format of the unformatted block to be recognized is determined as text; otherwise, the format of the unformatted block to be recognized is determined as formula.

[0092] Adjacent blocks to be recognized that are formatted as text are aggregated to form a text region, and adjacent blocks to be recognized that are formatted as formulas are aggregated to form a formula region.

[0093] Since nuclear power documents primarily consist of text, formulas, tables, or images, automatic annotation requires content recognition. Because text, formulas, tables, and images are different categories and may require separate annotation under certain needs, they need to be differentiated. However, recognition typically uses image comparison, making it difficult to distinguish between text, formulas, tables, and images. Therefore, this solution differentiates text, formulas, tables, and images based on their characteristics. When recognizing images, since most images are non-blank, the average pixel value differs significantly from the pixel value of blank areas. Text, formulas, and tables, being mostly blank, have a smaller difference in pixel value compared to blank areas. This allows for the identification of the corresponding block within the image using a threshold value. However, blank areas within the image are ignored. Since images are generally rectangular, the image area can be obtained by filling in the blank areas, and then all covered blocks can be recognized as images.

[0094] The table identifies the blocks to be identified within its covered area by recognizing its table frame. The edges of the table frame are horizontal or vertical lines, and L-shaped at the corners. Therefore, non-limited blocks can be identified in this way. However, non-limited blocks may be text or formulas. It is necessary to aggregate adjacent non-limited blocks. Non-limited blocks that are tables can always form closed frame shapes, while adjacent non-limited blocks that are text or formulas cannot form closed frame shapes. Therefore, the blocks to be identified as tables can be obtained accordingly.

[0095] The distinction between text and formulas relies on curvature. Formulas are mainly composed of letters and some symbols, which are almost all curved, so their curvature is relatively large. Text, on the other hand, is mainly straight. Therefore, curvature can be used as a dividing line to identify text and formulas.

[0096] The process of manually annotating sample data includes the following steps:

[0097] Obtain at least one sample nuclear power data. Based on the annotation requirements, use manual annotation to form annotated sample data. Determine the annotation range and annotation attributes of the content within the annotation range in the sample nuclear power data.

[0098] The labeled sample data is mainly used as a comparison for subsequent data training.

[0099] Reference Figure 3 As shown, establishing a labeling task recognition model includes the following steps:

[0100] At least one requirement for forming labeled sample data of sample nuclear power data is taken as a benchmark requirement. The benchmark requirement is split into word groups to obtain at least one benchmark word. The benchmark words are combined to form a benchmark vocabulary library.

[0101] Based on semantics, at least one set of similar words is formed, and the similar words in the set of similar words are all synonyms of the same base word;

[0102] An autonomous comparison and replacement mechanism is formed. Specifically, the acquired annotation task is split into at least one task word group, and the benchmark word corresponding to the set of similar words that are consistent with the task word group is obtained as the feature benchmark word. The corresponding task word group is replaced by the feature benchmark word.

[0103] When identifying annotation tasks, an autonomous comparison and replacement mechanism is used to replace the identified annotation task with a standard annotation task.

[0104] The annotation task input end and the standard annotation task output end are set up and combined with the autonomous comparison and replacement mechanism to form the annotation task recognition model.

[0105] When performing automatic annotation, it is necessary to identify the annotation requirements. These requirements may not be expressed in a strict manner. The baseline requirements may be scattered when actually expressed, and they may be expressed in other similar ways. Therefore, during identification, it is necessary to replace synonyms to obtain the required expression. When decomposing, since the expression is not strict, there may be repetitions, which will interfere with the identification. Therefore, the target baseline requirements are identified in sequence according to the identification ratio, so as to ensure accurate identification of the actual target baseline requirements.

[0106] Reference Figure 4 As shown, the steps to form a labeling task decomposition mechanism are as follows:

[0107] Summarize at least one baseline requirement into a baseline requirement set;

[0108] Obtain at least one part that overlaps with the baseline requirements and standard annotation tasks, and summarize it as a feature part;

[0109] The proportion of the calculated feature components in the baseline requirements is used to obtain the recognition ratio.

[0110] Baseline requirements with overlapping features are grouped into the same category to obtain a baseline requirement classification.

[0111] According to the recognition ratio, the baseline requirements in the baseline requirement category are sorted from largest to smallest. In the order of sorting, the baseline requirements in the baseline requirement category are deleted one by one, and the feature parts corresponding to the baseline requirements are deleted in the standard annotation task at the same time, until there are no feature parts corresponding to the baseline requirements in the baseline requirement category in the standard annotation task.

[0112] The baseline requirements that were deleted from the baseline requirement classification are respectively used as target baseline requirements, and the target baseline requirements are used as the output results of the annotation task decomposition.

[0113] Reference Figure 5 As shown, the mechanism for setting annotation precision includes the following steps:

[0114] Set the allowable error value for the annotation;

[0115] Feature extraction is performed on the contents contained in the labeled range of the labeled sample data to obtain the identified features;

[0116] Obtain the labeled sample data corresponding to the benchmark requirements. When performing recognition, if the proportion of the recognized content that is not included in the recognized feature is less than the allowable error value, then the recognized content is labeled as the labeled attribute in the labeled sample data corresponding to the recognized feature.

[0117] The allowable error value is mainly a setting for the accuracy of recognition. The higher the accuracy, the higher the precision. However, in actual use, such high precision may not be required to complete all annotations without errors. Therefore, the allowable error value needs to be determined during training so that the set allowable error value meets the requirements. This is because the smaller the allowable error value, the greater the computing power required for recognition and annotation.

[0118] Reference Figure 6 As shown, the annotation implementation mechanism includes the following steps:

[0119] A classification and recognition scheme is used to identify the format of nuclear power documents, resulting in text, formula, table, and image parts.

[0120] Extract the part that describes the formula from the text adjacent to the formula part, and use it as formula comments;

[0121] Extract the parts of the text adjacent to the table section that describe the table section, and use them as table notes;

[0122] Extract the descriptive text portion of the image from the text adjacent to the image portion, and use it as image captions;

[0123] Text, formula notes, table notes, and image notes are used as the content to be identified, and a labeling precision setting mechanism is used for identification and labeling.

[0124] Text recognition can be accomplished through semantic recognition, which is an existing technology. However, it is difficult to use semantic recognition for formulas, tables, or images. But according to common sense, formulas, tables, or images do not appear out of thin air. Therefore, the descriptions of their neighbors can be used as attributes of the formulas, tables, or images. Based on the semantics of the attributes, the annotation of the formulas, tables, or images can be determined.

[0125] The nuclear power intelligent annotation platform integrates the task identification model, annotation task decomposition mechanism, and annotation accuracy setting mechanism using an annotation implementation mechanism, including the following steps:

[0126] Based on the annotation accuracy setting mechanism, the annotation implementation mechanism identifies the results of the task identification model and the annotation task decomposition mechanism. According to the identified topology, the annotation implementation mechanism, the task identification model, the annotation task decomposition mechanism and the annotation accuracy setting mechanism are integrated to form a nuclear power intelligent annotation platform.

[0127] Reference Figure 7 As shown, the iterative training of the nuclear power intelligent annotation platform based on the annotation self-verification results includes the following steps:

[0128] At least one requirement test set is formed by randomly combining the baseline requirements. The labeled sample data corresponding to the baseline requirements in the requirement test set are summarized to obtain the labeled sample data set.

[0129] Input the demand test set and sample nuclear power data into the nuclear power intelligent annotation platform to obtain the annotated test data set, and compare the annotated sample data set with the annotated test data set;

[0130] When the labeled sample data set and the labeled test data set do not overlap, the allowable error value is updated to half of the original value, and the updated labeled sample data set and the labeled test data set are compared. This step is repeated until the labeled sample data set and the labeled test data set overlap.

[0131] When the labeled sample data set and the labeled test data set overlap, the allowable error value is updated to twice the original value, and the updated labeled sample data set and the labeled test data set are compared. This step is repeated until the labeled sample data set and the labeled test data set no longer overlap, and then the allowable error value is updated to half of the current value.

[0132] The comparison is complete, and the iterative training is finished.

[0133] During training, the allowable error value may be too large, causing inconsistencies between the annotations made by the nuclear power intelligent annotation platform and the annotation sample data. Therefore, it is necessary to reduce the allowable error value. Alternatively, if the allowable error value is too small, it will consume too much computing power. Therefore, the allowable error value is increased. Here, since the annotation sample data set and the annotation test data set do not overlap at the deadline, the previous allowable error value met the requirements, while the current allowable error value does not. Therefore, the allowable error value is updated to half of the current value.

[0134] Furthermore, this solution also proposes a storage medium on which a computer-readable program is stored. When the computer-readable program is invoked, the above-mentioned method for constructing a nuclear power intelligent annotation platform based on semantic understanding is executed.

[0135] It is understandable that the storage medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid-state drive (SSD).

[0136] In summary, the advantages of this invention are as follows: by forming a classification and recognition scheme, a labeling task recognition model, a labeling task decomposition mechanism, a labeling precision setting mechanism, and a labeling implementation mechanism, it can accurately recognize text in various formats and determine the labeling based on different text formats. Furthermore, through the generation of various mechanisms and models, different labeling can be performed according to different labeling requirements. Moreover, to control the workload of labeling, the precision in the labeling precision setting mechanism is controlled, thereby minimizing the computational power consumed by labeling while ensuring sufficient accuracy, thus improving labeling efficiency.

[0137] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for constructing a nuclear power intelligent annotation platform based on semantic understanding, characterized in that, include: Obtain at least one nuclear power file format, which can be text, formula, table, or image. Determine a classification and identification scheme for nuclear power plant document formats; Manual annotation creates annotation sample data, which consists of the content included in the annotation range and annotation attributes; Establish a labeling task identification model, form a labeling task decomposition mechanism, and form a labeling accuracy setting mechanism; A labeling implementation mechanism is formed, and using this mechanism, the task identification model, labeling task decomposition mechanism, and labeling accuracy setting mechanism are integrated into a nuclear power intelligent labeling platform. Based on the self-verification results of the annotation, the nuclear power intelligent annotation platform is iteratively trained, and after iterative training, the construction of the nuclear power intelligent annotation platform is completed. The nuclear power intelligent annotation platform uses a nuclear power file format recognition scheme to identify the format of nuclear power files, and then annotates the content of different formats separately after format recognition.

2. The method for constructing a nuclear power intelligent annotation platform based on semantic understanding according to claim 1, characterized in that, The classification and identification scheme for determining the format of nuclear power files includes the following steps: Obtain sample nuclear power files, and count the spacing between adjacent text, formulas, tables or images in the sample nuclear power files to obtain at least one sample spacing. The minimum value of the sample spacing is taken as the preset distance. The sample nuclear power file is uniformly segmented to obtain at least one sample identification block, and the format of the sample identification block is manually determined. The content of the nuclear power file format is uniformly divided to obtain at least one block to be identified. The block to be identified is a square and the side length of the block to be identified is a preset distance. Obtain the pixel values ​​of the blank pixels as feature pixel values; The average pixel value of each pixel in the sample recognition block is taken to obtain the sample mean; The sample mean of the sample recognition block, which is a text, formula or table, is subtracted from the feature pixel value and the absolute value is taken to obtain at least one sample difference. The maximum value of the sample difference is taken as the critical value. The average pixel values ​​of the pixels in the block to be identified are taken to obtain the mean value to be identified. The difference between the mean value to be identified and the feature pixel value is taken and the absolute value is obtained to obtain the difference to be identified. When the difference to be identified exceeds the threshold value, the format of the block to be identified corresponding to the difference to be identified is recognized as an image. Aggregate adjacent blocks to be identified that are in the format of images to obtain at least one image region. Identify the shape of the image region. If it is not a rectangle, update the image region using the smallest rectangle that contains the image region. The format of the blocks to be identified within the image region is identified as images. Blocks whose internal lines are of undefined format (horizontal, vertical, or L-shaped) are classified as unrestricted blocks. Adjacent undefined blocks are aggregated to obtain at least one aggregated region. When the aggregated region is a closed box, the aggregated region and the area it encloses are taken as a table region. The format of the blocks to be identified within the table region is recognized as a table. The maximum value of the average curvature of the lines in the sample recognition block formatted as text is taken as the first curvature; Calculate the average curvature of the lines in the unformatted block to be recognized as the feature curvature. If the feature curvature is less than the first curvature, the format of the unformatted block to be recognized is determined as text; otherwise, the format of the unformatted block to be recognized is determined as formula. Adjacent blocks to be recognized that are formatted as text are aggregated to form a text region, and adjacent blocks to be recognized that are formatted as formulas are aggregated to form a formula region.

3. The method for constructing a nuclear power intelligent annotation platform based on semantic understanding according to claim 2, characterized in that, The manual annotation process for generating labeled sample data includes the following steps: Obtain at least one sample nuclear power data. Based on the annotation requirements, use manual annotation to form annotated sample data. Determine the annotation range and annotation attributes of the content within the annotation range in the sample nuclear power data.

4. The method for constructing a nuclear power intelligent annotation platform based on semantic understanding according to claim 3, characterized in that, The establishment of the annotation task recognition model includes the following steps: At least one requirement for forming labeled sample data of sample nuclear power data is taken as a benchmark requirement. The benchmark requirement is split into word groups to obtain at least one benchmark word. The benchmark words are combined to form a benchmark vocabulary library. Based on semantics, at least one set of similar words is formed, and the similar words in the set of similar words are all synonyms of the same base word; An autonomous comparison and replacement mechanism is formed. Specifically, the acquired annotation task is split into at least one task word group, and the benchmark word corresponding to the set of similar words that are consistent with the task word group is obtained as the feature benchmark word. The corresponding task word group is replaced by the feature benchmark word. When identifying annotation tasks, an autonomous comparison and replacement mechanism is used to replace the identified annotation task with a standard annotation task. The annotation task input end and the standard annotation task output end are set up and combined with the autonomous comparison and replacement mechanism to form the annotation task recognition model.

5. The method for constructing a nuclear power intelligent annotation platform based on semantic understanding according to claim 4, characterized in that, The mechanism for forming the annotation task decomposition includes the following steps: Summarize at least one baseline requirement into a baseline requirement set; Obtain at least one part that overlaps with the baseline requirements and standard annotation tasks, and summarize it as a feature part; The proportion of the calculated feature components in the baseline requirements is used to obtain the recognition ratio. Baseline requirements with overlapping features are grouped into the same category to obtain a baseline requirement classification. According to the recognition ratio, the baseline requirements in the baseline requirement category are sorted from largest to smallest. In the order of sorting, the baseline requirements in the baseline requirement category are deleted one by one, and the feature parts corresponding to the baseline requirements are deleted in the standard annotation task at the same time, until there are no feature parts corresponding to the baseline requirements in the baseline requirement category in the standard annotation task. The baseline requirements that were deleted from the baseline requirement classification are respectively used as target baseline requirements, and the target baseline requirements are used as the output results of the annotation task decomposition.

6. The method for constructing a nuclear power intelligent annotation platform based on semantic understanding according to claim 5, characterized in that, The mechanism for setting annotation accuracy includes the following steps: Set the allowable error value for the annotation; Feature extraction is performed on the contents contained in the labeled range of the labeled sample data to obtain the identified features; Obtain the labeled sample data corresponding to the benchmark requirements. When performing recognition, if the proportion of the recognized content that is not included in the recognized feature is less than the allowable error value, then the recognized content is labeled as the labeled attribute in the labeled sample data corresponding to the recognized feature.

7. The method for constructing a nuclear power intelligent annotation platform based on semantic understanding according to claim 6, characterized in that, The annotation formation mechanism includes the following steps: A classification and recognition scheme is used to identify the format of nuclear power documents, resulting in text, formula, table, and image parts. Extract the part that describes the formula from the text adjacent to the formula part, and use it as formula comments; Extract the parts of the text adjacent to the table section that describe the table section, and use them as table notes; Extract the descriptive text portion of the image from the text adjacent to the image portion, and use it as image captions; Text, formula notes, table notes, and image notes are used as the content to be identified, and a labeling precision setting mechanism is used for identification and labeling.

8. The method for constructing a nuclear power intelligent annotation platform based on semantic understanding according to claim 7, characterized in that, The aforementioned annotation implementation mechanism, which integrates the task identification model, annotation task decomposition mechanism, and annotation accuracy setting mechanism into a nuclear power intelligent annotation platform, includes the following steps: Based on the annotation accuracy setting mechanism, the annotation implementation mechanism identifies the results of the task identification model and the annotation task decomposition mechanism. According to the identified topology, the annotation implementation mechanism, the task identification model, the annotation task decomposition mechanism and the annotation accuracy setting mechanism are integrated to form a nuclear power intelligent annotation platform.

9. The method for constructing a nuclear power intelligent annotation platform based on semantic understanding according to claim 8, characterized in that, The iterative training of the nuclear power intelligent annotation platform based on the annotation self-verification results includes the following steps: At least one requirement test set is formed by randomly combining the baseline requirements. The labeled sample data corresponding to the baseline requirements in the requirement test set are summarized to obtain the labeled sample data set. Input the demand test set and sample nuclear power data into the nuclear power intelligent annotation platform to obtain the annotated test data set, and compare the annotated sample data set with the annotated test data set; When the labeled sample data set and the labeled test data set do not overlap, the allowable error value is updated to half of the original value, and the updated labeled sample data set and the labeled test data set are compared. This step is repeated until the labeled sample data set and the labeled test data set overlap. When the labeled sample data set and the labeled test data set overlap, the allowable error value is updated to twice the original value, and the updated labeled sample data set and the labeled test data set are compared. This step is repeated until the labeled sample data set and the labeled test data set no longer overlap, and then the allowable error value is updated to half of the current value. The comparison is complete, and the iterative training is finished.

Images