Private placement fund announcement information key element intelligent extraction and labeling method and system based on natural language processing

Abstract

The invention provides an intelligent extraction and labeling method and system for key elements of private funds announcement information based on natural language processing, and relates to the technical field of natural language processing. Document image data under different wavelengths are acquired by scanning private funds announcement information and fused to obtain a fused image; a text area is separated from the text data by using a support vector machine, and an image text in the text area is converted into machine-readable text data through optical character recognition; performing word segmentation, part-of-speech tagging and named entity recognition on the text data through a natural language processing technology to obtain structured character feature parameters; and then analyzing the parameter through a naive Bayesian algorithm to identify and label key elements to obtain an element labeling result. According to the method, the key information in the announcement can be efficiently obtained, so that the accuracy and efficiency of private fund announcement information processing are improved.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method and system for intelligent extraction and annotation of key elements of private equity fund announcement information based on natural language processing. Background Technology

[0002] In the financial sector, private equity fund announcements serve as the core basis for regulatory agencies to conduct filing and verification, for investors to understand product information, and for fund managers to carry out compliance management. Key elements contained in these announcements, such as the amount raised, the name of the manager, the duration, and the investment scope, are important foundations for decision-making and supervision by various entities. As the scale of the private equity fund industry expands, the number of announcements is growing explosively. Manually extracting key elements is inefficient, time-consuming, and susceptible to errors due to subjective factors. Therefore, there is an urgent need for an intelligent technology solution that can automatically and accurately extract and label key elements.

[0003] Currently, existing technologies for extracting document information typically involve first acquiring image data of private equity fund announcements using ordinary image acquisition equipment, then converting the text content in the image into editable text data; subsequently, performing simple word segmentation on the text data and filtering possible entity information using preset fixed keywords; finally, manually verifying the filtered information, and manually marking key elements after confirmation to form a key element record result.

[0004] However, existing technologies have significant drawbacks: First, image acquisition relies on a single wavelength. If the private equity fund announcement has faded paper, ink smudges, or surface stains, the acquired image clarity is insufficient, directly reducing the accuracy of subsequent optical character recognition. Second, there is a lack of precise text region separation methods, which easily includes non-text content such as logos, table lines, and page numbers in the announcement within the text extraction scope, thus interfering with the quality of effective text data. Third, key element identification relies on fixed keyword matching, which cannot handle the flexible key information in the announcement. This easily leads to high rates of missed detections and false detections, and the final annotation results are mostly unstructured data, thus failing to meet the requirements for accurate key element extraction. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for intelligent extraction and annotation of key elements of private equity fund announcement information based on natural language processing, so as to solve the problem of low accuracy and efficiency in the extraction of key elements in the existing technology.

[0006] To address the aforementioned technical problems, firstly, this application provides a method for intelligent extraction and annotation of key elements in private equity fund announcements based on natural language processing, comprising:

[0007] The target announcement information is scanned by a multispectral image acquisition device to obtain document image data at different wavelengths. The document image data at all wavelengths are then fused to obtain a fused image.

[0008] The text region is separated from the fused image by a support vector machine, and optical character recognition processing is performed on the text region to convert the image text within the text region into machine-readable text data.

[0009] The text data is segmented, part-of-speech tagging is performed, and named entity recognition is performed using natural language processing techniques to obtain structured character feature parameters;

[0010] The character feature parameters are combined and analyzed using the Naive Bayes algorithm to identify key elements in the target announcement information. These key elements are then labeled to obtain the element labeling results.

[0011] Optionally, the step of performing word segmentation, part-of-speech tagging, and named entity recognition on the text data using natural language processing techniques to obtain structured character feature parameters includes:

[0012] The text data is segmented using a bidirectional long short-term memory network in natural language processing to obtain a word sequence. The word sequence is then labeled with part-of-speech tags to obtain words with part-of-speech information.

[0013] By using a recognition model that integrates a domain dictionary in natural language processing technology, named entity recognition is performed on the words with part-of-speech information to obtain words containing entity category tags. Information integration processing is then performed on the words containing entity category tags to obtain associated information groups.

[0014] All the associated information groups are sorted in an orderly manner to obtain structured character feature parameters.

[0015] Optionally, the step of performing named entity recognition processing on the words with part-of-speech information using a recognition model that integrates a domain dictionary in natural language processing technology to obtain words containing entity category tags includes:

[0016] A recognition model for a fusion domain dictionary is constructed, the model comprising a word embedding layer, a feature fusion layer, a context network layer, and a conditional random field layer;

[0017] Through the word embedding layer, the words with part-of-speech information are divided into paragraphs according to their part-of-speech categories to obtain multiple word segments. All word segments are then vectorized to obtain word vector representations.

[0018] The word vector representation is input into the feature fusion layer. The feature fusion layer performs domain feature matching on the word vector representation using a preset domain dictionary to generate a domain feature vector. The word vector representation and the domain feature vector are then fused to obtain a fused feature vector.

[0019] The fused feature vector is input into the context network layer for context feature extraction to obtain the context feature vector;

[0020] The context feature vector is input into the conditional random field layer for sequence labeling to obtain the entity category label corresponding to each word, thus obtaining words containing entity category labels.

[0021] Optionally, the step of combining and analyzing the character feature parameters using the Naive Bayes algorithm to identify key elements in the target announcement information, and then labeling the key elements to obtain element labeling results includes:

[0022] The character feature parameters are divided according to entity category to obtain multiple entity category groups;

[0023] The entity category groups are combined and analyzed using the Naive Bayes algorithm to obtain related words. The related words are then semantically integrated to obtain key elements.

[0024] The key elements are labeled with type to obtain key elements with type identifiers. The key elements with type identifiers are then arranged in order to obtain the element labeling result.

[0025] Optionally, the entity category group is combined and analyzed using the Naive Bayes algorithm to obtain related words, and the related words are then semantically integrated to obtain key elements, including:

[0026] Word extraction is performed on the entity category group to obtain a word set. The word set is then analyzed for correlation using the Naive Bayes algorithm to obtain correlation data.

[0027] The correlation data is filtered, and the filtered correlation data is subjected to an integrity check to obtain multiple related words;

[0028] All the associated words are sequentially checked to obtain an initial word sequence. The initial word sequence is then checked for coherence to obtain the target word sequence.

[0029] The target word sequence is semantically merged to obtain key elements.

[0030] Optionally, the step of separating the text region from the fused image using a support vector machine and performing optical character recognition processing on the text region to convert the image text within the text region into machine-readable text data includes:

[0031] The fused image is divided into regions according to a preset size to obtain multiple sub-images. Feature information is extracted from each sub-image to obtain the feature information of each sub-image.

[0032] The feature information is classified using a support vector machine to obtain a category identifier for each sub-image. Based on the category identifier, sub-images that meet preset conditions are filtered and aggregated to obtain the text region.

[0033] The image text within the text area is enhanced to obtain enhanced image text, and the enhanced image text is then split to obtain individual character images;

[0034] Optical character recognition (OCR) is performed on the individual character images to obtain character data corresponding to each individual character image. All the character data are then sequentially combined to obtain machine-readable text data.

[0035] Optionally, the step of fusing the document image data at all wavelengths to obtain a fused image includes:

[0036] The document image data at all wavelengths is divided into regions to obtain the text distribution area and non-text distribution area in each document image data.

[0037] Image information extraction processing is performed on all the text distribution areas and the non-text distribution areas respectively to obtain multiple text distribution information and multiple non-text distribution information. All the text distribution information is superimposed to obtain a first combined image.

[0038] Select one of the non-text distribution information as the reference information, and then stitch the remaining non-text distribution information with the reference information to obtain the second combined image;

[0039] The first combined image and the second combined image are integrated to obtain a fused image.

[0040] Secondly, this application provides an intelligent extraction and annotation system for key elements of private equity fund announcement information based on natural language processing, including:

[0041] The fusion module is used to scan the target announcement information through a multispectral image acquisition device to obtain document image data at different wavelengths, and to fuse the document image data at all wavelengths to obtain a fused image.

[0042] The conversion module is used to separate the text region from the fused image using a support vector machine, and to perform optical character recognition processing on the text region to convert the image text within the text region into machine-readable text data.

[0043] The recognition module is used to perform word segmentation, part-of-speech tagging, and named entity recognition on the text data using natural language processing technology to obtain structured character feature parameters;

[0044] The analysis module is used to perform combined analysis on the character feature parameters using the Naive Bayes algorithm to identify key elements in the target announcement information, and to annotate the key elements to obtain element annotation results.

[0045] Thirdly, this application provides an electronic device, comprising:

[0046] Memory, used to store computer programs;

[0047] A processor, used to execute the computer program to implement the steps of the intelligent extraction and annotation method for key elements of private equity fund announcement information based on natural language processing as described in the first aspect above.

[0048] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the intelligent extraction and annotation method for key elements of private equity fund announcement information based on natural language processing as described in the first aspect above.

[0049] The intelligent extraction and annotation method for key elements of private equity fund announcements based on natural language processing provided in this application has the following beneficial effects:

[0050] This application firstly overcomes the limitations of single-wavelength image acquisition by using multispectral acquisition and fusion processing. Even if the announcement has paper fading, ink bleeding, or surface stains, image data from different wavelengths can complementarily present the differences between text and background. Moreover, the fused image can significantly improve the clarity and contrast of the text area, thus laying a high-quality image foundation for subsequent accurate separation of the text area. Then, based on the strong classification and recognition capabilities of support vector machines, the text area in the fused image can be accurately distinguished from non-text areas such as logos, table lines, and page numbers, thereby avoiding interference from non-text content in subsequent processing.

[0051] Then, optical character recognition based on the separated clean text region can significantly reduce the recognition error caused by non-text interference, thereby ensuring that the converted machine-readable text data is complete and accurate, and meets the input requirements of subsequent natural language processing. After that, word segmentation can break down the coherent text into semantically independent word units and perform part-of-speech tagging, thereby assigning grammatical attributes to each word. Then, named entity recognition can locate key entities in the text.

[0052] Subsequently, the Naive Bayes algorithm can efficiently combine and analyze the relationships between words in the character feature parameters to identify key elements such as the amount raised and the name of the manager. At the same time, the element annotation results formed by the annotation process not only clarify the content and type of key elements, but also maintain their original order in the announcement, thereby providing directly usable structured annotation information for subsequent regulatory verification, data statistics and other scenarios.

[0053] Furthermore, this application utilizes a bidirectional long short-term memory network to achieve word segmentation and part-of-speech tagging, whose contextual semantic capture capabilities improve the accuracy of word segmentation and part-of-speech tagging. Subsequently, the recognition model, which integrates a dictionary of private equity funds, can specifically identify domain-specific terms and solve the problem of insufficient recognition of domain terms by general models. Then, through multi-layer collaboration, words containing entity category tags are obtained, and combined with structured character feature parameters formed by information integration and orderly arrangement, further improving the domain adaptability and structure of the data. This provides higher-quality data that is more suitable for the private equity fund scenario for the subsequent Naive Bayes algorithm to accurately identify key elements, thereby significantly reducing the missed detection and false detection rate of key elements. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of this application 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 of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating a method for intelligent extraction and annotation of key elements of private equity fund announcement information based on natural language processing, provided for an embodiment of this application;

[0056] Figure 2 A schematic diagram illustrating a specific implementation of a method for intelligent extraction and annotation of key elements of private equity fund announcement information based on natural language processing, provided in this application embodiment;

[0057] Figure 3A schematic diagram of the structure of a key element extraction and annotation system for private equity fund announcement information based on natural language processing, provided in an embodiment of this application;

[0058] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0059] Against the backdrop of the rapid development of the private equity industry, the efficiency and accuracy of extracting key elements from announcements directly impact regulatory verification, investor decision-making, and manager compliance management. However, current reliance on single-wavelength scanning is insufficient to address issues such as fading and stains in announcements, easily leading to blurred images and affecting text recognition. Furthermore, the lack of precise text region separation methods often results in the inclusion of non-text content such as logos and table lines within the processing scope, thus interfering with the extraction of effective information. Fixed keyword matching cannot adapt to the flexible expression methods in announcements, easily leading to missed or false detections of key elements, and the final results are mostly unstructured data, making it difficult to use directly for subsequent analysis. These problems collectively make it difficult for existing technologies to meet the demand for efficient and accurate extraction of key elements.

[0060] To address the aforementioned issues, this application proposes an intelligent extraction and annotation method for key elements of private equity fund announcements based on natural language processing. This method captures announcement images at different wavelengths using a multispectral image acquisition device and fuses them to form a clearer fused image, effectively overcoming the blurring problem of single-wavelength images. Then, it utilizes a support vector machine to accurately separate text regions, avoiding interference from non-textual content and improving the accuracy of subsequent optical character recognition. Subsequently, natural language processing techniques are used to segment, tag, and name entity recognize the recognized text, transforming it into structured character feature parameters. Finally, a Naive Bayes algorithm is employed to combine and analyze the feature parameters, accurately identifying key elements and completing the annotation. This solution specifically addresses the shortcomings of existing technologies, such as poor image quality, non-textual interference, low recognition accuracy, and unstructured data, enabling automatic and accurate extraction and annotation of key elements in private equity fund announcements, balancing efficiency and practicality.

[0061] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] The core of this application is to provide a method for intelligent extraction and annotation of key elements in private equity fund announcements based on natural language processing. A flowchart of one specific implementation is shown below. Figure 1As shown, the method includes:

[0063] S101. The target announcement information is scanned by a multispectral image acquisition device to obtain document image data at different wavelengths. The document image data at all wavelengths are fused to obtain a fused image.

[0064] In the financial field, the target announcement information may refer to private equity fund announcements, other announcements in the same field, or announcements from other fields. This application embodiment does not specifically limit the type and content of the announcement information, and can set it accordingly based on the actual situation.

[0065] Furthermore, a multispectral image acquisition device refers to a specialized device capable of emitting two or more different wavelength ranges of spectrum and scanning the target announcement information to acquire announcement image data under multiple different wavelengths; document image data of different wavelength ranges refers to the image data generated after scanning by the device, with each wavelength range corresponding to one image containing all the visual content of the private equity fund announcement.

[0066] In one specific implementation, step S101 involves fusing the document image data across all wavelengths to obtain a fused image, including:

[0067] Step 1011: Perform region segmentation processing on the document image data under all wavelengths to obtain the text distribution area and non-text distribution area in each document image data.

[0068] In this application example, the document image data under all wavelengths is divided into regions, and the image gray value analysis method is used to mark the regions that meet the text features in the image under all wavelengths as text distribution areas and the regions that meet the non-text features as non-text distribution areas, thereby obtaining the text distribution areas and non-text distribution areas in each document image data.

[0069] Step 1012: Extract image information from all text distribution areas and non-text distribution areas to obtain multiple text distribution information and multiple non-text distribution information. Superimpose all the text distribution information to obtain a first combined image.

[0070] The first combined image refers to the image formed by superimposing the text distribution information under all different wavelengths according to the corresponding pixel positions. Compared with the image before combination, the text features in this image are clearer and have higher contrast.

[0071] In step 1012, the grayscale value and edge contour information of the text are extracted from the text distribution area to form multiple text distribution information; then, the color pixel value and contour information of the non-text content are extracted from the non-text distribution area to form multiple non-text distribution information; then, all text distribution information is superimposed according to the pixel coordinate position, and the grayscale value of the same pixel position is calculated by the arithmetic mean method during the superposition process, and finally the first combined image is obtained.

[0072] Step 1013: Select one of the non-text distribution information as the reference information, and stitch the remaining non-text distribution information with the reference information to obtain the second combined image.

[0073] Among them, the non-text distribution information can be selected based on the core fixed content in the target announcement information. The core fixed content refers to the non-text content in the private equity fund announcement whose position and style remain unchanged for a long time, and usually has an identifying function; the benchmark information refers to a non-text distribution information that can most clearly present the core fixed content selected from multiple non-text distribution information; the second combined image refers to the non-text area image with complete core fixed content and no missing details formed by splicing the remaining non-text distribution information and the benchmark information according to the corresponding pixel positions.

[0074] In step 1013, the non-text distribution information with the clearest core fixed content is selected as the reference information; then the remaining non-text distribution information is aligned with the reference information according to pixel coordinates, and the clear area of ​​the core fixed content in the reference information is preserved during the stitching process, while the missing details in the reference information in the remaining non-text distribution information are supplemented, and finally the second combined image is obtained.

[0075] Step 1014: Integrate the first combined image with the second combined image to obtain a fused image.

[0076] Among them, the fused image refers to an image formed by integrating the first combination image and the second combination image according to the original visual layout of the private equity fund announcement, which includes clear text areas and complete non-text areas and can completely restore the visual content of the announcement.

[0077] In step 1014, the pixel coordinate system of the first combined image and the second combined image is first confirmed to ensure that there is no positional deviation when the two are integrated; then the first combined image and the second combined image are integrated according to the original visual layout of the private equity fund announcement, and the clear features of the text area in the first combined image and the complete features of the non-text area in the second combined image are preserved during the integration process. Then, the text details are preserved in the coordinate overlapping area according to the text priority principle, and finally the fused image is obtained.

[0078] This application uses multispectral scanning to compensate for the information loss caused by paper fading and stains in single-wavelength scanning; it can also accurately distinguish between text and non-text areas through region division, thereby avoiding interference between the two; then, the text areas are superimposed to form the first combined image, which can improve the text blurring problem. The integrated fused image takes into account both text clarity and non-text integrity, and can completely restore the visual content of the private equity fund announcement.

[0079] S102. Separate the text region from the fused image using a support vector machine, and perform optical character recognition processing on the text region to convert the image text within the text region into machine-readable text data.

[0080] In one specific implementation, step S102 includes:

[0081] Step 1021: Divide the fused image into regions according to a preset size to obtain multiple sub-images, and extract feature information from each sub-image to obtain feature information of each sub-image.

[0082] The preset size refers to the fixed pixel size set in advance for dividing the fused image. This application does not specifically limit the size of this size and can set it according to the actual situation. The sub-image refers to the small image block formed after the fused image is divided according to the preset size. The feature information refers to the information extracted from the sub-image that can distinguish between text and non-text attributes.

[0083] In this application example, step 1021 divides the fused image into regions according to a preset size. During the division, the image is divided horizontally and vertically according to the preset size, starting from the image's starting coordinates, to ensure that the divided regions do not overlap and completely cover the fused image, resulting in multiple sub-images. Subsequently, feature information extraction is performed on each sub-image, and image feature extraction technology is used to calculate the grayscale mean and edge density of each sub-image. Then, texture features of the sub-image are extracted using a texture analysis algorithm, and these information are integrated to obtain the feature information of each sub-image.

[0084] For example, consider the fused image of a company's series of private equity fund raising announcements. The image has a pixel size of 2000×3000. It is then divided into regions according to a preset size of 20×20 pixels. When dividing, start from the horizontal coordinate 0 and the vertical coordinate 0. Move 20 pixels horizontally to determine the horizontal boundary of a sub-image, and move 20 pixels vertically to determine the vertical boundary of a sub-image. Thus, the number of sub-images that can be divided horizontally is 2000÷20=100, and the number of sub-images that can be divided vertically is 3000÷20=150, resulting in 100×150=15000 sub-images.

[0085] Subsequently, feature information extraction processing is performed on each sub-image: using the gray-scale mean calculation formula. Calculate the grayscale mean, where, denoted as the average grayscale value of the sub-image, m as the horizontal pixel count of the sub-image (here, m is 20), and n as the vertical pixel count of the sub-image (here, n is 20). Let the gray value of the pixel in the i-th row and j-th column of the sub-image be given. Then, the edge detection algorithm is used to identify the edge pixels in the sub-image and calculate the ratio of the number of edge pixels to the total number of pixels in the sub-image to obtain the edge density. The total number of pixels in the sub-image is 20×20=400. Then, the texture feature value of the sub-image is extracted by the gray-level co-occurrence matrix algorithm. Finally, the gray-level mean, edge density and texture feature value are integrated to obtain the feature information of each sub-image.

[0086] Step 1022: Classify the feature information using a support vector machine to obtain the category identifier corresponding to each sub-image. Based on the category identifier, filter and aggregate sub-images that meet preset conditions to obtain the text region.

[0087] Support Vector Machine (SVM) is a technical tool used for data classification. It can build a classification model by learning the data features of known categories, and then make a category judgment for data of unknown categories. Category label refers to the label assigned to each sub-image after classification by SVM, which is used to indicate its attributes. Preset conditions refer to the rules set in advance for filtering sub-images. Here, the preset condition is that the category label of the sub-image is text.

[0088] In this application example, step 1022 inputs the feature information of each sub-image into a pre-trained support vector machine classification model. The model determines the attributes of each sub-image by comparing the differences between the input feature information and the training data, and assigns a corresponding category label to each sub-image. Then, based on the category label, sub-images that meet the preset conditions are filtered out. After filtering out all sub-images with the category label as text, adjacent text sub-images are sequentially connected and integrated according to the original coordinate positions of these sub-images in the fused image to exclude scattered non-text sub-images and obtain the text region.

[0089] For example, the feature information of sub-images from a series of private equity fund raising announcements is input into a pre-trained support vector machine classification model. The training data of this model contains 10,000 text-based sub-image features and 10,000 non-text-based sub-image features, and after training, it has the ability to distinguish between the two types of sub-images. Then, by comparing the differences between the input feature information and the training data, the model assigns text-based labels to sub-images containing text such as "duration: 5 years" and "custodian: bank", and assigns non-text-based labels to sub-images containing the company logo and page number "page 1". Then, based on the category labels, sub-images that meet the preset conditions are selected, and all sub-images with the category label of text are selected. The original coordinates of these sub-images are then counted, and it is found that sub-images with horizontal coordinates of 20-1800 and vertical coordinates of 200-2700 are mostly text-based. Then, these sub-images are aggregated in order of coordinates, and non-text-based sub-images with scattered coordinates are deleted, finally forming a complete text region.

[0090] Step 1023: Enhance the image text within the text area to obtain enhanced image text, and then split the enhanced image text to obtain individual character images.

[0091] In this application example, the image text within the text area is enhanced in step 1023, and the grayscale range of the image text is adjusted by grayscale stretching technology, thereby expanding the grayscale difference between the text and the background. Then, the small noise points in the text area are eliminated by noise removal algorithm to obtain the enhanced image text. Subsequently, the enhanced image text is split, and the blank area between characters in the enhanced image text is identified by character segmentation algorithm. The continuous image text is segmented character by character with the blank area as the segmentation boundary to obtain a single character image.

[0092] For example, enhancing the text area of ​​the company's series of private equity fund raising announcements: using a grayscale stretching algorithm to adjust the grayscale range of the text area image from the original 50-200 to 0-255, using the formula... Calculate the adjusted pixel grayscale value, where, G represents the adjusted pixel grayscale value, and G represents the original pixel grayscale value. The minimum grayscale value of the original text region image, such as the minimum grayscale value of the original text region image. It is 50. The maximum grayscale value of the original text region image, such as the maximum grayscale value of the original text region image. The value is 200. Taking a pixel with an original grayscale value of 80 as an example, substituting it into the formula yields... This is to enhance the contrast between the text and the background.

[0093] Then, the median filtering algorithm is used to remove noise. First, the filtering window size is set to 3×3, and the enhanced image text is obtained after removal; Subsequently, the enhanced image text is split: The projection method is used to identify the character boundaries, and the horizontal gray projection value of the image text is calculated. The areas with projection values close to 0 are the character spacings. Subsequently, the continuous text such as "Raised Amount: 5 million yuan" is split into multiple single character images along these areas, and the size of each single character image is uniformly 30×30 pixels.

[0094] Step 1024: Perform optical character recognition processing on the single character images to obtain the character data corresponding to each single character image, and perform sequential combination processing on all the character data to obtain machine-readable text data.

[0095] In the example of this application, through step 1024, optical character recognition processing is performed on each single character image, and the single character image is input into the optical character recognition system. The system can determine the character data corresponding to each character image by comparing the shape and contour features of the character image with the feature templates in the preset character library; Then, sequential combination processing is performed on all the character data according to the arrangement order of the single character images in the original text area, and the character data is spliced in sequence to form a continuous text sequence, obtaining machine-readable text data.

[0096] For example, perform optical character recognition processing on the single character images of the company's series of private equity fund raising announcements: Input each single character image into the optical character recognition system, and the system calls the preset character library, which contains the feature templates of common Chinese characters, 0-9 digits, and common punctuation marks, and compares the similarity between the character image and the template through the image matching algorithm. When the similarity reaches the preset threshold, the corresponding character is determined. For example, the similarity between the "募" character image and the "募" character template in the character library meets the requirements, and it is recognized as the character data "募". Finally, the character data corresponding to all single character images is obtained; Subsequently, the character data is spliced in sequence according to the order of the single character images in the original text area. For example, "募", "集", "金", "额", ":", "5", "0", "0", "万", "元" are spliced in sequence into "募集金额: 5 million yuan", and machine-readable text data is obtained after all the character data is spliced.

[0097] This application reduces the processing complexity by dividing the fused image according to the preset size, and can accurately distinguish the text and non-text attributes through feature extraction; Subsequently, after support vector machine classification, a complete text area is aggregated, and the text and non-text can be completely separated; Then, the clarity is improved through text enhancement, the character interference is split and eliminated, and then it is converted into character data through optical character recognition, and the machine-readable text is obtained by combining in the original order, thus completely getting rid of the image form limitation.

[0098] S103. The text data is segmented, part-of-speech tagging and named entity recognition are performed using natural language processing technology to obtain structured character feature parameters.

[0099] In the embodiments of this application, such as Figure 2 As shown, step S103 includes:

[0100] Step 1031: The text data is segmented using a bidirectional long short-term memory network in natural language processing to obtain a word sequence. The word sequence is then labeled with part-of-speech tags to obtain words with part-of-speech information.

[0101] Among them, bidirectional long short-term memory network refers to a technical tool that can capture the semantic relationship of text context. It can accurately split continuous text by analyzing the content before and after words and is used for word segmentation of text data. It should be noted that the embodiments of this application do not specifically limit the structure and specific implementation process of the network.

[0102] In this application example, text data is input into a bidirectional long short-term memory network in natural language processing technology through step 1031. The network analyzes the collocation and semantic association of words in the text to split the continuous text into independent words and obtain a word sequence. Then, a part-of-speech tagging tool based on a combination of rules and statistics is used to tag each word in the word sequence according to the grammatical function and semantic attributes of the words in the text, so that each word is bound to a corresponding part-of-speech tag, and finally words with part-of-speech information are obtained.

[0103] For example, machine-readable text data of a company's series of private equity fund raising announcements, which reads "The company's series of private equity funds raised 5 million yuan, with a bank as the custodian and a term of 5 years," is then input into a bidirectional long short-term memory network for word segmentation. The network learns common word combinations in the private equity fund field, such as "private equity fund," "fundraising amount," "custodian," and "term," and breaks down the text into a sequence of words: "company," "series," "private equity fund," "fundraising," "amount," "5 million," "million yuan," "custodian," "as," "bank," "term," "5," and "year."

[0104] Subsequently, the word sequence was annotated using a part-of-speech tagging tool: "company" was annotated with proper noun PN, "series" with proper noun PN, "private equity fund" with noun phrase NP, "fundraising" with verb V, "amount" with noun N, "500" with quantifier CD, "ten thousand yuan" with quantifier M, "custodian" with noun N, "for" with verb V, "bank" with proper noun PN, "duration" with noun phrase NP, "5" with quantifier CD, and "year" with quantifier M, thus obtaining words with part-of-speech information.

[0105] Step 1032: Using the recognition model that integrates domain dictionaries in natural language processing technology, perform named entity recognition processing on the words with part-of-speech information to obtain words containing entity category tags. Then, perform information integration processing on the words containing entity category tags to obtain associated information groups.

[0106] The domain dictionary refers to a dictionary containing terminology specific to the private equity fund industry. For example, a domain dictionary may include type terms, regulatory terms, and financial terms. This application embodiment does not specifically limit the content and structure of this dictionary; it can be configured according to actual circumstances.

[0107] Step 1032 may specifically include the following steps:

[0108] Step a1: Construct a recognition model for a fusion domain dictionary, the model including a word embedding layer, a feature fusion layer, a context network layer, and a conditional random field layer.

[0109] Among them, the word embedding layer refers to the model input layer, which is used to convert words into fixed-dimensional numerical vectors that can be processed by computers. The feature fusion layer can adopt the feature concatenation method. The context network layer can be set as a 2-layer hidden layer. The conditional random field layer can adopt the linear chain conditional random field algorithm.

[0110] For example, an identification model can be constructed to integrate domain-specific dictionaries, which include type terms such as "private equity fund" and "securities investment fund", regulatory terms such as "custodian" and "duration", and financial terms such as "fundraising amount" and "ten thousand yuan".

[0111] Step a2: Through the word embedding layer, the words with part-of-speech information are divided into paragraphs according to their part-of-speech categories to obtain multiple word segments. All word segments are then vectorized to obtain word vector representations.

[0112] Among them, a word segment refers to a continuous set of words formed by dividing words with part-of-speech information according to their part-of-speech categories; the word vector representation refers to the fixed-dimensional numerical vector that the word embedding layer transforms into a word segment.

[0113] For example, the word embedding layer uses a word vector algorithm to set the vector dimension to 128 dimensions, which can balance semantic expression and computational efficiency. For words with part-of-speech information, such as "company proper noun PN", "private equity fund noun phrase NP", "500 quantifier CD", "ten thousand yuan quantifier M", "bank proper noun PN", "5 quantifier CD", and "annual quantifier M", the word embedding layer divides the words into segments according to their part-of-speech categories. That is, "company proper noun PN + private equity fund noun phrase NP" is divided into segment 1, "500 quantifier CD + ten thousand yuan quantifier M" is divided into segment 2, "bank proper noun PN" is divided into segment 3, and "5 quantifier CD + annual quantifier M" is divided into segment 4. Then, each segment is vectorized, and the word vector algorithm converts each segment into a 128-dimensional word vector representation.

[0114] Step a3: Input the word vector representation into the feature fusion layer. The feature fusion layer performs domain feature matching on the word vector representation using a preset domain dictionary to generate a domain feature vector. Then, it fuses the word vector representation with the domain feature vector to obtain a fused feature vector.

[0115] Among them, the fusion feature vector refers to the vector output by the feature fusion layer that combines general semantics and domain features.

[0116] For example, based on the 128-dimensional word vector representation mentioned above, a semantic feature corresponding to each dimension of the word is input into the feature fusion layer. In the word segment 1, "private equity fund" is matched with the type term in the domain dictionary, thereby generating a 128-dimensional domain feature vector. The dimension of this vector is consistent with the dimension of the word vector. Then, the word vector representation of word segment 1 is fused with the domain feature vector through feature concatenation to obtain a 256-dimensional fused feature vector. Similarly, other word segments are processed to obtain the corresponding fused feature vectors.

[0117] Step a4: Input the fused feature vector into the context network layer to extract context features and obtain a context feature vector; input the context feature vector into the conditional random field layer for sequence labeling to obtain the entity category label corresponding to each word and obtain words containing entity category labels.

[0118] Among them, the context feature vector refers to the vector extracted by the context network layer that contains the semantic relationship between words before and after; the entity category label refers to the label assigned to each word by the conditional random field layer that indicates the type of entity it belongs to.

[0119] In this embodiment of the application, the fused feature vector is input into the context network layer in step a4, and the semantic association between the word segment and the preceding and following content is extracted to obtain the context feature vector; then the context feature vector is input into the conditional random field layer for sequence labeling, and each word is assigned an entity category label to obtain words containing entity category labels. The information of the words containing entity category labels is integrated, and each word is bound to its corresponding part-of-speech information and entity category label to obtain the association information group.

[0120] For example, the fused feature vector is input into the context network layer, and contextual associations are extracted. For instance, the semantic association between "5 million yuan" in word segment 2 and the preceding "fundraising amount" is obtained, resulting in a 128-dimensional context feature vector. The context feature vector is then input into the conditional random field layer and labeled with entity category tags: "company" in word segment 1 is labeled with "manager name", "private equity fund" is labeled with "fund type", word segment 2 is labeled with "fundraising amount", word segment 3 "bank" is labeled with "custodian name", and word segment 4 is labeled with "duration", thus obtaining words containing entity category tags. Finally, the information of each word is integrated to form multiple related information groups such as "company-proper noun PN-manager name", "500-quantifier CD-fundraising amount", and "ten thousand yuan-quantifier M-fundraising amount".

[0121] Step 1033: Arrange all the associated information groups in an orderly manner to obtain structured character feature parameters.

[0122] Among them, the associated information group refers to the structured data unit formed by integrating words containing entity category tags with their corresponding part-of-speech information and entity category tags; the structured character feature parameters refer to the structured data set that contains multiple associated information groups and maintains the semantic order of the original text after being processed in an ordered manner.

[0123] In this embodiment of the application, step 1033 verifies the original order of appearance of all related information groups in the original text data and determines the order of the words corresponding to each related information group in the original text; then, all related information groups are arranged in order according to the original order of appearance to ensure that the arranged set of related information groups is consistent with the semantic logic of the original text; finally, the arranged related information groups are integrated to form structured character feature parameters.

[0124] For example, regarding the related information groups in the company's series of private equity fund raising announcements, first check the order in which each related information group appears in the original text "The company's series of private equity funds raised 5 million yuan, the custodian is a bank, and the term is 5 years". The result is that "company" appears first, followed by "series", "private equity fund", "raising", "amount", "5 million", "ten thousand yuan", "custodian", "is", "bank", "term", "5", "years". Then arrange the related information groups in this order to obtain an ordered set of related information groups. This set is the structured character feature parameter.

[0125] This application utilizes a bidirectional long short-term memory network for word segmentation, which captures contextual semantics more easily than traditional methods, reducing segmentation errors and ensuring that word sequences accurately reflect the original meaning. Furthermore, by integrating a domain dictionary into the model, it addresses the issue of inaccurate terminology recognition in general models for private placement scenarios. Subsequently, a word embedding layer converts text into numerical values, improving entity recognition. The combination of a context network layer and a conditional random field layer reduces mislabeling. Finally, by structurally binding word attributes through associated information groups, the orderly arrangement ensures semantic logic, and the final character feature parameters are well-organized, clearly presenting attributes while preserving textual semantic relationships.

[0126] S104. The character feature parameters are combined and analyzed using the Naive Bayes algorithm to identify key elements in the target announcement information, and the key elements are labeled to obtain the element labeling results.

[0127] In one specific implementation, step S104 includes:

[0128] Step 1041: Divide the character feature parameters according to entity category to obtain multiple entity category groups.

[0129] In this application example, step 1041 identifies the entity category marker of each associated information group in the character feature parameters and sorts out all different entity categories; then, based on the identified entity categories, the associated information groups in the character feature parameters are classified into the corresponding entity categories one by one, thereby ensuring that each associated information group is classified into a group consistent with its entity category marker, and finally multiple entity category groups are obtained.

[0130] For example, the structured character feature parameters for a company's series of private equity fund raising announcements include related information groups: "Company - proper noun PN - manager name", "Series - proper noun PN - fund type", "Private equity fund - noun phrase NP - fund type", "500 - quantifier CD - fundraising amount", "ten thousand yuan - quantifier M - fundraising amount", "Bank - proper noun PN - custodian name", "5 - quantifier CD - duration", "year - quantifier M - duration".

[0131] First, identify the entity category markers for each associated information group to determine the entity category as "Manager Name", "Fund Type", "Fund Raised Amount", "Custodian Name", and "Duration". Then, according to the entity category, group "Company - Proper Noun PN - Manager Name" into the "Manager Name Group", "Series - Proper Noun PN - Fund Type" and "Private Equity Fund - Noun Phrase NP - Fund Type" into the "Fund Type Group", "500 - Quantifier CD - Fund Raised Amount" and "Ten Thousand Yuan - Quantifier M - Fund Raised Amount" into the "Fund Raised Amount Group", "Bank - Proper Noun PN - Custodian Name" into the "Custodian Name Group", and "5 - Quantifier CD - Duration" and "Year - Quantifier M - Duration" into the "Duration" group, resulting in five entity category groups.

[0132] Step 1042: Perform combination analysis on the entity category group using the Naive Bayes algorithm to obtain related words, and perform semantic integration processing on the related words to obtain key elements.

[0133] Among them, the Naive Bayes algorithm is a technical tool used to analyze the probability of association between words. It can determine the degree of association between words by calculating the co-occurrence probability of words in the same entity category.

[0134] Step 1042 may specifically include the following steps:

[0135] Step b1: Extract words from the entity category group to obtain a word set, and perform association analysis on the word set using the Naive Bayes algorithm to obtain association data.

[0136] Among them, the word set refers to the unordered word set formed after extracting all words from the entity category group; the correlation data refers to the numerical result that reflects the degree of correlation between words, calculated by the Naive Bayes algorithm.

[0137] In this embodiment of the application, word extraction processing is performed on each entity category group in step b1 to extract all words from the association information group within the group, thereby obtaining the word set corresponding to each entity category group; then the word set is input into the Naive Bayes algorithm to calculate the co-occurrence probability between words under the corresponding entity category, thereby obtaining the association degree data.

[0138] For example, regarding the "fundraising amount group" in the company's series of private equity fund announcements, this group includes related information groups such as "500-quantifier CD-fundraising amount" and "ten thousand yuan-quantifier M-fundraising amount". The first step is to extract the words to form a word set {500, ten thousand yuan}. Then, the word set is input into a Naive Bayes algorithm to calculate the relevance. Event A is defined as "word 500 appears", event B as "word ten thousand yuan appears", and event C as "belongs to the fundraising amount category". The conditional probability is then calculated using the algorithm, with the formula: ,in, Represents conditional probability. This indicates the probability of both 500 and 10,000 yuan appearing under the fundraising amount category. This indicates the probability of 500 appearing under the fundraising amount category. According to statistics from the sample database of private equity fund announcements to which this announcement belongs, 95% of the samples with 500 appearing under the fundraising amount category also include tens of thousands of yuan. The calculation yields... =0.95. Since 0.95 is greater than the preset threshold of 0.8, it meets the preset correlation standard and obtains the correlation data, that is, 500 and 10,000 yuan have a high correlation.

[0139] Step b2: Filter the correlation data and perform an integrity check on the filtered correlation data to obtain multiple related words.

[0140] Among them, multiple related words refer to the ordered combination of words that are closely related and have complete information after filtering the correlation data.

[0141] In step a3, a correlation threshold is set based on the textual features of the private equity fund field. Then, the correlation data is determined by statistically analyzing the correlation distribution of "effective key information word combinations" in a large number of private equity fund announcement samples. If the correlation data is greater than or equal to the correlation threshold, it means that the combination is a high-frequency effective collocation in the field and is retained. Otherwise, it is removed because it may be a random co-occurrence or semantically irrelevant. This judgment process can ensure that the screened word combinations are both reliable in their correlation and have the basis to form complete key elements.

[0142] Then, in step b2, the relevance data is filtered according to the preset relevance criteria, and word combinations that meet the relevance criteria are retained. The completeness of the filtered word combinations is then checked to ensure that they can fully express the key information, thus obtaining related words.

[0143] For example, the correlation data is further filtered to retain "500-ten thousand yuan". Then, the completeness is checked to confirm that "500" needs to be paired with "ten thousand yuan" to fully express the amount, thus obtaining the related term "500-ten thousand yuan".

[0144] Step b3: Perform sequential verification on all the associated words to obtain an initial word sequence, and perform a coherence check on the initial word sequence to obtain the target word sequence.

[0145] In this embodiment of the application, step b3 checks the order of words in the related words based on the order of appearance of words in the original text, and adjusts them to the same order as the original text to obtain an initial word sequence. Then, the semantic coherence of the initial word sequence is checked to confirm that the word collocation conforms to the language logic, and the target word sequence is obtained.

[0146] For example, check the order of words in the original text. Since "500" comes before "ten thousand yuan", we get the initial word sequence "500-ten thousand yuan". Then check the coherence. Since "5 million yuan" conforms to the logic of expressing monetary amounts, we get the target word sequence "500-ten thousand yuan".

[0147] Step b4: Perform semantic merging processing on the target word sequence to obtain key elements.

[0148] Among them, key elements refer to phrases or sentences that can fully express a certain type of key information by semantically merging the target word sequence. In this application embodiment, the type and specific content of key elements are not specifically limited.

[0149] In this embodiment of the application, the target word sequence is semantically merged in step b4, and the ordered words are combined into a complete phrase to obtain the key elements.

[0150] For example, semantic merging is performed on the target word sequence "500-ten thousand yuan" to obtain the key element "5 million yuan"; similarly, other entity category groups are processed: "manager name group" extracts "company", which has no multi-word association and directly obtains the key element "company"; "fund type group" extracts "series" and "private equity fund", calculates that the conditional probability of the two under the fund type category meets the preset association standard, and merges them into "series private equity fund"; "custodian name group" extracts "bank" to obtain "bank"; "duration group" extracts "5" and "year", calculates the association degree and merges them into "5 years".

[0151] Step 1043: Perform type labeling on the key elements to obtain key elements with type identifiers, and arrange the key elements with type identifiers in sequence to obtain the element labeling results.

[0152] In this embodiment of the application, step 1043 adds a corresponding type identifier to each key element according to the entity category to which it belongs, so as to ensure that the type identifier is consistent with the attribute of the key element, and obtains key elements with type identifiers; then, the order of appearance of each key element with type identifier in the original text is checked, and all key elements with type identifiers are arranged in this order to ensure that the arranged order is consistent with the semantic logic of the original text, and finally all key elements with type identifiers are integrated and arranged to form the element annotation result.

[0153] For example, the key elements of the company's series of private equity fund announcements, including "Company", "Series of Private Equity Funds", "5 million yuan", "Bank", and "5 years", are labeled with the following types: "Company" corresponds to "Manager Name", so it is labeled as "Manager Name: Company"; "Series of Private Equity Funds" corresponds to "Fund Type", so it is labeled as "Fund Type: Series of Private Equity Funds"; "5 million yuan" corresponds to "Amount Raised", so it is labeled as "Amount Raised: 5 million yuan"; "Bank" corresponds to "Custodian Name", so it is labeled as "Custodian Name: Bank"; and "5 years" corresponds to "Duration", so it is labeled as "Duration: 5 years".

[0154] Subsequently, the order of these key elements in the original text "Company series private equity fund raised 5 million yuan, custodian is a bank, duration is 5 years" was checked. It was found that "company" appeared first, followed by "series private equity fund", "5 million yuan", "bank", and "5 years". The key elements with type identifiers were arranged in this order to obtain the element labeling results: [Manager name: Company, Fund type: Series private equity fund, Raised amount: 5 million yuan, Custodian name: Bank, Duration: 5 years].

[0155] This application avoids analytical interference caused by the mixing of different types of key information by dividing character feature parameters according to entity categories and classifying scattered related information groups according to attributes. By focusing on a single type of information in each entity category group, subsequent combination analysis can be carried out on words with the same attribute, thereby improving the focus and accuracy of key element identification. Furthermore, the Naive Bayes algorithm is used to accurately analyze word association, avoiding misassociation of irrelevant words. Subsequently, screening and integrity checks ensure the integrity of related word information, and sequence verification and coherence checks ensure that word combinations conform to the logic and language habits of the original text. The final element annotation results are clear in structure, clear in attributes, and reasonable in order, and can be directly used in scenarios such as regulatory verification, investor access, and data statistics, thereby significantly improving the efficiency of using key information in private equity fund announcements.

[0156] Figure 3This application provides a schematic diagram of a specific implementation of a system for intelligent extraction and annotation of key elements in private equity fund announcements based on natural language processing, as illustrated in the embodiments of this application. Figure 3 The system may include:

[0157] The fusion module 31 is used to scan the target announcement information through a multispectral image acquisition device to obtain document image data at different wavelengths, and to perform fusion processing on the document image data at all wavelengths to obtain a fused image.

[0158] The conversion module 32 is used to separate the text region from the fused image using a support vector machine, and to perform optical character recognition processing on the text region to convert the image text within the text region into machine-readable text data.

[0159] The recognition module 33 is used to perform word segmentation, part-of-speech tagging, and named entity recognition on the text data using natural language processing technology to obtain structured character feature parameters.

[0160] Analysis module 34 is used to perform combined analysis on the character feature parameters using the Naive Bayes algorithm to identify key elements in the target announcement information, and to label the key elements to obtain element labeling results.

[0161] The intelligent extraction and annotation system for key elements of private equity fund announcements based on natural language processing in this application embodiment is used to implement the aforementioned intelligent extraction and annotation method for key elements of private equity fund announcements based on natural language processing. Therefore, the specific implementation of the intelligent extraction and annotation system for key elements of private equity fund announcements based on natural language processing can be found in the embodiment section of the intelligent extraction and annotation method for key elements of private equity fund announcements based on natural language processing mentioned above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0162] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of the above-described method for intelligent extraction and annotation of key elements of private equity fund announcement information based on natural language processing.

[0163] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for intelligent extraction and annotation of key elements of private equity fund announcement information based on natural language processing.

[0164] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0165] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the intelligent extraction and annotation method for key elements of private equity fund announcement information based on natural language processing.

[0166] Those skilled in the art will further 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, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. 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 implementations should not be considered beyond the scope of this invention.

[0167] The foregoing has provided a detailed description of the intelligent extraction and annotation method and system for key elements of private equity fund announcement information based on natural language processing, as provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for intelligent extraction and annotation of key elements in private equity fund announcements based on natural language processing, characterized in that, include: The target announcement information is scanned by a multispectral image acquisition device to obtain document image data at different wavelengths. The document image data at all wavelengths are then fused to obtain a fused image. The text region is separated from the fused image by a support vector machine, and optical character recognition processing is performed on the text region to convert the image text within the text region into machine-readable text data. The text data is segmented, part-of-speech tagging is performed, and named entity recognition is performed using natural language processing techniques to obtain structured character feature parameters; The character feature parameters are combined and analyzed using the Naive Bayes algorithm to identify key elements in the target announcement information. These key elements are then labeled to obtain the element labeling results.

2. The method according to claim 1, characterized in that, The text data is segmented, part-of-speech tagging, and named entity recognition using natural language processing techniques to obtain structured character feature parameters, including: The text data is segmented using a bidirectional long short-term memory network in natural language processing to obtain a word sequence. The word sequence is then labeled with part-of-speech tags to obtain words with part-of-speech information. By using a recognition model that integrates a domain dictionary in natural language processing technology, named entity recognition is performed on the words with part-of-speech information to obtain words containing entity category tags. Information integration processing is then performed on the words containing entity category tags to obtain associated information groups. All the associated information groups are sorted in an orderly manner to obtain structured character feature parameters.

3. The method according to claim 2, characterized in that, The method involves using a recognition model that integrates a domain dictionary in natural language processing to perform named entity recognition on the words containing part-of-speech information, resulting in words with entity category tags, including: A recognition model for a fusion domain dictionary is constructed, the model comprising a word embedding layer, a feature fusion layer, a context network layer, and a conditional random field layer; Through the word embedding layer, the words with part-of-speech information are divided into paragraphs according to their part-of-speech categories to obtain multiple word segments. All word segments are then vectorized to obtain word vector representations. The word vector representation is input into the feature fusion layer. The feature fusion layer performs domain feature matching on the word vector representation using a preset domain dictionary to generate a domain feature vector. The word vector representation and the domain feature vector are then fused to obtain a fused feature vector. The fused feature vector is input into the context network layer for context feature extraction to obtain the context feature vector; The context feature vector is input into the conditional random field layer for sequence labeling to obtain the entity category label corresponding to each word, thus obtaining words containing entity category labels.

4. The method according to claim 1, characterized in that, The method involves combining and analyzing the character feature parameters using a Naive Bayes algorithm to identify key elements in the target announcement information, and then annotating these key elements to obtain the element annotation results, including: The character feature parameters are divided according to entity category to obtain multiple entity category groups; The entity category groups are combined and analyzed using the Naive Bayes algorithm to obtain related words. The related words are then semantically integrated to obtain key elements. The key elements are labeled with type to obtain key elements with type identifiers. The key elements with type identifiers are then arranged in order to obtain the element labeling result.

5. The method according to claim 4, characterized in that, The entity category groups are combined and analyzed using the Naive Bayes algorithm to obtain related words. These related words are then semantically integrated to obtain key elements, including: Word extraction is performed on the entity category group to obtain a word set. The word set is then analyzed for correlation using the Naive Bayes algorithm to obtain correlation data. The correlation data is filtered, and the filtered correlation data is subjected to an integrity check to obtain multiple related words; All the associated words are sequentially checked to obtain an initial word sequence. The initial word sequence is then checked for coherence to obtain the target word sequence. The target word sequence is semantically merged to obtain key elements.

6. The method according to claim 1, characterized in that, The step of separating text regions from the fused image using a support vector machine and performing optical character recognition processing on the text regions to convert the image text within the text regions into machine-readable text data includes: The fused image is divided into regions according to a preset size to obtain multiple sub-images. Feature information extraction is performed on the sub-images to obtain the feature information of each sub-image. The feature information is classified using a support vector machine to obtain a category identifier for each sub-image. Based on the category identifier, sub-images that meet preset conditions are filtered and aggregated to obtain the text region. The image text within the text area is enhanced to obtain enhanced image text, and the enhanced image text is then split to obtain individual character images; Optical character recognition (OCR) is performed on the individual character images to obtain character data corresponding to each individual character image. All the character data are then sequentially combined to obtain machine-readable text data.

7. The method according to claim 1, characterized in that, The process of fusing the document image data across all wavelengths to obtain a fused image includes: The document image data at all wavelengths is divided into regions to obtain the text distribution area and non-text distribution area in each document image data. Image information extraction processing is performed on all the text distribution areas and the non-text distribution areas respectively to obtain multiple text distribution information and multiple non-text distribution information. All the text distribution information is superimposed to obtain a first combined image. Select one of the non-text distribution information as the reference information, and then stitch the remaining non-text distribution information with the reference information to obtain the second combined image; The first combined image and the second combined image are integrated to obtain a fused image.

8. A system for intelligent extraction and annotation of key elements of private equity fund announcement information based on natural language processing, characterized in that, include: The fusion module is used to scan the target announcement information through a multispectral image acquisition device to obtain document image data at different wavelengths, and to fuse the document image data at all wavelengths to obtain a fused image. The conversion module is used to separate the text region from the fused image using a support vector machine, and to perform optical character recognition processing on the text region to convert the image text within the text region into machine-readable text data. The recognition module is used to perform word segmentation, part-of-speech tagging, and named entity recognition on the text data using natural language processing technology to obtain structured character feature parameters; The analysis module is used to perform combined analysis on the character feature parameters using the Naive Bayes algorithm to identify key elements in the target announcement information, and to annotate the key elements to obtain element annotation results.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the intelligent extraction and annotation method for key elements of private equity fund announcement information based on natural language processing as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the intelligent extraction and annotation method for key elements of private equity fund announcement information based on natural language processing as described in any one of claims 1 to 7.

Images