Machine learning driven electronic mail attachment intelligent classification method and system

Abstract

The present application relates to the technical field of machine learning and electronic mail management, and discloses a machine learning driven electronic mail attachment intelligent classification method and system, which comprises the following steps: obtaining an electronic mail attachment original data stream, extracting business keywords through three-stage slicing, calculating a semantic concentration index to obtain semantic features, mapping to a semantic feature collection space, and generating a semantic fingerprint code; obtaining a three-stage area through modeling recognition algorithm, extracting a structured feature collection, calculating an association strength value to construct a fusion feature collection; extracting a context element collection, calculating a total association strength value, setting a hierarchical decision rule to construct a directed association collection. The method solves the problems of difficulty in converting unstructured attachments, insufficient multi-modal understanding, weak context association, low classification accuracy and low query efficiency in electronic mail attachment classification.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and email management technology, and more specifically, to a machine learning-driven intelligent classification method and system for email attachments. Background Technology

[0002] With the deepening of enterprise digital transformation and the popularization of distributed office models, email has become a core hub for internal and external business communication and information transmission. Email attachments carry high-value business data such as contracts, invoices, and reports, serving as a crucial carrier for data access in a distributed environment. However, current email systems only process the main text, relying on manual downloading and archiving of attachments. This often leads to version inconsistencies and inconsistent statuses due to multi-user operations, and also suffers from subjective categorization, time-series breaks, and information silos. Furthermore, static rules based on file extensions and keywords are easily circumvented and cannot resolve status and time-series issues, highlighting significant drawbacks that are key problems currently requiring solutions.

[0003] In the prior art, Chinese patent CN120388388B discloses an automatic identification method and system for correspondence-related regulations. The core process includes file uploading, preprocessing, text recognition, chapter recognition, content extraction, tree structure construction, and result display. It focuses on image preprocessing and text recognition, and relies on hierarchical tree structure construction to achieve full-process automation, providing users with an intuitive display of regulatory information. Chinese patent CN118779459B discloses an email classification method and system based on a multimodal large language model. Its process involves: preprocessing the multimodal text content of the email data to be classified; randomly sampling labeled emails from preset categories as samples; analyzing the samples using a multimodal large language model to obtain category feature text; and then inputting the email text to be classified and the category feature text into the model using prompt word templates to determine the email category. This can achieve spam detection and classification by industry or discipline, meeting the needs of email data governance, improving classification accuracy, and reducing costs.

[0004] However, while the two existing technologies mentioned above have some value in identifying the correlation between letters and regulations and classifying multimodal emails, they fail to solve the core problem of managing electronic letter attachments. Specifically, the patent with authorization announcement number CN120388388B focuses on the correlation between letters and regulations, image preprocessing, and tree structure construction, but does not involve the fusion of multimodal features of attachments, nor does it integrate the three-dimensional context of email body, personnel relationships, and time. It cannot resolve attachment version conflicts in parallel operations by multiple entities, nor does it manage the correctness of the time sequence. The patent with authorization announcement number CN118779459B focuses on email classification, but it does not design a mechanism for consistent attachment status, break down information silos, or manage the time sequence features of attachments. Neither of these technologies has a collaborative attachment status and time sequence management system; either the classification is out of sync with the business scenario, or it causes version confusion, failing to meet the needs of reliable, stable, and orderly attachment management in distributed office environments. Summary of the Invention

[0005] This invention is applicable to electronic mail attachment management scenarios for enterprises of all sizes, such as attachment-intensive business scenarios like financial invoice archiving, HR resume screening, and procurement contract review. It can meet the compatibility management needs of multi-format attachments and multi-departmental collaboration within an enterprise. Through three-level slicing processing, semantic fingerprint encoding, and context element set, it achieves the dual goals of transforming unstructured content of electronic mail attachments into structured semantic representations and improving classification accuracy. By integrating feature matrices with total association strength value calculation, it transforms subjective judgment of attachment classification into quantitative analysis, accurately identifying core and auxiliary semantic features, solving the problems of attachment version confusion and temporal sequence breaks under multi-subject operations, and reducing the misclassification rate. Directed association sets combined with hierarchical decision rules enable on-demand information association, ensuring the matching degree between attachment classification and business scenarios, providing accurate data support for enterprise information retrieval, shortening attachment search time, and reducing the management and maintenance costs of electronic mail attachments.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Machine learning-driven intelligent classification methods for email attachments include:

[0008] Obtain the original data stream of email attachments, perform three-level slicing on the original data stream of email attachments, obtain semantic features based on the three-level slicing, map the semantic features to the semantic feature set space, and obtain semantic fingerprint encoding;

[0009] Semantic fingerprint encoding and three-level slices are identified to obtain three-level regions. A structured feature set is obtained based on the three-level regions. The structured feature set is then transformed to obtain a fused structured feature set.

[0010] Based on the original data stream of the email attachments, a set of contextual elements is extracted, and the total association strength value is calculated from the set of contextual elements. A directed association set is then defined for the total association strength value.

[0011] Furthermore, the third-level slicing includes:

[0012] The first-level slice identifies page separators from the raw data stream of email attachments, divides the data into multiple independent page slices according to page boundaries, and assigns a unique page number to each page slice;

[0013] The second-level slice identifies paragraph terminators for all page slices obtained from the first-level slice, divides them into multiple independent paragraph slices with the terminators as boundaries, assigns a paragraph number to each paragraph slice, and associates each paragraph number with the page number of the page to which it belongs.

[0014] The third-level slicing process identifies sentence terminators in all paragraph slices obtained from the second-level slicing process. It then divides the text into multiple independent sentence slices based on the sentence terminators as boundaries. Each sentence slice is assigned a sentence number according to the text reading order, and each sentence number is associated with the page number of the page to which it belongs and the paragraph number.

[0015] Furthermore, the method for obtaining semantic features based on three-level slicing includes:

[0016] Identify the business keywords in each sentence slice, calculate the semantic concentration index of the business keywords;

[0017] Set a concentration threshold for the semantic concentration index, and compare the semantic concentration index with the concentration threshold to obtain semantic features.

[0018] Furthermore, the method for identifying the semantic fingerprint encoding and the three-level slice to obtain the three-level region includes:

[0019] The target slice is selected from the three-level slice for detection to obtain the number of recognizable characters and the total number of pixels. The number of recognizable characters and the total number of pixels are judged to obtain the text region. The target slice is extracted to obtain the edge pixels. The edge density is obtained based on the edge pixels. The color entropy value is calculated for the target slice. The image region is obtained by combining the edge density and the color entropy value. The target slice is detected to obtain a regular grid. The table region is obtained based on the regular grid.

[0020] The table areas, image areas, and text areas of all target slices are combined into a three-level region.

[0021] Furthermore, the method for obtaining the structured feature set based on the three-level regions includes:

[0022] Text regions in the three-level regions are extracted to obtain a set of text structured features;

[0023] The image regions of the three-level regions are extracted to obtain a set of structured features of the image;

[0024] Extract the table regions from the three-level regions to obtain a set of structured table features;

[0025] The extracted text structured feature set, image structured feature set, and table structured feature set are combined to form a structured feature set.

[0026] Furthermore, the method for transforming the structured feature set to obtain the fused structured feature set includes:

[0027] The structured feature set is normalized to obtain normalized values ​​for text annotation density, numerical cell proportion, and clarity.

[0028] Based on the structured feature set, the association strength value of each pair of text, image and table is calculated, and all association strength values ​​among the three are combined to form a fused structured feature set.

[0029] Furthermore, the method for extracting the set of contextual elements based on the original data stream of email attachments includes:

[0030] The raw data stream of email attachments is identified to obtain a basic set of verbs and an intent classification.

[0031] The business scenario is determined based on the semantic feature set and the text structure feature set. Based on the intent classification and business scenario, the analysis results of the semantic elements of the main text are obtained.

[0032] The personnel identity attributes are obtained from the raw data stream of email attachments, and the results of personnel relationship network element analysis are obtained based on the personnel identity attributes.

[0033] The basic time information of the email is obtained from the raw data stream of the email attachments, and the time pattern feature element analysis results are obtained based on the basic time information.

[0034] The results of semantic element analysis, personnel relationship network element analysis, and time pattern feature element analysis constitute the context element set.

[0035] Furthermore, the method for calculating the total association strength value from the set of context elements includes:

[0036] The intent classification and business scenario are converted into a set representation, and the content relevance is calculated from the set.

[0037] By analyzing personnel identity attributes, we can obtain the specific work scenarios and historical collaboration relationships of personnel, and then determine the relevance of personnel based on these specific scenarios and historical collaboration relationships.

[0038] Based on the business scenario, set the time relevance score to obtain the time relevance.

[0039] The total correlation strength value is obtained by weighting the content relevance, personnel relevance, and time relevance.

[0040] Furthermore, the directed association set includes:

[0041] Directed association sets include strong association scenarios, medium association scenarios, and weak association scenarios;

[0042] In strongly correlated scenarios, when it is necessary to judge elements, if there are conclusions inferred from the context and other single, isolated information, the conclusions inferred from the context shall be accepted.

[0043] In the case of medium relevance scenarios, the conclusions output by context inference and content analysis are compared by weighting based on the total relevance strength value, and the conclusion with the higher weighting ratio is accepted.

[0044] In the case of a weak association, the analysis depends on the content of the attachment.

[0045] A machine learning-driven intelligent classification system for email attachments, used to implement the aforementioned machine learning-driven intelligent classification method for email attachments, the system comprising:

[0046] Semantic fingerprint module: used to obtain the original data stream of email attachments, perform three-level slicing processing on the original data stream of email attachments, obtain semantic features based on the three-level slicing, and map the semantic features to the semantic feature set space to obtain the semantic fingerprint encoding;

[0047] The fusion feature module is used to identify semantic fingerprint encoding and three-level slices to obtain three-level regions. Based on the three-level regions, a structured feature set is obtained, and the structured feature set is transformed to obtain a fused structured feature set.

[0048] Directed Association Module: This module is used to extract the set of context elements from the original data stream of email attachments, calculate the total association strength value from the set of context elements, and define a directed association set for the total association strength value.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention transforms unstructured content in email attachments into structured semantic representations through three-level slicing, semantic fingerprinting, and domain dictionary extraction of business keywords. This addresses the pain points of traditional solutions, such as low efficiency due to manual classification and high misjudgment rates due to static rules relying on file extensions or keywords. By fusing feature matrices, it transforms the fuzzy understanding of multimodal attachments into quantitative association analysis, accurately identifying the core relationships between different modal attachments and addressing the weakness of existing technologies in handling composite attachments such as text, images, and tables. The combination of contextual element sets and weighted calculation of total association strength values ​​enables context-aware classification decisions, resolving issues such as traditional classification ignoring email context, leading to classifications detached from business scenarios and inconsistent operational states among multiple entities. Directed association sets enable on-demand information association and retrieval, breaking down information silos caused by scattered attachment storage, providing precise data support for operations and maintenance, shortening attachment search and fault location time, and reducing enterprise email attachment management and maintenance costs. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 Flowchart of a machine learning-driven intelligent classification method for email attachments provided in an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram illustrating the principle of step S10 in an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram illustrating the principle of step S20 in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram illustrating the principle of step S30 in an embodiment of the present invention;

[0056] Figure 5 A functional block diagram of a machine learning-driven intelligent classification system for email attachments provided in an embodiment of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1:

[0059] Please see Figure 1 As shown, this embodiment provides a machine learning-driven intelligent classification method for email attachments, including:

[0060] Step S10: Obtain the original data stream of email attachments, perform three-level slicing on the original data stream of email attachments, obtain semantic features based on the three-level slicing, map the semantic features to the semantic feature set space, and obtain semantic fingerprint encoding.

[0061] Furthermore, such as Figure 2 As shown, step S10 includes:

[0062] Step S11: Obtain the original data stream of the email attachment and perform three-level slicing processing on the original data stream of the email attachment.

[0063] The original attachment is extracted from the email system. The underlying file interface of the operating system for the original attachment is called to obtain the original data stream of the email attachment. This original data stream includes header information, main content data, and metadata. The header information represents the format and basic attributes of the original attachment; the main content data represents the core parts of the original attachment, such as text content, image data, and table information; and the metadata represents auxiliary data related to the attachment's basic attributes, such as the attachment's creation time, modification time, and size.

[0064] By using the Magic Number tool to analyze the file header information, the true file type of the attachment can be determined, avoiding misjudgments caused by relying solely on the file extension.

[0065] The purpose of using a three-level slicing process to identify attachments of real document types is to transform unstructured attachment content into structured data units. The first-level slicing uses document structure parsing technology to identify page separators from the attachment content's main data, dividing it into multiple independent page slices based on page boundaries. Each page slice is assigned a unique page number, for example, P1 represents the first page. Page boundaries are set according to the document type's margins. The second-level slicing process identifies paragraph termination characters, such as double line breaks and paragraph marks, from all page slices obtained in the first-level slicing. The text is divided into multiple independent paragraph slices based on the terminator. Each paragraph slice is assigned a paragraph number in a left-to-right, top-to-bottom order, for example, S1 represents the first paragraph and is associated with the page number of its parent page, such as P2-S3 representing the third paragraph on the second page. The third-level slices identify sentence terminators (such as periods, question marks, and exclamation marks) from all paragraph slices obtained from the second-level slices and divide them into multiple independent sentence slices. Each sentence slice is assigned a sentence number according to the text reading order, for example, L1 represents the first sentence and is associated with the page number of its parent page and the paragraph number, such as P2-S3-L5 representing the fifth sentence of the third paragraph on the second page.

[0066] Step S12: Based on the three-level slicing, the domain dictionary method is used to extract business keywords from the sentence slices; the business keywords are calculated to obtain the semantic concentration index; the semantic concentration index is judged to obtain the semantic features of the sentence slices.

[0067] The domain dictionary method is used to identify business keywords in each sentence slice of the three-level slice, such as accounts receivable, balance sheet, force majeure, and liability for breach of contract.

[0068] The frequency (F) and distribution density (D) of business keywords in sentence segments are statistically analyzed. Frequency represents the number of times the same business keyword appears across all sentence segments. Distribution density represents the number of consecutive sentence segment groups containing a business keyword; for example, if "force majeure" appears consecutively in the first and second sentences, it forms a consecutive sentence segment group. The semantic concentration index (C) of business keywords is also calculated. The total number of sentence segments represents the total number of sentence segments in a single attachment, while the semantic concentration index of business keywords measures the importance of business keywords in the attachment file.

[0069] Statistical methods were used to collect labeled core and non-core words from historical documents, and a concentration threshold A was set for the semantic concentration index. When, it indicates that the sentence slice corresponding to the business keyword is the core semantic feature, when In this context, the sentence slices corresponding to the business keywords are considered as auxiliary semantic features. The core semantic features represent the main content and purpose of the attachments, such as payment methods; the auxiliary semantic features represent the background information of the attachments, such as company names; and the core semantic features and auxiliary semantic features are combined into a single semantic feature.

[0070] Step S13: Map the semantic features of the sentence slice to the semantic feature set space, and apply the locality-sensitive hashing algorithm to generate semantic fingerprint encoding.

[0071] The semantic feature set space comprises four subspaces: document type subspace, content theme subspace, structure subspace, and style subspace. The document type feature subspace represents the category attributes of the encoded document, such as contracts, reports, and invoices; the content theme subspace represents the specific content of the encoded document, such as financial data and project progress; the structure subspace represents the organizational structure of the encoded document, such as the number of chapters and paragraph distribution; and the style subspace represents the writing style of the encoded document, used to identify the writing context, such as contracts in the legal field.

[0072] The domain dictionary method is used to analyze semantic features to determine attribute types, which are then assigned to subspaces, and mapping values ​​are assigned to each subspace. The mapping values ​​are set based on a semantic concentration index. When a feature is a core semantic feature, the mapping value is the semantic concentration index; when it is an auxiliary semantic feature, the mapping value is the product of the semantic concentration index and a decay coefficient. The decay coefficient is obtained by cross-validating various domains in history to find the value with the highest accuracy and least interference. When multiple semantic features are mapped to the same dimension, a weighted average method is used for judgment. The weights are determined based on metadata, with earlier semantic features having higher weights.

[0073] The semantic feature set space is transformed into a semantic fingerprint code using the Locality Sensitive Hash (LSH) algorithm. The semantic fingerprint code includes document type code, content topic code, structural feature code, and style feature code. The semantic fingerprint code corresponds one-to-one with the four subspaces, enabling partial matching for different needs and improving query efficiency.

[0074] Step S10 solves the current attachment classification problem through three-level slicing, semantic features, semantic feature set space, and semantic fingerprint encoding, realizing the transformation from unstructured attachments to structured semantic representation. Three-level slicing not only locates content but also implies the document's structural information. By analyzing the distribution pattern of identifiers, the document's organization can be inferred. For example, frequent changes in paragraph numbers in the identifiers indicate shorter paragraphs, suggesting a clause-based document such as a contract. The semantic concentration index demonstrates robustness to noise when distinguishing core features; even with a large amount of irrelevant content in the document, the core business content can still be accurately identified. Semantic fingerprint encoding brings significant advantages in storage and transmission.

[0075] Step S20: Identify the semantic fingerprint encoding and the three-level slices to obtain three-level regions. Obtain a structured feature set based on the three-level regions, and transform the structured feature set to obtain a fused structured feature set.

[0076] Furthermore, such as Figure 3 As shown, step S20 includes:

[0077] Step S21: Establish a modality recognition algorithm to identify semantic fingerprint encoding and three-level slices to obtain three-level regions.

[0078] A modality recognition algorithm is established, including text modality recognition, image modality recognition, and table modality recognition. For text modality recognition, the target slice is selected from a three-level slice selection based on accuracy requirements; for example, a page slice is selected if a macroscopic judgment is needed. Optical character recognition (OCR) and document structure analysis techniques are used to detect the target slice, avoiding misclassification of non-text slices as text regions. The number of identifiable characters and the total number of pixels are obtained. The text modality ratio is calculated by dividing the number of identifiable characters by the total number of pixels. Only when the text modality ratio is greater than a text modality threshold is the target slice determined as a text region. The text modality threshold is determined based on the character density data of previously labeled text and non-text regions in historical documents.

[0079] Image modality recognition extracts edge pixels from target slices using edge detection algorithms. Dividing these edge pixels by the total number of pixels in the target slice yields the edge density, reflecting the sparseness of the region's edge distribution. The color entropy value is then calculated using the information entropy formula. Only when the color entropy value is greater than a color threshold and the edge density is less than the edge threshold is the target slice considered an image region, ensuring the exclusion of text and table interference. The edge threshold is set based on the edge density data of labeled image and non-image regions in historical documents, and the color entropy value is set based on the color entropy data of labeled image and non-image regions in historical documents.

[0080] Table modality recognition uses Hough transform to detect straight lines in the horizontal and vertical directions of the target slice, ensuring that only straight lines are considered and eliminating interference from images and text. The straight lines form a regular grid, and the target slice is identified as a table region only when the regular grid is larger than a grid threshold. The grid threshold is set based on the grid structure data of previously labeled table and non-table regions in historical documents.

[0081] The table areas, image areas, and text areas of all target slices are combined into a three-level region.

[0082] Step S22: Extract the three-level regions to obtain a structured feature set.

[0083] Five dimensions are extracted from the text region to form a set of structured text features. The five dimensions include keyword set, sentiment value, density of technical terms, sentence complexity, and paragraph structure.

[0084] The keyword set is extracted from the text region using the TF-IDF algorithm to obtain potential core words. Differential position weights are applied based on the location of these core words, assigning them to headings, first paragraphs, last paragraphs, or other positions according to their ranking from highest to lowest. The TF-IDF value is then calculated, with the word having the highest TF-IDF value representing the core keyword of the text. Sentiment value is obtained by calculating the text region content using a hybrid method of dictionary and rule-based analysis, indicating the emotional attitude of the text region content. Terminology density value is obtained by identifying specialized words in the text region content using a domain dictionary and calculating their density using terminology density measurement, indicating the level of specialization of the text region content. Sentence complexity is calculated from the text region content using syntactic analysis techniques, reflecting the complexity of the sentence structure. Paragraph structure is calculated from independent paragraph slices in the text region using statistical analysis methods, representing a set of features related to the organization of paragraphs in the text region.

[0085] The image structured feature set includes chart-type features, photograph-type features, and icon-type features. Computer vision tools are used to determine the image content of image regions, resulting in image types such as charts, photographs, and icons. Clustering algorithms, contrast formulas, and Laplacian variance are then used to calculate the dominant color tone, contrast, and sharpness of the image type. The dominant color tone represents the core color attribute of the image, contrast represents the degree of difference in brightness, and sharpness represents the sharpness of image details.

[0086] The table structured feature set includes table structure features, data attribute features, and business pattern features. Basic structural information is extracted from table areas using table parsing tools and then broken down. Specifically, table structure features include the number of rows, the number of columns, the row-to-column ratio, and the percentage of merged cells. Data attribute features are obtained through regular expression matching to determine the number and percentage of numeric cells, the number and percentage of text cells, and the distribution of numeric types. Business pattern features are obtained through a domain dictionary to determine keyword matching degree and the total number of rows. Keyword matching degree indicates the closeness of the relationship between the table content and a specific business domain, while the total number of rows reflects the summary analysis attributes of the table data.

[0087] The structured feature set consists of the table structured feature set, the image structured feature set, and the text structured feature set.

[0088] Step S23: The structured feature set is transformed by a normalization method to obtain a fused structured feature set.

[0089] A normalization method is used to map quantified indicators in the structured feature set, such as image contrast and the proportion of numerical cells in a table, to the [0,1] interval, resulting in normalized values ​​for text annotation density, numerical cell proportion, and sharpness. Non-quantified indicators are converted to numerical form using a label encoding method to avoid differences in numerical range. Specifically, the normalized value for text annotation density represents the relative density of text annotations in the image; the closer to 1, the higher the proportion of text annotations in the image; the closer to 0, the lower the proportion of text annotations in the image. The normalized value for numerical cell proportion represents the relative proportion of numerical cells in the table; the closer to 1, the more numerical data is in the table; the closer to 0, the higher the proportion of text description cells in the table. The normalized value for sharpness represents the relative strength of image sharpness; the closer to 1, the sharper the image details; the closer to 0, the blurrier the image.

[0090] Based on a set of structured features, the association strength value of each pair of text, image, and table is calculated.

[0091] The association strength between text and image is determined by transforming the topic word set and image type into a semantic set of fixed-dimensional numerical arrays using a word set model. The cosine of the angle between the two semantic sets is calculated, with the range being [-1, 1]. The smaller or equal to 0 the cosine of the angle, the weaker the semantic association between the two sets. Text annotations in the image region are extracted using optical character recognition technology. The number of pixels in the text annotations is counted against the total number of pixels in the image. The ratio of the number of pixels in the text annotations to the total number of pixels in the image is calculated to obtain the original text annotation density value. The original text annotation density value is then normalized to obtain the normalized text annotation density value. Multiplying the cosine of the angle with the normalized text annotation density value yields the association strength between text and image. Combining the cosine of the angle with the normalized text annotation density value can accurately determine the semantic closeness of the association between text and image, eliminating interference from images without text annotations or with semantic mismatch.

[0092] The association strength value between text and table is calculated by counting the number of numerical pairs in both the text structured feature set and the table structured feature set. For example, if the text structured feature set mentions a cost of 500,000 yuan, and the table structured feature set also shows a cost of 500,000 yuan, that's considered a numerical pair. The ratio of the number of numerical pairs to the total number of text values ​​is then calculated to obtain the numerical pair percentage. A table parsing tool is used to extract the total number of cells and the number of numerical cells in the table area. The ratio of the number of numerical cells to the total number of cells is then calculated to obtain the raw numerical cell percentage. This raw numerical cell percentage is then normalized to obtain the normalized numerical cell percentage. Multiplying this normalized numerical cell percentage by the numerical pair percentage yields the association strength value between text and table. The combination of the normalized numerical cell percentage and the numerical pair percentage effectively verifies the consistency between text and table at the data level, preventing tables with low data percentages or mismatched values ​​from affecting association judgments.

[0093] The association strength between the image and the table is determined by detecting row and column dividing lines in the image type within the image structured feature set using Hough transform, yielding the visual row and column count. This visual row and column count represents the visual row and column distribution of the table screenshot. For example, if the table screenshot visually contains 5 rows and 3 columns, then the visual row and column count is 5 rows and 3 columns. The fit is calculated based on the basic structural information in the table structured feature set, using the following formula: The Laplacian operator is used to calculate the image region, and the variance of the Laplacian operator, i.e., the original sharpness value, is obtained. The original sharpness value is then normalized to obtain the sharpness normalized value. The product of the fit and the sharpness normalized value is the association strength value between the image and the table. The combination of the fit and the sharpness normalized value ensures that the association strength value is calculated only based on visually reliable and structurally matched images and tables, avoiding association deviations caused by blurry images or structural misalignments.

[0094] The correlation strength values ​​among the three are combined to form a fused structured feature set.

[0095] Step S20 addresses the issue of insufficient multi-dimensional understanding of attachment content through modal recognition algorithms, three-level regions, structured feature sets, and fused structured feature sets, achieving the transformation from unstructured attachment layout information to structured fused features. The modal recognition algorithm not only distinguishes between text, image, and table modalities but also accurately eliminates interference from non-target modalities. Threshold judgments ensure accurate region recognition; for example, non-text regions are excluded based on the text modality ratio for text modalities, and text and table interference is excluded based on both color entropy and edge density for image modalities. The structured feature set demonstrates the ability to mine unique modal information during feature extraction; even if attachments contain multiple modalities, the core features of each modality can still be fully extracted. The fused structured feature set improves the efficiency of subsequent classification calculations and the accuracy of association analysis, providing support for a deeper understanding of multimodal attachments.

[0096] Step S30: Extract the context element set based on the original data stream of the email attachments, calculate the total association strength value based on the context element set, and formulate a directed association set for the total association strength value.

[0097] Furthermore, such as Figure 4 As shown, step S30 includes:

[0098] Step S31: Extract the set of contextual elements based on the original data stream of the email attachments.

[0099] Intent verb recognition is performed on the main content data of the original data stream of email attachments to obtain a basic verb set and intent classification. The basic verb set includes universal action verbs for all scenarios and basic processing verbs for cross-scenario applications. Universal action verbs represent verbs of the most basic behaviors, such as submit, send, receive, and view. Basic processing verbs for cross-scenario applications represent verbs that perform fundamental processing functions in multiple business scenarios, such as approve, review, verify, and summarize. Intent classification includes instruction, information, and request categories. Instruction categories indicate actions with clear targets that require execution, such as approve and sign. Information categories indicate actions that involve information transmission without execution, such as refer to and understand. Request categories indicate actions that involve initiating a request that requires a response from others, such as assist and arrange. Data extracted from the original data stream of the email attachments is used to analyze the attachment files based on three elements: semantic elements of the main text, personnel relationship network elements, and time pattern feature elements.

[0100] In the semantic element analysis of the main text, the business scenario is determined by combining the semantic feature set and the text structure feature set. The basic types identified in the document type subspace of the semantic feature set are used to initially match the business scenario; for example, if the basic type is a contract attachment, it is initially matched to a legal signing scenario. After the initial matching, a secondary verification is performed based on the density of technical terms in the text structure feature set. If the technical term density exceeds a preset threshold, the specific business scenario is confirmed, and the scenario is associated with the verbs. The technical term density threshold is set based on the total number of technical terms identified in the attachments. The semantic element analysis of the main text yields the basic verb set, intent classification, and the association between verbs and scenarios.

[0101] In the analysis of personnel relationship network elements, the email domain name of the attached file is extracted by regular expression to determine the organization affiliation. The personnel identity attributes of the organization affiliation are obtained through the organizational structure database query interface, such as personnel department, job level, and reporting relationship information. Personnel collaboration relationships and type tags are generated, such as superior to subordinate, subordinate to superior, peer collaboration, and external communication. The personnel relationship network element analysis obtains personnel identity attributes and personnel collaboration relationships.

[0102] In the time pattern feature element analysis, basic time information of the email is extracted from the metadata of the original data stream of the email attachments. This includes, for example, the sending timestamp, sender address, recipient list, CC list, email subject, body content, and email priority marker. Time dimension characteristics are determined based on predefined time intervals. Business cycles and deadlines are analyzed in conjunction with industry characteristics. Time sensitivity markers are generated by combining time distance and business importance. The predefined time intervals are determined based on business needs. For example, for an e-commerce company's finance department to efficiently process monthly sales data accounting, requiring the completion of the previous month's sales reconciliation and payment statistics by the 5th of the following month, the predefined time interval is set to the 1st to the 5th of each month. The time pattern feature element analysis yields basic time information, predefined time intervals, and time sensitivity markers.

[0103] The data obtained from the analysis of the three elements are combined into a context element set.

[0104] Step S32: Calculate the total association strength value by analyzing the set of context elements.

[0105] Calculate content relevance The intention verbs and business scenarios are converted into set representations. For example, approval and financial reimbursement forms are converted into the set representation [approval-financial scenario]. The set is then converted into a computable, structured vector set using a pre-trained word set model, such as the BERT model. The cosine of the angle between two elements in the vector set is calculated. The cosine of the angle ranges from -1 to 1. A cosine value greater than 0 indicates a higher semantic similarity between the two vectors, while a value less than or equal to 0 indicates a lower semantic similarity.

[0106] Calculate the relevance of personnel By analyzing personnel identity attributes, the specific work scenarios and historical collaboration relationships of personnel are obtained. A base score is assigned based on the relevance of the main text and attachments according to different relationship types. The relevance of different relationship types is categorized from high to low as strong correlation, medium correlation, weak correlation, and no correlation. Strong correlation indicates that the purpose of the main text is achieved through the attachment; for example, the main text contains approval for supplier payments, and the attachment contains a supplier payment application form. Medium correlation indicates that the main text and attachments can be referenced; for example, the main text contains checking budget deviations, and the attachment contains a budget execution comparison table. Weak correlation indicates that the main text and attachments are indirectly related; for example, the main text contains synchronized project documents, and the attachment contains a project team contact list. No correlation indicates that the main text and attachments are unrelated; for example, the main text contains employee training, and the attachment contains customer payment details.

[0107] Calculate time correlation Based on business pattern analysis and through large-scale historical data analysis, scenarios are matched with specific time intervals according to the inherent time patterns of different business scenarios. Simultaneously, a time relevance score is set based on the business urgency of the scenario, ranging from 0 to 1. For example, in a financial scenario, the specific time for the settlement and reimbursement approval period is the last three working days of the month. If a match is found within the last three working days of the month, the match is considered successful; otherwise, the match fails. The business urgency is set based on the time constraints for achieving business goals and the impact of delays.

[0108] The total correlation strength value is calculated using a weighted average. The total correlation strength value ranges from [0,1]. Among them, These are the weighting coefficients for content relevance, personnel relevance, and time relevance, with the weighting coefficients set based on the cost of misjudging each relevance level.

[0109] Step S33: Develop hierarchical decision rules for the total association strength value and construct a directed association set.

[0110] Based on the total association strength value, a directed association set is defined. The directed association graph includes strong association scenarios, medium association scenarios, and weak association scenarios. When the total association strength value is less than the median value within the range, it is considered a weak association scenario; when it is greater than or equal to the median value within the range but less than the strong association threshold, it is considered a medium association scenario; and when it is greater than the strong association threshold, it is considered a strong association scenario. The strong association threshold is set based on industry standards and the risk tolerance of actual business operations.

[0111] In strongly correlated scenarios, when it is necessary to judge elements, if there are conclusions inferred from the context and other single, isolated information, the conclusions inferred from the context should be accepted.

[0112] In scenarios with moderate relevance, the conclusions from context inference and content analysis are compared using a weighted average based on the overall relevance strength value, and the conclusion with the higher weighting ratio is accepted. For context inference, weights are primarily assigned based on the coherence of the scenario and the completeness of the information. If the context includes supplementary information that forms a logical loop, it is assigned a high weight; for example, a coherent process from a purchase request to a supplier's quotation to contract confirmation. For the conclusions from content analysis, weights are assigned based on the semantic consistency of feature matching. For example, if the extracted keywords are VAT or reimbursement items, and the financial statement format matches the obtained scenario, it can be assigned a high weight. If the features are fragmented, such as documents or materials, they are assigned a low weight.

[0113] In weakly related scenarios, the analysis mainly relies on the content of attachments.

[0114] Step S30 addresses the issues of insufficient utilization of contextual information and singular association judgments in attachment classification by using contextual element sets, total association strength values, and directed association sets. This transforms attachment classification from isolated information to multi-dimensional contextual association decisions. The contextual element set not only covers the semantics, personnel collaboration, and temporal characteristics of the attachment content itself, but also constructs a complete business scenario logic through association analysis between elements. For example, by combining intent verbs with business scenarios, the business purpose of attachments can be accurately identified. The total association strength value, through differentiated weighting, effectively balances the influence of content, personnel, and time dimensions. Even if there are deviations in information within a single dimension, weighted calculations can still ensure the accuracy of association judgments. The directed association set, based on the hierarchical decision rules of the total association strength value, provides targeted judgment strategies for different association scenarios, avoiding the pitfalls of a one-size-fits-all classification approach. This significantly improves the accuracy and reliability of attachment classification in complex business scenarios, providing a scientific basis for subsequent intelligent classification decisions.

[0115] Example 2:

[0116] This embodiment, based on Embodiment 1, provides a machine learning-driven intelligent classification system for email attachments, such as... Figure 5As shown, it includes:

[0117] Semantic fingerprint module: used to obtain the original data stream of email attachments, perform three-level slicing processing on the original data stream of email attachments, obtain semantic features based on the three-level slicing, and map the semantic features to the semantic feature set space to obtain the semantic fingerprint encoding;

[0118] The fusion feature module is used to identify semantic fingerprint encoding and three-level slices to obtain three-level regions. Based on the three-level regions, a structured feature set is obtained, and the structured feature set is transformed to obtain a fused structured feature set.

[0119] Directed Association Module: This module is used to extract the set of context elements from the original data stream of email attachments, calculate the total association strength value from the set of context elements, and define a directed association set for the total association strength value.

[0120] In the semantic fingerprint module, the process of acquiring the original data stream of email attachments, performing three-level slicing on the original data stream of email attachments, obtaining semantic features based on the three-level slicing, and mapping the semantic features to a semantic feature set space to obtain semantic fingerprint encoding includes:

[0121] Step S11: Obtain the raw data stream of the email attachments and perform three-level slicing processing on the raw data stream of the email attachments;

[0122] Step S12: Based on the three-level slicing, the domain dictionary method is used to extract business keywords from the sentence slices; the business keywords are calculated to obtain the semantic concentration index; the semantic concentration index is judged to obtain the semantic features of the sentence slices.

[0123] Step S13: Map the semantic features of the sentence slice to the semantic feature set space, and apply the locality-sensitive hashing algorithm to generate semantic fingerprint encoding.

[0124] In the feature fusion module, the semantic fingerprint encoding and three-level slice recognition are performed to obtain three-level regions. A structured feature set is obtained based on the three-level regions. This structured feature set is then transformed to obtain a fused structured feature set, including:

[0125] Step S21: Establish a modality recognition algorithm to identify semantic fingerprint encoding and three-level slices to obtain three-level regions;

[0126] Step S22: Extract the three-level regions to obtain a structured feature set;

[0127] Step S23: The structured feature set is transformed by a normalization method to obtain a fused structured feature set.

[0128] In the directed association module, the step of extracting a set of contextual elements based on the original data stream of email attachments, calculating the total association strength value from the set of contextual elements, and defining a directed association set for the total association strength value includes:

[0129] Step S31: Extract the set of contextual elements based on the original data stream of email attachments;

[0130] Step S32: Calculate the total association strength value by analyzing the set of context elements;

[0131] Step S33: Develop hierarchical decision rules for the total association strength value and construct a directed association set.

Claims

1. A machine learning-driven intelligent classification method for email attachments, characterized in that, The method includes: Obtain the original data stream of email attachments, perform three-level slicing on the original data stream of email attachments, obtain semantic features based on the three-level slicing, map the semantic features to the semantic feature set space, and obtain semantic fingerprint encoding; Semantic fingerprint encoding and three-level slices are identified to obtain three-level regions. A structured feature set is obtained based on the three-level regions. The structured feature set is then transformed to obtain a fused structured feature set. Based on the original data stream of the email attachments, the context element set is extracted, the context element set is calculated to obtain the total association strength value, and a directed association set is defined for the total association strength value; The third-level slice includes: The first-level slice identifies page separators from the raw data stream of email attachments, divides the data into multiple independent page slices according to page boundaries, and assigns a unique page number to each page slice; The second-level slice identifies paragraph terminators for all page slices obtained from the first-level slice, divides them into multiple independent paragraph slices with the terminators as boundaries, assigns a paragraph number to each paragraph slice, and associates each paragraph number with the page number of the page to which it belongs. The third-level slicing process identifies sentence terminators in all paragraph slices obtained from the second-level slicing process. It then divides the text into multiple independent sentence slices based on the sentence terminators as boundaries. Each sentence slice is assigned a sentence number according to the text reading order, and each sentence number is associated with the page number of the page to which it belongs and the paragraph number.

2. The machine learning-driven intelligent classification method for email attachments according to claim 1, characterized in that, The method for obtaining semantic features based on three-level slicing includes: Identify the business keywords in each sentence slice, calculate the semantic concentration index of the business keywords; Set a concentration threshold for the semantic concentration index, and compare the semantic concentration index with the concentration threshold to obtain semantic features.

3. The machine learning-driven intelligent classification method for email attachments according to claim 2, characterized in that, The method for identifying the semantic fingerprint encoding and the three-level slice to obtain the three-level region includes: The target slice is selected from the three-level slice for detection to obtain the number of recognizable characters and the total number of pixels. The number of recognizable characters and the total number of pixels are judged to obtain the text region. The target slice is extracted to obtain the edge pixels. The edge density is obtained based on the edge pixels. The color entropy value is calculated for the target slice. The image region is obtained by combining the edge density and the color entropy value. The target slice is detected to obtain a regular grid. The table region is obtained based on the regular grid. The table areas, image areas, and text areas of all target slices are combined into a three-level region.

4. The machine learning-driven intelligent classification method for email attachments according to claim 3, characterized in that, The method for obtaining the structured feature set based on the three-level regions includes: Text regions in the three-level regions are extracted to obtain a set of text structured features; The image regions of the three-level regions are extracted to obtain a set of structured features of the image; Extract the table regions from the three-level regions to obtain a set of structured table features; The extracted text structured feature set, image structured feature set, and table structured feature set are combined to form a structured feature set.

5. The machine learning-driven intelligent classification method for email attachments according to claim 4, characterized in that, The method for transforming the structured feature set to obtain the fused structured feature set includes: The structured feature set is normalized to obtain normalized values ​​for text annotation density, numerical cell proportion, and clarity. Based on the structured feature set, the association strength value of each pair of text, image and table is calculated, and all association strength values ​​among the three are combined to form a fused structured feature set.

6. The machine learning-driven intelligent classification method for email attachments according to claim 5, characterized in that, The method for extracting the set of contextual elements based on the original data stream of email attachments includes: The raw data stream of email attachments is identified to obtain a basic set of verbs and an intent classification. The business scenario is determined based on the semantic feature set and the text structure feature set. Based on the intent classification and business scenario, the analysis results of the semantic elements of the main text are obtained. The personnel identity attributes are obtained from the raw data stream of email attachments, and the results of personnel relationship network element analysis are obtained based on the personnel identity attributes. The basic time information of the email is obtained from the raw data stream of the email attachments, and the time pattern feature element analysis results are obtained based on the basic time information. The results of semantic element analysis, personnel relationship network element analysis, and time pattern feature element analysis constitute the context element set.

7. The machine learning-driven intelligent classification method for email attachments according to claim 6, characterized in that, The method for calculating the total association strength value from the set of context elements includes: The intent classification and business scenario are converted into a set representation, and the content relevance is calculated from the set. By analyzing personnel identity attributes, we can obtain the specific work scenarios and historical collaboration relationships of personnel, and then determine the relevance of personnel based on these specific scenarios and historical collaboration relationships. Based on the business scenario, set the time relevance score to obtain the time relevance. The total correlation strength value is obtained by weighting the content relevance, personnel relevance, and time relevance.

8. The machine learning-driven intelligent classification method for email attachments according to claim 7, characterized in that, The directed association set includes: Directed association sets include strong association scenarios, medium association scenarios, and weak association scenarios; In strongly correlated scenarios, when it is necessary to judge elements, if there are conclusions inferred from the context and other single, isolated information, the conclusions inferred from the context shall be accepted. In the case of medium relevance scenarios, the conclusions output by context inference and content analysis are compared by weighting based on the total relevance strength value, and the conclusion with the higher weighting ratio is accepted. In the case of a weak association, the analysis depends on the content of the attachment.

9. A machine learning-driven intelligent classification system for email attachments, used to implement the machine learning-driven intelligent classification method for email attachments as described in any one of claims 1-8, characterized in that, The system includes: Semantic fingerprint module: used to obtain the original data stream of email attachments, perform three-level slicing processing on the original data stream of email attachments, obtain semantic features based on the three-level slicing, and map the semantic features to the semantic feature set space to obtain the semantic fingerprint encoding; The fusion feature module is used to identify semantic fingerprint encoding and three-level slices to obtain three-level regions. Based on the three-level regions, a structured feature set is obtained, and the structured feature set is transformed to obtain a fused structured feature set. Directed Association Module: This module is used to extract the set of context elements from the original data stream of email attachments, calculate the total association strength value from the set of context elements, and define a directed association set for the total association strength value.

Images