Salary automatic accounting system based on deep learning

Abstract

The invention belongs to the field of artificial intelligence, and provides a salary automatic accounting system based on deep learning, comprising an input module used for receiving structured data from an enterprise salary management system and unstructured data related to salary; the data processing module is used for performing embedding processing on the structured data to generate structured data feature representation; performing data conversion on the unstructured data to generate unstructured data feature representation; performing two-way information interaction on the structured data feature representation and the unstructured data feature representation by utilizing interactive condition modeling, and dynamically adjusting an information weight related to salary accounting in the unstructured data; the salary accounting module is used for extracting key features related to the salary through a multi-modal fusion network based on the unstructured data feature representation and the information weight related to the salary accounting; and the output module is used for generating a salary accounting report which comprises salary composition details, bonus and deduction explanation.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and in particular relates to an automatic salary calculation system based on deep learning. Background Art

[0002] With the continuous development of enterprise management informatization, salary accounting has become one of the important links in human resource management. The salary accounting process involves a variety of data types such as basic information of employees, working time records, performance evaluation results, labor contract terms, etc. These data are not only complex, but also have the characteristics of diverse formats. For example, structured data (such as working time records, performance scores, etc.) and unstructured data (such as labor contract texts, performance evaluation reports, scanned receipts, etc.) often need to be processed at the same time. However, most traditional salary accounting systems rely on rule-driven and lack the ability to efficiently parse complex unstructured data, resulting in limited accounting efficiency and accuracy.

[0003] Payroll accounting involves the integration of structured and unstructured data. For example, the terms in the labor contract and the descriptive text in the performance evaluation report need to be calculated together with the employee's working time record and performance score, but the existing system often only supports the processing of structured data and has difficulty in parsing the key information in unstructured data. Parsing and extracting rules for unstructured data (such as the text of the labor contract) usually requires manual operation. Especially when dealing with scenarios involving legal terms such as overtime pay rules and deduction policies, manual intervention is prone to inefficiency and high error rates. At the same time, the existing payroll accounting system mostly uses static logic based on predefined rules, lacking dynamic learning and adaptive capabilities. When the company's salary policy or labor regulations change (usually notified in the form of unstructured data), the system needs to be manually adjusted and cannot quickly adapt to complex payroll accounting scenarios. Summary of the invention

[0004] In order to solve the problems in the prior art, the present invention provides a salary automatic calculation system based on deep learning, comprising the following modules:

[0005] An input module, used to receive structured data from the enterprise salary management system and unstructured data related to salary;

[0006] A data processing module is used to embed the structured data to generate a structured data feature representation; perform data conversion on the unstructured data to generate an unstructured data feature representation; use interactive conditional modeling to perform two-way information interaction between the structured data feature representation and the unstructured data feature representation, and dynamically adjust the information weight related to salary accounting in the unstructured data;

[0007] A salary accounting module, for extracting key features related to salary through a multimodal fusion network based on the unstructured data feature representation and the information weights related to salary accounting; and calculating salary details based on the key features and corresponding values ​​in combination with enterprise salary rules and laws and regulations;

[0008] Output module, used to generate salary accounting reports, including salary composition details, bonus and deduction explanations.

[0009] Furthermore, the structured data refers to organized data that is stored in a fixed format and can be directly parsed and processed by algorithms; the unstructured data refers to data types that have no fixed format, diverse content, and require further parsing to extract useful information.

[0010] Furthermore, the embedding process of the structured data to generate a feature representation of the structured data includes: using a multi-layer perceptron network of fixed dimension to extract features of different types of structured data, and normalizing or standardizing the embedded representation thereof.

[0011] Furthermore, the data conversion of the unstructured data to generate unstructured data feature representation includes: textualizing the unstructured data, and then converting the text data into contextual semantic embedding through a pre-trained deep learning model, or converting the data into a feature vector representation through a convolutional neural network.

[0012] Furthermore, the two-way information interaction includes positive attention interaction and reverse attention interaction.

[0013] Furthermore, the positive attention interaction includes:

[0014] Using structured data feature representation S to guide unstructured data feature representation X, highlighting the content related to salary accounting in unstructured data;

[0015]

[0016] Among them, Q X and K X They are the query and key vectors for unstructured data, respectively;

[0017] f(S) and g(S) are conditional vectors generated by structured data and used to adjust Q X and K X The weight of

[0018] d is the dimension of the embedding vector, that is, the length of the vector used to represent the features of the input data;

[0019] T represents matrix transpose;

[0020] softmax() represents the softmax function;

[0021] A X is the generated attention matrix used to dynamically assign the importance of each part in unstructured data.

[0022] Furthermore, the reverse attention interaction includes:

[0023] Use the unstructured data feature representation X to optimize the structured data feature representation S, and dynamically adjust the expression of structured data features;

[0024]

[0025] Among them, Q S and K S They are the query and key vectors of structured data, respectively;

[0026] h(X) and k(X) are conditional vectors generated from unstructured data and used to optimize Q S and K S The weight of

[0027] d is the dimension of the embedding vector, that is, the length of the vector used to represent the features of the input data;

[0028] softmax() represents the softmax function;

[0029] A S is the generated attention matrix used to adjust the structured data features.

[0030] Furthermore, the two-way information interaction further includes:

[0031] Intermodal interaction further strengthens the correlation between the two features by stacking multiple layers of interaction networks:

[0032] The first layer: completes the forward information flow and injects structured data conditions into unstructured data;

[0033] The second layer: completes the reverse information flow and adjusts the unstructured data to act on the structured data;

[0034] The third layer: Generate the final optimized representation by fusing features for subsequent salary calculation tasks.

[0035] Furthermore, the dynamically adjusting the weight of information related to salary accounting in the unstructured data includes:

[0036] The results of the forward and reverse information interactions are used to comprehensively generate the final weights of the unstructured data:

[0037] w i =α·AX [i]+β·A S [j]

[0038] Among them, w i represents the final weight of the i-th part of the unstructured data;

[0039] A X [i] represents the weight of the i-th part of unstructured data in the positive attention;

[0040] A S [j] represents the weight of the reverse attention applied to the structured feature j of the i-th part of the unstructured data;

[0041] α and β represent weight parameters, which are used to balance the impact of forward and reverse information flows.

[0042] Furthermore, the key features related to salary extracted by the multimodal fusion network include:

[0043] Combine features from structured and unstructured data to form a unified data representation;

[0044] Strengthen information that is highly relevant to payroll tasks based on dynamically adjusted unstructured data weights;

[0045] The key features related to the accounting task are extracted based on the multimodal fusion network.

[0046] The present invention provides a deep learning-based automatic salary calculation system, which deeply integrates structured data with unstructured data (such as text, images, and voice) through a multimodal fusion network to extract key features that are highly relevant to salary calculation. Multimodal interaction not only solves the problem that traditional systems are difficult to process unstructured data, but also ensures the semantic relevance between data, greatly improving the processing capabilities of complex salary calculation scenarios.

[0047] The system automatically extracts the information required for salary calculation through dynamic weight adjustment, generates calculation results that comply with corporate salary rules and laws and regulations, and reduces the frequency of manual intervention and the possibility of human error. At the same time, the system can promptly detect and mark abnormal data, further improving the reliability of the results.

[0048] Based on the characteristics of deep learning, the system can adapt to changes in corporate salary rules and laws and regulations through continuous learning. For example, when labor laws or corporate policies are updated, the system can quickly adapt to new accounting needs by adjusting the rule base and retraining some model parameters to ensure compliance and flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 It is a system diagram of the method of the present invention. DETAILED DESCRIPTION

[0051] Below, the invention is preferably described in conjunction with the accompanying drawings and specific implementation methods.

[0052] This embodiment solves the above problem through the following steps:

[0053] In one embodiment, reference Figure 1 The present invention provides an automatic salary accounting system based on deep learning, which comprehensively utilizes enterprise structured data and unstructured data related to salary, and realizes high automation and intelligence of the salary accounting process through multimodal fusion and interactive conditional modeling methods in deep learning models, effectively reduces manual participation, improves accounting efficiency, and ensures the accuracy and compliance of accounting results.

[0054] Specifically, the system includes the following modules:

[0055] The input module is used to receive structured data from the enterprise salary management system and unstructured data related to salary.

[0056] Structured data refers to organized data stored in a fixed format that can be directly parsed and processed by algorithms, including but not limited to the following:

[0057] Basic employee information: such as name, employee number, department, position, time of employment, and salary grade;

[0058] Attendance records: such as daily or monthly working hours, lateness and early departure records, overtime hours, leave types and duration;

[0059] Performance ratings: such as employee annual performance appraisal scores, quarterly or monthly target completion rates, and project delivery indicators;

[0060] Salary changes: such as salary adjustment time and adjustment range, reward records, and deduction records;

[0061] Reimbursement records: such as reimbursement items, specific amounts, etc.

[0062] Other records: such as wages paid, tax deductions, social security payment records, etc.

[0063] For example, an employee's attendance record can be presented in a table format, including date, attendance hours, and overtime hours; performance ratings may be represented by database fields, including project completion rate and rating level.

[0064] Unstructured data refers to data types that are not in a fixed format, have diverse content, and require further analysis to extract useful information, including but not limited to the following:

[0065] Labor contract text: employee salary terms, overtime pay standards, performance reward rules, etc. expressed in natural language. For example, a labor contract in PDF format contains the clause "hourly overtime pay is 1.5 times the normal hourly wage";

[0066] Performance evaluation report: describes the employee's work performance and assessment results over a period of time, such as free text records of work achievements or comments from superiors;

[0067] Invoice image data: such as paper invoices or scanned images, including overtime meal allowances, transportation reimbursements and other salary-related invoices, which usually require optical character recognition (OCR) technology to extract their text information;

[0068] Email content: such as internal email records about salary adjustments and bonus payments, which may contain key information such as the reasons for the adjustment and the specific amount;

[0069] Voice recordings: For example, the recordings of performance appraisal interviews can be used to extract descriptions of employee appraisal results through speech-to-text technology;

[0070] Instant messaging records: such as chat records with supervisors that contain communication text regarding salary adjustments.

[0071] For example, a scanned copy of an overtime application form may contain the employee's name, overtime hours and reason for overtime, and these contents need to be extracted into structured information through image processing technology and natural language analysis.

[0072] The data processing module is used to embed the structured data to generate a structured data feature representation; perform data conversion on the unstructured data to generate an unstructured data feature representation; use interactive conditional modeling to perform two-way information interaction between the structured data feature representation and the unstructured data feature representation, and dynamically adjust the information weight related to salary accounting in the unstructured data.

[0073] The first function of the data processing module is to perform embedding processing on structured data.

[0074] Structured data is usually stored in tables, databases or other regularized formats, including basic information of employees, attendance records, performance scores, etc. This module uses embedding technology to convert these structured data into vectorized feature representations suitable for deep learning model processing, forming structured data feature representations.

[0075] Embedding processing uses a fixed-dimensional multi-layer perceptron (MLP) network or other deep learning methods to extract features from different types of structured data (such as numerical and categorical data) and normalize or standardize their embedded representations.

[0076] For example, overtime hours and leave days in attendance records are embedded into multidimensional vectors, with each dimension representing an attribute, so that they can be fused with unstructured data features in subsequent calculations.

[0077] The second function of the data processing module is to perform data conversion on unstructured data and generate unstructured data feature representation suitable for deep learning model processing.

[0078] For text data (such as labor contracts and performance evaluation reports), we use natural language processing (NLP) technology to perform word segmentation, syntactic analysis, and named entity recognition (NER) to extract key sentences and information related to salary. For example, we can extract a clause such as "overtime pay is 1.5 times the basic salary" from a labor contract.

[0079] For image data (such as scanned receipts and application forms), optical character recognition (OCR) technology is used to convert images into text information and further extract structured fields. For example, the amount, date, and project description can be extracted from the scanned image of an expense report.

[0080] Through speech-to-text technology, voice recordings are converted into text, and combined with NLP technology, information related to salary accounting can be extracted. For example, the decisive content of bonuses or deductions can be extracted from performance interview recordings.

[0081] The text data is converted into contextual semantic embeddings through pre-trained deep learning models (such as BERT, GPT), or the data is converted into feature vector representations through convolutional neural networks (CNN).

[0082] The goal of data transformation is to standardize unstructured data with inconsistent formats and embed them into a unified feature space to form an unstructured data feature representation for joint modeling with the structured data feature representation.

[0083] The third function of the data processing module is to establish a dynamic relationship between structured data feature representation and unstructured data feature representation through interactive conditional modeling, thereby achieving two-way information interaction.

[0084] Interactive conditional modeling constructs a bidirectional information flow model to dynamically interact the feature representation of structured data with the feature representation of unstructured data in a deep learning network.

[0085] In the positive interaction, structured data feature representations (such as employees’ attendance records, performance scores, etc.) are used as conditional information to guide the attention distribution of unstructured data (such as labor contracts and performance reports), highlighting important information related to salary accounting.

[0086] In the reverse interaction, unstructured data (such as the semantic embedding of contract terms) reacts to the feature representation of structured data, dynamically adjusting the feature expression of structured data to more accurately reflect the underlying rules or constraints in the unstructured data.

[0087] Two-way information interaction ensures that the model can extract key features related to salary accounting from multimodal data, thereby improving the accuracy and efficiency of accounting.

[0088] In positive attention, the structured data feature representation S is used to guide the unstructured data feature representation X, highlighting the content related to salary accounting in the unstructured data.

[0089]

[0090] Among them, Q X and K X They are the query and key vectors for unstructured data, respectively;

[0091] f(S) and g(S) are conditional vectors generated by structured data and used to adjust Q X and K X The weight of

[0092] d is the dimension of the embedding vector, that is, the length of the vector used to represent the features of the input data. For structured and unstructured data, the model maps the input data to the same dimension through the feature embedding process for calculation;

[0093] softmax() represents the softmax function;

[0094] A X is the generated attention matrix used to dynamically assign the importance of each part in unstructured data.

[0095] For example, if the structured data shows that an employee "works 20 hours of overtime", the model will dynamically increase the weight of the clause in the labor contract that describes the overtime pay rules (such as "overtime pay is 1.5 times the normal salary").

[0096] If the structured data contains an employee's "performance score is 95 points", the model will prioritize extracting key sentences in the performance report that describe the reasons for the high score or the reward terms.

[0097] In reverse attention, the unstructured data feature representation X is used to optimize the structured data feature representation S, and the expression of structured data features is dynamically adjusted.

[0098]

[0099] Among them, Q S and K S They are the query and key vectors of structured data, respectively;

[0100] h(X) and k(X) are conditional vectors generated from unstructured data and used to optimize Q S and K S The weight of

[0101] d is the dimension of the embedding vector, that is, the length of the vector used to represent the features of the input data;

[0102] softmax() represents the softmax function;

[0103] A S is the generated attention matrix used to adjust the structured data features.

[0104] For example, if the contract text clearly states that "employees with annual performance scores above 90 points can receive additional bonuses", the model will reverse the information of this clause to the structured performance score feature, dynamically increasing the importance of this feature in salary calculation.

[0105] When high traffic subsidies are extracted from the ticket image, the model will dynamically adjust the weights of the structured data fields related to traffic.

[0106] Intermodal interaction further strengthens the correlation between the two features by stacking multiple layers of interaction networks:

[0107] The first layer: completes the forward information flow and injects structured data conditions into unstructured data;

[0108] The second layer: completes the reverse information flow and adjusts the unstructured data to act on the structured data;

[0109] The third layer: Generate the final optimized representation by fusing features for subsequent salary calculation tasks.

[0110] For example, when analyzing a contract text, the first layer focuses on the “bonus rules” clause (positive interaction), the second layer adjusts the weight of the “annual performance score” based on the clause (reverse interaction), and the third layer generates an optimized feature representation to ensure that the rules are consistent with the score.

[0111] The purpose of dynamic weight adjustment is to automatically assign the importance of different contents in unstructured data so that the system can focus on the information related to payroll accounting.

[0112] The results of the forward and reverse information interactions are used to comprehensively generate the final weights of the unstructured data:

[0113] w i =α·A X [i]+β·A S [j]

[0114] Among them, w i represents the final weight of the i-th part of the unstructured data;

[0115] A X [i] represents the weight of the i-th part of unstructured data in the positive attention;

[0116] A S [j] represents the weight of the reverse attention applied to the structured feature j of the i-th part of the unstructured data;

[0117] α and β represent weight parameters, which are used to balance the impact of forward and reverse information flows.

[0118] According to the results of positive interaction, the unstructured information that is strongly related to the characteristics of structured data is dynamically enhanced; according to the results of reverse interaction, the importance of unstructured data to structured data is comprehensively considered to further optimize the weight distribution of unstructured data.

[0119] After the weight adjustment is completed, the feature representation of unstructured data is updated as follows:

[0120]

[0121] Among them, V X [i] is the vector of values ​​for the i-th part of the unstructured data.

[0122] The updated X′ more accurately expresses the information related to the salary accounting task and provides optimized feature input for subsequent modules.

[0123] Example application scenarios:

[0124] Input data:

[0125] Structured data: basic employee information (such as overtime hours of 30 hours), performance score (such as 92 points).

[0126] Unstructured data: Text of labor contracts, including clauses such as “overtime pay is twice the hourly wage” and “employees with performance scores above 90 points receive a 20% bonus.”

[0127] Positive interaction:

[0128] Based on the characteristics of overtime hours, the model increases the weight of overtime clauses.

[0129] Based on the performance score characteristics, the model prioritizes extracting clauses related to bonus rules.

[0130] Reverse interaction:

[0131] Optimize the importance of the performance rating field in structured data based on high performance reward clauses in the contract.

[0132] Dynamic Adjustment:

[0133] By combining the results of positive and reverse interactions, the final weight of the "overtime pay rules" and "bonus rules" clauses in the contract can be improved.

[0134] In this module, through two-way information interaction and dynamic weight adjustment: automatically focus on information related to payroll accounting and filter redundant content; optimize the importance of features to make them more suitable for task requirements; comprehensively consider the semantic relationship between the two data types to provide highly relevant feature input for subsequent payroll accounting.

[0135] The salary accounting module is used to extract key features related to salary through a multimodal fusion network based on the unstructured data feature representation and the information weight related to salary accounting; based on the key features and corresponding values, the salary details are calculated in combination with the enterprise salary rules and laws and regulations.

[0136] This module first combines structured data with unstructured data through a multimodal fusion network to extract key features related to the salary accounting task.

[0137] Input:

[0138] Dynamically adjusted representation of unstructured data features (such as overtime pay clauses in labor contracts and reward rules in performance reports).

[0139] Optimized structured data features (such as basic employee information, overtime hours, and performance scores).

[0140] Key tasks of multimodal fusion:

[0141] Concatenate features from structured and unstructured data into a unified data representation.

[0142] Reinforce information that is highly relevant to payroll tasks based on dynamically adjusted unstructured data weights.

[0143] Feature interaction methods are used to optimize information expression and ensure that multimodal data can complement each other.

[0144] The specific implementation of the multimodal fusion network can use the Transformer architecture in the existing technology, take the feature representations of different modalities as the input sequence, use the aforementioned information weights to adjust its attention mechanism to capture the deep correlation between modalities, and extract key features that are highly relevant (high weight) to the accounting task.

[0145] Through multimodal fusion, the system can extract clear key features, such as:

[0146] Rules for calculating overtime pay in labor contracts.

[0147] Clauses in the performance evaluation report related to bonus payment.

[0148] The conditions for calculating deductions in the company's salary rules.

[0149] After extracting key features, the module completes the itemized calculation of salary details based on the company's salary rules and laws and regulations, combined with specific data values.

[0150] Calculation content:

[0151] Base salary: directly read from structured data or calculated by algorithm.

[0152] Overtime pay: calculated based on the employee's overtime hours, wage standards and the overtime coefficient stipulated by the company.

[0153] Performance bonus: calculated based on performance rating and bonus payment rules.

[0154] Deductions: calculated based on the proportion of unfulfilled targets and deduction rules.

[0155] Calculation logic:

[0156] The system matches the corresponding corporate salary rules or legal clauses from the features and applies them to the relevant data values. For example, for an employee who works 30 hours of overtime, the matching rule may be "1.5 times the hourly wage" and the overtime pay is calculated based on this.

[0157] If an abnormal value is detected (such as an abnormal bonus amount or negative deduction), the module will mark the result and prompt for review.

[0158] In this module, the system can accurately extract the key information required for salary calculation tasks by dynamically adjusting the weight of unstructured data, ensuring the reliability and explainability of each part of the calculation results. Through the deep integration of structured and unstructured data, the automatic calculation of salary details in complex scenarios can be realized, greatly improving efficiency and accuracy.

[0159] Output module, used to generate salary accounting reports, including salary composition details, bonus and deduction explanations.

[0160] The output module of this module organizes the calculation results of the salary accounting module into a structured report to ensure complete content and clear logic, which is easy for employees or managers to understand and consult.

[0161] The generated report includes

[0162] Salary composition details:

[0163] Basic salary.

[0164] Overtime pay.

[0165] Performance bonus.

[0166] Other allowances (such as transportation subsidies, catering subsidies).

[0167] Deduction items (such as deductions for failure to achieve targets, fines for lateness).

[0168] Bonuses and deductions explained:

[0169] The source rules, calculation basis and corresponding value of each bonus or deduction.

[0170] References to clauses (e.g. employment contracts, corporate policies).

[0171] A brief description of the calculation process.

[0172] Furthermore, the output module can visualize the salary data to help users intuitively understand the salary composition and changing trends.

[0173] Furthermore, the output module supports exporting payroll reports into multiple file formats to suit different usage scenarios and needs.

[0174] Furthermore, for abnormal values ​​or special cases in the accounting process, the module will highlight them in the report and provide additional explanations, such as the type of abnormal item (such as excessive bonuses or negative deductions). Possible reasons for the abnormality (such as mismatched terms or data entry errors). Specific terms or data items that are recommended for review.

[0175] The prior art mentioned in the above background technology section and specific embodiment section of the present invention can be used as part of the present invention to understand the meaning of some technical features or parameters.

Claims

1. A deep learning-based automatic salary calculation system, characterized in that: The system includes the following modules: An input module, used to receive structured data from the enterprise salary management system and unstructured data related to salary; A data processing module, used for embedding the structured data to generate a feature representation of the structured data; Performing data conversion on the unstructured data to generate unstructured data feature representation; By using interactive conditional modeling, bidirectional information interaction is performed on the structured data feature representation and the unstructured data feature representation, and the weight of information related to salary accounting in the unstructured data is dynamically adjusted; A salary accounting module, for extracting key features related to salary through a multimodal fusion network based on the unstructured data feature representation and the information weights related to salary accounting; and calculating salary details based on the key features and corresponding values ​​in combination with enterprise salary rules and laws and regulations; Output module, used to generate salary accounting reports, including salary composition details, bonus and deduction explanations.

2. The deep learning-based automatic salary calculation system according to claim 1 is characterized in that: The structured data refers to organized data that is stored in a fixed format and can be directly parsed and processed by algorithms; the unstructured data refers to data types that have no fixed format, diverse content, and require further parsing to extract useful information.

3. The deep learning-based automatic salary calculation system according to claim 1 is characterized in that: The embedding process of the structured data to generate a feature representation of the structured data includes: using a multi-layer perceptron network of fixed dimension to extract features of different types of structured data, and normalizing or standardizing the embedded representation thereof.

4. The deep learning-based automatic salary calculation system according to claim 1 is characterized in that: The step of performing data conversion on the unstructured data to generate unstructured data feature representation includes: converting the unstructured data into text, and then converting the text data into contextual semantic embedding through a pre-trained deep learning model, or converting the data into a feature vector representation through a convolutional neural network.

5. The deep learning-based automatic salary calculation system according to claim 1 is characterized in that: The two-way information interaction includes positive attention interaction and reverse attention interaction.

6. The deep learning-based automatic salary calculation system according to claim 5 is characterized in that: The positive attention interaction includes: Using structured data feature representation S to guide unstructured data feature representation X, highlighting the content related to salary accounting in unstructured data; Among them, Q X and K X They are the query and key vectors for unstructured data, respectively; f(S) and g(S) are conditional vectors generated by structured data and used to adjust Q X and K X The weight of d is the dimension of the embedding vector, that is, the length of the vector used to represent the features of the input data; T represents matrix transpose; softmax() represents the softmax function; A X is the generated attention matrix used to dynamically assign the importance of each part in unstructured data.

7. The deep learning-based automatic salary calculation system according to claim 6 is characterized in that: The reverse attention interaction includes: Use the unstructured data feature representation X to optimize the structured data feature representation S, and dynamically adjust the expression of structured data features; Among them, Q S and K S They are the query and key vectors of structured data, respectively; h(X) and k(X) are conditional vectors generated from unstructured data and used to optimize Q S and K S The weight of d is the dimension of the embedding vector, that is, the length of the vector used to represent the features of the input data; softmax() represents the softmax function; A S is the generated attention matrix used to adjust the structured data features.

8. The deep learning-based automatic salary calculation system according to claim 7 is characterized in that: The two-way information interaction further includes: Intermodal interaction further strengthens the correlation between the two features by stacking multiple layers of interaction networks: The first layer: completes the forward information flow and injects structured data conditions into unstructured data; The second layer: completes the reverse information flow and adjusts the unstructured data to act on the structured data; The third layer: Generate the final optimized representation by fusing features for subsequent salary calculation tasks.

9. The deep learning-based automatic salary calculation system according to claim 8 is characterized in that: The dynamically adjusting the information weight related to salary calculation in the unstructured data includes: The results of the forward and reverse information interactions are used to comprehensively generate the final weights of the unstructured data: w i =α·A X [i]+β·A S [j] Among them, w i represents the final weight of the i-th part of the unstructured data; A X [i] represents the weight of the i-th part of unstructured data in the positive attention; A S [j] represents the weight of the reverse attention applied to the structured feature j of the i-th part of the unstructured data; α and β represent weight parameters, which are used to balance the impact of forward and reverse information flows.

10. The deep learning-based automatic salary calculation system according to claim 1 is characterized in that: The key features related to salary extracted by the multimodal fusion network include: Combine features from structured and unstructured data to form a unified data representation; Strengthen information that is highly relevant to payroll tasks based on dynamically adjusted unstructured data weights; The key features related to the accounting task are extracted based on the multimodal fusion network.

Images