Financial reconciliation intelligent processing method and system based on neural network

Abstract

According to the financial reconciliation intelligent processing method and system based on the neural network, the features of the reconciliation task description data can be accurately extracted by using the feature extraction network in the neural network model, the reconciliation task description vector is generated, and accurate basic data is provided for subsequent processing. By performing feature extraction on the to-be-learned financial reconciliation guide data and combining the reconciliation task description vector and the template session flow graph convolution features, the model can effectively learn key knowledge points of financial reconciliation, and the reconciliation accuracy is improved. Through feature distance determination and target training cost calculation, accurate matching of the sample financial reconciliation guiding features and the candidate matching features is realized, then financial reconciliation guiding knowledge points in the financial reconciliation guiding data to be learned are optimized according to the target training cost, and the reconciliation performance of the neural network model is continuously improved. Therefore, the intelligent level and accuracy of financial reconciliation can be remarkably improved, the manual operation cost is reduced, and the working efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method and system for intelligent processing of financial reconciliation based on neural networks. Background Art

[0002] In the current financial reconciliation process, traditional methods often rely on a large amount of manual operations and complex rule settings, which is not only inefficient but also prone to errors. With the expansion of enterprise scale and the complexity of business, traditional financial reconciliation methods can no longer meet the needs of modern enterprises. Therefore, it is particularly important to develop an intelligent and automated financial reconciliation processing method. Summary of the invention

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a method for intelligent processing of financial reconciliation based on a neural network, the method comprising: Determine the reconciliation task description data corresponding to the template financial reconciliation session flow; Using the feature extraction network in the neural network model, feature extraction is performed on the reconciliation task description data to generate a reconciliation task description vector; using the feature extraction network, feature extraction is performed on the financial reconciliation guidance data to be learned to generate sample financial reconciliation guidance features, wherein the financial reconciliation guidance data to be learned includes the financial reconciliation guidance knowledge points to be learned; Using the graph convolution unit in the neural network model, extracting features from the template financial reconciliation conversation flow to generate graph convolution features of the template conversation flow; Determine the feature distance based on the sample financial reconciliation guide feature and the candidate matching feature, and determine the target training cost based on the feature distance result; the candidate matching feature is determined based on the reconciliation task description vector and the template conversation flow graph convolution feature; The financial reconciliation guidance knowledge points in the to-be-learned financial reconciliation guidance data are optimized according to the target training cost.

[0004] In a possible implementation of the first aspect, the to-be-learned financial reconciliation guidance data includes first financial reconciliation guidance data, wherein the first financial reconciliation guidance data includes first financial reconciliation guidance knowledge points to be learned and each reference financial reconciliation label information in a financial reconciliation branch; the sample financial reconciliation guidance feature includes a first sample financial reconciliation guidance feature; The determining of the feature distance based on the sample financial reconciliation guide feature and the candidate matching feature, and determining the target training cost based on the feature distance result, includes: Using a feature interaction network, perform feature interaction aggregation on the account reconciliation task description vector and the template conversation flow graph convolutional features to generate a first candidate matching feature; Determine the feature distance based on the first sample financial reconciliation guidance feature and the first candidate matching feature to generate a first feature distance; the first feature distance reflects the deviation between the template financial reconciliation conversation flow and each reference financial reconciliation tag information in the financial reconciliation branch; A first target training cost is determined based on the first feature distance and a reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch.

[0005] In a possible implementation of the first aspect, the using a feature interaction network to perform feature interaction aggregation on the reconciliation task description vector and the template conversation flow graph convolution feature to generate a first candidate matching feature includes: Determine financial reconciliation docking features by using the interactive attention unit in the feature interaction network based on the reconciliation task description vector as a search vector and the template conversation flow graph convolution features as indexes and contents; The self-attention unit in the feature interaction network is used to perform self-attention processing on the financial reconciliation docking feature to generate the first candidate matching feature.

[0006] In a possible implementation of the first aspect, optimizing the financial reconciliation guidance knowledge points in the to-be-learned financial reconciliation guidance data according to the target training cost includes: optimizing the first financial reconciliation guidance knowledge points in the first financial reconciliation guidance data according to the first target training cost; The method also includes: optimizing parameter information of the feature interaction network according to the first target training cost.

[0007] In a possible implementation of the first aspect, the financial reconciliation guidance data to be learned includes second financial reconciliation guidance data and third financial reconciliation guidance data, the second financial reconciliation guidance data includes second financial reconciliation guidance knowledge points to be learned and reconciliation task description knowledge points extracted from the reconciliation task description data, and the third financial reconciliation guidance data includes the second financial reconciliation guidance knowledge points and each reference financial reconciliation label information in a financial reconciliation branch; The method of using the feature extraction network to extract features from the financial reconciliation guidance data to be learned and generating sample financial reconciliation guidance features includes: The feature extraction network is used to extract features from the second financial reconciliation guidance data and the third financial reconciliation guidance data, respectively, to generate second sample financial reconciliation guidance features corresponding to the second financial reconciliation guidance data and third sample financial reconciliation guidance features corresponding to the third financial reconciliation guidance data.

[0008] In a possible implementation of the first aspect, determining the feature distance based on the sample financial reconciliation guiding feature and the candidate matching feature, and determining the target training cost based on the feature distance result includes: Extracting a member vector reflecting the overall semantic feature from the reconciliation task description vector as the second candidate matching feature; extracting a member vector reflecting the overall financial reconciliation session semantic feature from the template conversation flow graph convolution feature as the third candidate matching feature; Determine a feature distance based on the second sample financial reconciliation guide feature and the second candidate matching feature to generate a second feature distance, and determine a first unit training cost based on the second feature distance; Determine a feature distance based on the third sample financial reconciliation guiding feature and the second candidate matching feature to generate a third feature distance, and determine a second unit training cost based on the third feature distance and a reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch; Determine a feature distance based on the third sample financial reconciliation guidance feature and the third candidate matching feature to generate a fourth feature distance, and determine a third unit training cost based on the fourth feature distance and a reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch; A second target training cost is determined based on the first unit training cost, the second unit training cost, and the third unit training cost.

[0009] In a possible implementation of the first aspect, optimizing the financial reconciliation guidance knowledge points in the to-be-learned financial reconciliation guidance data according to the target training cost includes: The second financial reconciliation guidance knowledge points in the second financial reconciliation guidance data and the third financial reconciliation guidance data are optimized according to the second target training cost.

[0010] In a possible implementation of the first aspect, the method further includes: Get the input target financial reconciliation session flow; Using the graph convolution unit in the neural network model, extracting features of the target financial reconciliation conversation flow to generate graph convolution features of the target conversation flow; The feature distance is determined based on the target conversation flow graph convolution features and the target financial reconciliation guidance features, and based on the feature distance results, the financial reconciliation label data corresponding to the target financial reconciliation conversation flow is determined, wherein the target financial reconciliation guidance features are generated by extracting features from the target financial reconciliation guidance data using the feature extraction network in the neural network model, and the target financial reconciliation guidance data includes the optimized financial reconciliation guidance knowledge points and each reference financial reconciliation label information in the financial reconciliation branch.

[0011] In a possible implementation of the first aspect, the target financial reconciliation guidance feature includes a first target financial reconciliation guidance feature and a second target financial reconciliation guidance feature, the first target financial reconciliation guidance feature corresponds to first target financial reconciliation guidance data including a first financial reconciliation guidance knowledge point, the second target financial reconciliation guidance feature corresponds to second target financial reconciliation guidance data including a second financial reconciliation guidance knowledge point, and the first financial reconciliation guidance knowledge point and the second financial reconciliation guidance knowledge point are obtained in different ways; The determining of the feature distance based on the target conversation flow graph convolution feature and the target financial reconciliation guidance feature, and determining the financial reconciliation label data corresponding to the target financial reconciliation conversation flow based on the feature distance result, includes: Determine a feature distance based on the target conversation flow graph convolution feature and the first target financial reconciliation guidance feature to generate a fifth feature distance result; Determine a feature distance based on the target conversation flow graph convolution feature and the second target financial reconciliation guidance feature to generate a sixth feature distance result; Determining the financial reconciliation tag data based on the fifth characteristic distance result and the sixth characteristic distance result; Among them, the first financial reconciliation guidance knowledge point is generated by learning based on the feature interaction fusion features of the reconciliation task description vector with mapping relationship and the template conversation flow graph convolution feature, and the second financial reconciliation guidance knowledge point is generated by deriving the first financial reconciliation guidance knowledge point using the first derivation strategy.

[0012] On the other hand, an embodiment of the present application also provides a neural network-based intelligent financial reconciliation processing system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0013] Based on the above aspects, the embodiment of the present application makes the financial reconciliation process more intelligent, reduces manual intervention, and improves work efficiency through the application of a neural network model. By utilizing the feature extraction network in the neural network model, the features of the reconciliation task description data can be accurately extracted, and a reconciliation task description vector can be generated to provide accurate basic data for subsequent processing. By extracting features from the financial reconciliation guidance data to be learned, and combining the reconciliation task description vector and the template conversation flow graph convolution features, the model can effectively learn the key knowledge points of financial reconciliation and improve the accuracy of reconciliation. By determining the feature distance and calculating the target training cost, the accurate matching of the sample financial reconciliation guidance features and the candidate matching features is achieved, and then the financial reconciliation guidance knowledge points in the financial reconciliation guidance data to be learned are optimized according to the target training cost, and the reconciliation performance of the neural network model is continuously improved. As a result, the intelligence level and accuracy of financial reconciliation can be significantly improved, the cost of manual operation can be reduced, and work efficiency can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the execution flow of the neural network-based intelligent processing method for financial reconciliation provided in an embodiment of the present application.

[0015] Figure 2 It is a schematic diagram of the hardware architecture of the neural network-based intelligent processing system for financial reconciliation provided in an embodiment of the present application. DETAILED DESCRIPTION

[0016] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a method for intelligent processing of financial reconciliation based on a neural network provided by an embodiment of the present application. The method for intelligent processing of financial reconciliation based on a neural network is introduced in detail below.

[0017] Step S110, determining the reconciliation task description data corresponding to the template financial reconciliation session flow.

[0018] In this embodiment, the template financial reconciliation conversation flow is first retrieved from the database. These template financial reconciliation conversation flows are usually previously defined and represent a series of standard financial reconciliation communication processes. The server then determines the reconciliation task description data corresponding to these template financial reconciliation conversation flows. The reconciliation task description data may include key information such as the specific objectives of the reconciliation, the account information involved, and the time range of the reconciliation.

[0019] For example, a reconciliation task description data can be: "Reconcile the sales data for the first quarter of 2023 with the records provided by the supplier to ensure that the records of both parties are consistent." This description provides a clear context for the subsequent feature extraction.

[0020] Step S120, using the feature extraction network in the neural network model, extracting features from the reconciliation task description data to generate a reconciliation task description vector. Using the feature extraction network, extracting features from the financial reconciliation guidance data to be learned to generate sample financial reconciliation guidance features, wherein the financial reconciliation guidance data to be learned includes the financial reconciliation guidance knowledge points to be learned.

[0021] In this embodiment, the server will use the feature extraction network in the neural network model to extract features from the above-mentioned reconciliation task description data. The feature extraction network can be a deep learning model based on natural language processing, such as a recurrent neural network (RNN) or a Transformer model, which can effectively capture the semantic information in text data.

[0022] Continuing with the above example, the feature extraction network will analyze the reconciliation task description data, extract key information such as "first quarter of 2023", "sales data", "supplier records", etc., and encode this information into a high-dimensional vector - the reconciliation task description vector, which will serve as the basis for subsequent processing.

[0023] At the same time, the server will also use this feature extraction network to process the financial reconciliation guidance data to be learned. These financial reconciliation guidance data contain the financial reconciliation guidance knowledge points that need to be learned, such as how to identify abnormal situations in the reconciliation process, how to communicate effectively, etc. Through feature extraction, the server also converts these financial reconciliation guidance knowledge points into high-dimensional sample financial reconciliation guidance features.

[0024] Step S130, using the graph convolution unit in the neural network model, extracting features from the template financial reconciliation conversation flow to generate graph convolution features of the template conversation flow.

[0025] In this embodiment, the server can use the graph convolution unit in the neural network model to process the template financial reconciliation session flow. A graph convolution network (GCN) is a neural network that can process graph structured data and can effectively capture the relationship and features between nodes.

[0026] In this scenario, the template financial reconciliation conversation flow can be viewed as a graph structure, where nodes represent different conversation turns or events, and edges represent the order or logical relationship between them. The graph convolution unit analyzes this graph structure, extracts the features of the entire conversation flow, and generates the template conversation flow graph convolution features.

[0027] For example, a template financial reconciliation conversation flow may include multiple links such as opening remarks, data verification, exception handling, and confirmation results. The graph convolution unit captures the associations and transformations between these links, as well as the features within each link, thereby generating a comprehensive graph convolution feature representation.

[0028] Step S140, determining a feature distance based on the sample financial reconciliation guide feature and the candidate matching feature, and determining a target training cost based on the feature distance result. The candidate matching feature is determined based on the reconciliation task description vector and the template conversation flow graph convolution feature.

[0029] In this embodiment, the server will determine the feature distance based on the previously extracted sample financial reconciliation guidance features and candidate matching features. The candidate matching features are determined based on the reconciliation task description vector and the template conversation flow graph convolution features, which can be achieved through feature fusion, attention mechanism and other methods.

[0030] Specifically, the server may calculate indicators such as cosine similarity and Euclidean distance between the sample financial reconciliation guiding features and the candidate matching features to measure the similarity or difference between them, which will serve as the basis for determining the target training cost.

[0031] For example, if a sample financial reconciliation guide feature is very similar to a candidate matching feature, it means that the guide knowledge point is highly relevant to the current reconciliation task and conversation flow, and therefore should be given more attention during training. On the contrary, if the two are very different, the representation or weight of the guide knowledge point may need to be adjusted.

[0032] Step S150: optimizing the financial reconciliation guidance knowledge points in the to-be-learned financial reconciliation guidance data according to the target training cost.

[0033] In this embodiment, the server will optimize the financial reconciliation guidance knowledge points in the financial reconciliation guidance data to be learned according to the target training cost calculated above, which usually involves a process of adjusting the parameters of the neural network model to minimize the training cost.

[0034] The optimization algorithm can be a common deep learning optimization algorithm such as gradient descent and Adam. Through continuous iteration and optimization, the server can gradually improve the accuracy and effectiveness of the financial reconciliation guidance knowledge points, making it better adapted to different reconciliation tasks and conversation scenarios.

[0035] For example, if a certain guiding knowledge point is found to perform poorly in a specific type of reconciliation task (i.e., the corresponding training cost is high), the server may adjust the representation of the knowledge point, add relevant training data, or adjust the model structure to improve its performance. After multiple iterations and optimizations, these guiding knowledge points will be able to better assist the server in performing efficient financial reconciliation work.

[0036] Based on the above steps, the embodiment of the present application makes the financial reconciliation process more intelligent, reduces manual intervention, and improves work efficiency through the application of a neural network model. By utilizing the feature extraction network in the neural network model, the features of the reconciliation task description data can be accurately extracted, and a reconciliation task description vector can be generated to provide accurate basic data for subsequent processing. By extracting features from the financial reconciliation guidance data to be learned, and combining the reconciliation task description vector and the template conversation flow graph convolution features, the model can effectively learn the key knowledge points of financial reconciliation and improve the accuracy of reconciliation. By determining the feature distance and calculating the target training cost, the accurate matching of the sample financial reconciliation guidance features and the candidate matching features is achieved, and then the financial reconciliation guidance knowledge points in the financial reconciliation guidance data to be learned are optimized according to the target training cost, and the reconciliation performance of the neural network model is continuously improved. As a result, the intelligence level and accuracy of financial reconciliation can be significantly improved, the cost of manual operation can be reduced, and work efficiency can be improved.

[0037] In a possible implementation, the financial reconciliation guidance data to be learned includes first financial reconciliation guidance data, wherein the first financial reconciliation guidance data includes first financial reconciliation guidance knowledge points to be learned and each reference financial reconciliation tag information in a financial reconciliation branch. The sample financial reconciliation guidance feature includes a first sample financial reconciliation guidance feature.

[0038] Step S140 includes: Step S141, using a feature interaction network, performs feature interaction aggregation on the reconciliation task description vector and the template conversation flow graph convolutional features to generate a first candidate matching feature.

[0039] Step S142: determining a feature distance based on the first sample financial reconciliation guidance feature and the first candidate matching feature to generate a first feature distance. The first feature distance reflects the deviation between the template financial reconciliation conversation flow and each reference financial reconciliation tag information in the financial reconciliation branch.

[0040] Step S143: determining a first target training cost based on the first feature distance and a reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch.

[0041] In this embodiment, the first financial reconciliation guidance data includes the first financial reconciliation guidance knowledge points to be learned and each reference financial reconciliation label information in the financial reconciliation branch. For example, a first financial reconciliation guidance data may include a knowledge point such as how to identify and handle differences in reconciliation, and mark the specific position of this knowledge point in the financial reconciliation process, such as the label of "difference handling branch".

[0042] In detail, the server extracted a typical case from the historical reconciliation records, which involved the process of difference processing. The case was marked as the "difference processing branch" and included specific steps and techniques on how to communicate effectively to resolve account differences. These data were organized into the first financial reconciliation guidance data for subsequent learning.

[0043] The server will then use a feature interaction network to perform feature interaction aggregation on the reconciliation task description vector and the template conversation flow graph convolutional features to generate the first candidate matching feature. This process is to find a comprehensive representation that can reflect the reconciliation task requirements and conversation flow characteristics.

[0044] For example, the server loads a pre-trained feature interaction network and inputs the reconciliation task description vector (such as "check the sales data for the first quarter") and the template conversation flow graph convolutional features (such as a standard difference processing dialogue process). Through the feature interaction aggregation operation of the feature interaction network, the server obtains a first candidate matching feature that combines task requirements and conversation features.

[0045] After obtaining the first candidate matching feature, the server can determine the feature distance based on the first sample financial reconciliation guidance feature and the candidate matching feature to generate a first feature distance, which reflects the deviation between the template financial reconciliation session flow and each reference financial reconciliation label information in the financial reconciliation branch.

[0046] For example, the server calculates the cosine similarity between the first sample financial reconciliation guidance feature (such as features on how to identify and handle reconciliation differences) and the first candidate matching feature, and obtains a specific value, namely, the first feature distance. The smaller the distance value, the more the guidance knowledge point matches the current reconciliation task and conversation flow.

[0047] Based on the first feature distance and the reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch, the server can determine a first target training cost, and the cost function will be used to optimize the parameters of the neural network to better learn and adapt to different reconciliation tasks and conversation scenarios.

[0048] For example, based on the calculated first feature distance and known reference financial reconciliation labels (such as "difference processing branch"), the server constructs a target training cost function that comprehensively considers the distance of the feature and the accuracy of the label, providing a clear direction for the subsequent optimization process.

[0049] Finally, the server may optimize the first financial reconciliation guidance knowledge points in the first financial reconciliation guidance data according to the first target training cost. This process is achieved by adjusting the weights and parameters of the neural network, with the aim of improving the accuracy and effectiveness of the guidance knowledge points.

[0050] For example, the server starts to execute an optimization algorithm (such as gradient descent) to gradually adjust the parameters of the neural network according to the first target training cost. As the iterations proceed, the first financial reconciliation guidance knowledge points are gradually optimized and improved to make them more suitable for specific reconciliation tasks and conversation scenarios. For example, when dealing with differences, the optimized guidance knowledge points may include more accurate and efficient communication skills and steps.

[0051] In a possible implementation, step S141 includes: Step S1411, using the interaction attention unit in the feature interaction network, based on the reconciliation task description vector as the exploration vector and the template conversation flow graph convolution features as the index and content, determine the financial reconciliation docking features.

[0052] Step S1412: using the self-attention unit in the feature interaction network, perform self-attention processing on the financial reconciliation and docking features to generate the first candidate matching features.

[0053] In this embodiment, the reconciliation task is described in the form of a vector, which contains key information such as the type, target, and time range of the task. At the same time, the server also holds a series of template financial reconciliation session flow data, which has been processed by the graph convolutional network and converted into template session flow graph convolution features, which characterize the structure and dynamic characteristics of different session templates.

[0054] The server now needs to use a feature interaction network to process these two features. The purpose of this feature interaction network is to find the correlation and complementarity between the two, so as to generate a more comprehensive and rich feature expression.

[0055] Inside the feature interaction network, a component of an interactive attention unit is configured. The function of the interactive attention unit is to use the reconciliation task description vector as the exploration vector to explore or focus on the template conversation features that are most relevant to the reconciliation task.

[0056] Specifically, the server can pass the reconciliation task description vector as input to the interaction attention unit, which will retrieve and match the most relevant information in the convolutional features of the template conversation flow graph based on this vector.

[0057] Through this interactive attention method, the server can accurately locate the template session features that best match the current reconciliation task, that is, the financial reconciliation docking features.

[0058] After determining the financial reconciliation docking features, the server can further use the self-attention unit in the feature interaction network to process these features. The function of the self-attention unit is to make the features "self-attention", that is, to make each feature pay attention to itself and other related features, thereby enhancing the expressiveness and context perception of the features.

[0059] Specifically, the server can input the financial reconciliation docking features into the self-attention unit, which calculates the correlation between each feature and other features and adjusts the weight of each feature based on these correlations. In this way, the most relevant and important features for the current reconciliation task will be given a larger weight, while relatively unimportant features will be given a smaller weight.

[0060] In a possible implementation, step S150 includes: optimizing the first financial reconciliation guidance knowledge point in the first financial reconciliation guidance data according to the first target training cost. The method also includes: optimizing parameter information of the feature interaction network according to the first target training cost.

[0061] In this embodiment, the server sets a target training cost during the learning and optimization of the financial reconciliation guidance knowledge. This cost function reflects the difference between the model prediction and the actual result. The first target training cost is set for the first financial reconciliation guidance data.

[0062] The server will first use the first financial reconciliation guidance data to train the model. This data contains various key knowledge points about financial reconciliation. During the training process, the server can continuously compare the model's predicted output with the actual results and calculate the first target training cost.

[0063] If the training cost of a financial reconciliation guidance knowledge point is high, it means that the model performs poorly on this knowledge point, and the server will optimize this knowledge point. Optimization can include adjusting the expression of the knowledge point, adding relevant instances, or adjusting the weight of the knowledge point, so that the model can understand and apply this knowledge point more accurately.

[0064] Through continuous iteration and optimization, the server can gradually reduce the first target training cost and improve the accuracy and efficiency of the model on the first financial reconciliation guidance knowledge point.

[0065] During the training process, in addition to optimizing the financial reconciliation guidance knowledge points, the server also optimizes the parameter information of the feature interaction network based on the first target training cost.

[0066] The feature interaction network is a key component for processing the convolutional features of the reconciliation task description vector and the template conversation flow graph. Its parameter settings directly affect the effect of feature interaction and the performance of the final model. Therefore, it is crucial to adjust these parameters according to the training cost.

[0067] The server can analyze the first target training cost and find out the possible reasons for the high cost, such as unreasonable weight settings of certain network layers, improper selection of activation functions, etc. Then, the server can adjust these parameters in a targeted manner, such as updating the network weights through the gradient descent algorithm, or trying different activation functions to observe the effects.

[0068] By continuously optimizing the parameter information of the feature interaction network, the server can improve the model's feature extraction and interaction capabilities, thereby reducing training costs and improving the accuracy of financial reconciliation. This optimization process is usually iterative and requires multiple attempts and adjustments to achieve the desired effect.

[0069] In a possible implementation, the financial reconciliation guidance data to be learned includes second financial reconciliation guidance data and third financial reconciliation guidance data, the second financial reconciliation guidance data includes second financial reconciliation guidance knowledge points to be learned and reconciliation task description knowledge points extracted from the reconciliation task description data, and the third financial reconciliation guidance data includes the second financial reconciliation guidance knowledge points and each reference financial reconciliation label information in the financial reconciliation branch.

[0070] Step S120 includes: using the feature extraction network to extract features from the second financial reconciliation guidance data and the third financial reconciliation guidance data, respectively, to generate second sample financial reconciliation guidance features corresponding to the second financial reconciliation guidance data and third sample financial reconciliation guidance features corresponding to the third financial reconciliation guidance data.

[0071] In this embodiment, first, it is necessary to clarify the composition of the financial reconciliation guidance data to be learned, which includes two types of data: second financial reconciliation guidance data and third financial reconciliation guidance data.

[0072] The second financial reconciliation guidance data mainly includes two parts: one is the second financial reconciliation guidance knowledge points to be learned, which can be key information about the reconciliation process, rules or skills; the other is the reconciliation task description knowledge points extracted from the reconciliation task description data, which reflect the characteristics and requirements of specific reconciliation tasks.

[0073] The third financial reconciliation guidance data combines the second financial reconciliation guidance knowledge points and the various reference financial reconciliation label information in the financial reconciliation branch. These reference financial reconciliation label information can be different classifications, states or results of reconciliation tasks, which are used to guide the model to better understand and handle various reconciliation scenarios.

[0074] Next, we enter the feature extraction stage. The server can use a feature extraction network to process the financial reconciliation guidance data to be learned.

[0075] The server first inputs the second financial reconciliation guidance data into the feature extraction network. The feature extraction network deeply analyzes each knowledge point in the second financial reconciliation guidance data, whether it is the reconciliation guidance knowledge point to be learned or the knowledge point extracted from the reconciliation task description. The network captures the key features of these knowledge points, such as their semantic information, contextual relationships, etc.

[0076] After being processed by the network, the server can output the second sample financial reconciliation guidance features, which are abstractions and compressions of the original data, retaining the key information in the data while removing redundancy and noise.

[0077] Similar to processing the second financial reconciliation guidance data, the server also inputs the third financial reconciliation guidance data into the feature extraction network.

[0078] But this time, the feature extraction network not only needs to pay attention to the second financial reconciliation guidance knowledge point, but also needs to consider the various reference financial reconciliation label information in the financial reconciliation branch. These reference financial reconciliation label information provide more context and guidance for the network, which helps the feature extraction network to extract features more accurately.

[0079] After processing, the server can output the third sample financial reconciliation guidance features, which integrate the key features of the second financial reconciliation guidance knowledge points and the reference financial reconciliation label information, providing a rich information basis for subsequent model training and reasoning.

[0080] Therefore, through the processing of the feature extraction network, the server can convert the key features extracted from the financial reconciliation guidance data to be learned into a form that is easier for the model to understand and use, thereby improving the training efficiency and reasoning accuracy of the model.

[0081] In a possible implementation, step S140 includes: Step S144, extracting a member vector reflecting the overall semantic feature from the reconciliation task description vector as the second candidate matching feature. Extracting a member vector reflecting the overall financial reconciliation session semantic feature from the template conversation flow graph convolution feature as the third candidate matching feature.

[0082] Step S145 , determining a feature distance based on the second sample financial reconciliation guiding feature and the second candidate matching feature to generate a second feature distance, and determining a first unit training cost based on the second feature distance.

[0083] Step S146, determining the feature distance based on the third sample financial reconciliation guiding feature and the second candidate matching feature to generate a third feature distance, and determining a second unit training cost based on the third feature distance and the reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch.

[0084] Step S147, determining the feature distance based on the third sample financial reconciliation guiding feature and the third candidate matching feature, generating a fourth feature distance, and determining a third unit training cost based on the fourth feature distance and the reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch.

[0085] Step S148: determining a second target training cost based on the first unit training cost, the second unit training cost, and the third unit training cost.

[0086] First, the server extracts a member vector reflecting the overall semantic features from the reconciliation task description vector. This vector is called the second candidate matching feature, which captures the overall intent and key information of the reconciliation task. At the same time, the server also extracts a member vector reflecting the semantic features of the overall financial reconciliation session from the template session flow graph convolutional features as the third candidate matching feature, which represents the overall structure and semantic information of the financial reconciliation session.

[0087] Next, the server uses the second sample financial reconciliation guide feature and the second candidate matching feature to determine the feature distance, which can be accomplished by calculating the similarity or difference between the two feature vectors, such as using cosine similarity, Euclidean distance, and other methods.

[0088] After the server calculates the second feature distance, it determines the first unit training cost based on this distance. If the second feature distance is large, it means that the features predicted by the model are quite different from the actual guided features, so the training cost will be relatively high. On the contrary, if the feature distance is small, it means that the model prediction is more accurate and the training cost is relatively low.

[0089] The server continues to use the third sample financial reconciliation guidance feature and the second candidate matching feature to determine the feature distance and generate a third feature distance. The purpose of this step is to compare the matching degree between the financial reconciliation guidance knowledge point and the reconciliation task description.

[0090] Based on the third feature distance and the reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch, the server determines a second unit training cost, which reflects the accuracy of the model in understanding and applying the financial reconciliation guidance knowledge points.

[0091] Then, the server uses the third sample financial reconciliation guidance feature and the third candidate matching feature to determine the feature distance to obtain a fourth feature distance. This step is intended to evaluate the model's overall understanding and application capabilities of the financial reconciliation conversation flow.

[0092] According to the fourth feature distance and the reference financial reconciliation label of the template financial reconciliation session flow in the financial reconciliation branch, the server calculates a third unit training cost, where the third unit training cost reflects the performance of the model when processing a specific financial reconciliation session.

[0093] Finally, the server comprehensively determines the second target training cost based on the first unit training cost, the second unit training cost and the third unit training cost. The second target training cost is the performance indicator of the model in the overall reconciliation guidance task, which is used to guide the subsequent model optimization and training process.

[0094] Through the above steps, the server can comprehensively evaluate the performance of the model in the financial reconciliation guidance task, and perform targeted optimization and improvement based on the training cost.

[0095] In a possible implementation, the step S150 includes: optimizing the second financial reconciliation guidance knowledge points in the second financial reconciliation guidance data and the third financial reconciliation guidance data according to the second target training cost.

[0096] In this embodiment, after completing the above feature extraction, feature distance determination and training cost calculation, the server will then optimize the financial reconciliation guidance data to be learned based on the obtained target training cost. In this scenario, the second target training cost and the second financial reconciliation guidance data related thereto and the second financial reconciliation guidance knowledge points in the third financial reconciliation guidance data are mainly focused on.

[0097] The server first analyzes the second objective training cost, which is a combination of the first, second, and third unit training costs calculated in the previous steps. It reflects the performance of the model in the overall financial reconciliation guidance task.

[0098] If it is found that the second objective training cost is high, it means that the model is insufficient in understanding and applying the second financial reconciliation guidance knowledge point. At this time, the server can optimize the second financial reconciliation guidance knowledge point.

[0099] Specifically, the server can review the description and examples of the second financial reconciliation guidance knowledge points in the second financial reconciliation guidance data and the third financial reconciliation guidance data. It may find that the description of some knowledge points is not clear enough, or the examples are not typical enough, making it difficult for the model to accurately understand and apply these knowledge points.

[0100] To address these issues, the server can perform a series of optimization operations. For example, it can reorganize the language to make the knowledge points more accurate and easier to understand; or add more representative examples to help the model better grasp the connotation and application scenarios of the knowledge points.

[0101] In addition, the server may also use machine learning methods, such as reinforcement learning or deep learning technology, to further refine and optimize the second financial reconciliation guidance knowledge points. By continuously adjusting the expression of knowledge points and instance content, the server aims to reduce the second objective training cost and improve the accuracy and efficiency of the model in the financial reconciliation guidance task.

[0102] These optimization operations are an iterative process, and the server can continuously try different optimization solutions until it finds the best combination that can effectively reduce the training cost of the second objective. In this way, the server can gradually improve the model's ability to understand and apply the second financial reconciliation guidance knowledge points, thereby improving the performance of the overall financial reconciliation guidance task.

[0103] In a possible implementation, the method further includes: Step S160: obtaining the input target financial reconciliation session flow.

[0104] Step S170, using the graph convolution unit in the neural network model, extracting features of the target financial reconciliation conversation flow to generate graph convolution features of the target conversation flow.

[0105] Step S180, determining the feature distance based on the target conversation flow graph convolution feature and the target financial reconciliation guidance feature, and determining the financial reconciliation label data corresponding to the target financial reconciliation conversation flow based on the feature distance result, wherein the target financial reconciliation guidance feature is generated by extracting the target financial reconciliation guidance data using the feature extraction network in the neural network model, and the target financial reconciliation guidance data includes the optimized financial reconciliation guidance knowledge points and each reference financial reconciliation label information in the financial reconciliation branch.

[0106] In this embodiment, the server first receives an input, which is a target financial reconciliation conversation flow, which can be a financial conversation record within an enterprise, or communication information about account reconciliation between a bank and a customer. The server uses the target financial reconciliation conversation flow as basic data for subsequent processing.

[0107] The server then uses the graph convolution unit in the trained neural network model to extract features from the target financial reconciliation conversation flow. The graph convolution unit can deeply understand the text structure and semantic information in the conversation flow and capture key features.

[0108] In this process, the server regards each sentence and each word in the target financial reconciliation conversation flow as a node in the graph, and extracts the relationship and features between these nodes through graph convolution operations. These features can include the semantics of words, the grammatical structure of sentences, and the logical flow of the entire conversation flow. After processing by the graph convolution unit, the server generates the target conversation flow graph convolution feature, which is a high-dimensional feature vector that contains key information in the target financial reconciliation conversation flow.

[0109] The server then needs to determine the financial reconciliation tag data corresponding to the target financial reconciliation session flow. To do this, the feature extraction network in the neural network model is first used to extract features from the target financial reconciliation guidance data to generate target financial reconciliation guidance features. The target financial reconciliation guidance data contains optimized financial reconciliation guidance knowledge points and various reference financial reconciliation tag information in the financial reconciliation branch.

[0110] The server then determines the feature distance between the target conversation flow graph convolutional features and the target financial reconciliation guidance features, which can be obtained by calculating the similarity between the two feature vectors. If the two feature vectors are very similar, then their feature distance will be close.

[0111] Based on this feature distance result, the server can determine the financial reconciliation tag data corresponding to the target financial reconciliation conversation flow. If the feature distance is very close, it means that the conversation flow is highly matched with a financial reconciliation tag, and the server will use this tag data as the output result. For example, if the target financial reconciliation conversation flow mentions "invoice verification is correct", the server may label it as "verified".

[0112] Through this process, the server can automatically add accurate tags to the input financial reconciliation session flow, thereby helping enterprises handle large amounts of financial reconciliation tasks more efficiently.

[0113] In a possible implementation, the target financial reconciliation guidance feature includes a first target financial reconciliation guidance feature and a second target financial reconciliation guidance feature, the first target financial reconciliation guidance feature corresponds to a first target financial reconciliation guidance data including a first financial reconciliation guidance knowledge point, the second target financial reconciliation guidance feature corresponds to a second target financial reconciliation guidance data including a second financial reconciliation guidance knowledge point, and the first financial reconciliation guidance knowledge point and the second financial reconciliation guidance knowledge point are obtained using different methods.

[0114] Step S180 includes: Step S181, determining a feature distance based on the target conversation flow graph convolution feature and the first target financial reconciliation guidance feature, and generating a fifth feature distance result.

[0115] Step S182, determining the feature distance based on the target conversation flow graph convolution feature and the second target financial reconciliation guidance feature, and generating a sixth feature distance result.

[0116] Step S183: determining the financial reconciliation tag data based on the fifth characteristic distance result and the sixth characteristic distance result.

[0117] Among them, the first financial reconciliation guidance knowledge point is generated by learning based on the feature interaction fusion features of the reconciliation task description vector with mapping relationship and the template conversation flow graph convolution feature, and the second financial reconciliation guidance knowledge point is generated by deriving the first financial reconciliation guidance knowledge point using the first derivation strategy.

[0118] In this embodiment, when the server processes the target financial reconciliation conversation flow, different financial reconciliation guidance features and knowledge points are involved. These financial reconciliation guidance features and knowledge points are obtained in different ways and are used to determine the financial reconciliation tag data of the conversation flow.

[0119] First, two important concepts need to be clarified: the first target financial reconciliation guidance feature and the second target financial reconciliation guidance feature, which correspond to data containing different financial reconciliation guidance knowledge points. Specifically, the first target financial reconciliation guidance feature corresponds to the first target financial reconciliation guidance data including the first financial reconciliation guidance knowledge point, while the second target financial reconciliation guidance feature corresponds to the second target financial reconciliation guidance data including the second financial reconciliation guidance knowledge point.

[0120] The first financial reconciliation guidance knowledge point and the second financial reconciliation guidance knowledge point are obtained in different ways. The first financial reconciliation guidance knowledge point is generated by learning based on the feature interaction fusion features of the reconciliation task description vector with mapping connection and the template conversation flow graph convolution feature, which means that the server extracts key information and fuses it by analyzing the description of the reconciliation task and the existing template conversation flow, thereby generating the first financial reconciliation guidance knowledge point.

[0121] The second financial reconciliation guidance knowledge point is generated by deriving the first financial reconciliation guidance knowledge point using the first derivative strategy. This means that the server further derives the second financial reconciliation guidance knowledge point based on the existing first financial reconciliation guidance knowledge point by applying a certain derivative strategy (such as rule reasoning, pattern recognition, etc.).

[0122] Next, the server will use these features and knowledge points to determine the label data of the target financial reconciliation conversation flow. The server first determines the feature distance based on the target conversation flow graph convolution feature and the first target financial reconciliation guide feature to generate the fifth feature distance result. This step is to evaluate the similarity or matching degree between the conversation flow and the first financial reconciliation guide knowledge point. Then, the server determines the feature distance based on the target conversation flow graph convolution feature and the second target financial reconciliation guide feature to generate the sixth feature distance result. This step is to evaluate the similarity or matching degree between the conversation flow and the second financial reconciliation guide knowledge point. Finally, based on the fifth feature distance result and the sixth feature distance result, the server comprehensively determines which financial reconciliation guide knowledge point has a higher match with the target financial reconciliation conversation flow, thereby determining the financial reconciliation label data corresponding to the conversation flow. For example, if the fifth feature distance result is smaller, it means that the conversation flow is more matched with the first financial reconciliation guide knowledge point, and therefore may be labeled with a label related to the first financial reconciliation guide knowledge point.

[0123] In this way, the server can use different financial reconciliation guidance features and knowledge points to accurately label the input financial reconciliation session flow, thereby helping enterprises handle financial reconciliation tasks more efficiently.

[0124] Figure 2 The hardware structure of the neural network-based intelligent financial reconciliation processing system 100 provided in the embodiment of the present application for implementing the above-mentioned neural network-based intelligent financial reconciliation processing method is shown as follows: Figure 2 As shown, the neural network-based financial reconciliation intelligent processing system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0125] In one possible design, the neural network-based intelligent processing system 100 for financial reconciliation can be a single server or a server group. The server group can be centralized or distributed (for example, the neural network-based intelligent processing system 100 for financial reconciliation can be a distributed system). In some embodiments, the neural network-based intelligent processing system 100 for financial reconciliation can be local or remote. For example, the neural network-based intelligent processing system 100 for financial reconciliation can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the neural network-based intelligent processing system 100 for financial reconciliation can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the neural network-based intelligent processing system 100 for financial reconciliation can be implemented on a neural network-based intelligent processing system for financial reconciliation. As an example only, the neural network-based intelligent processing system for financial reconciliation can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.

[0126] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the neural network-based financial reconciliation intelligent processing system 100 uses to execute or use to complete the exemplary method described in this application.

[0127] In the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the intelligent processing method of financial reconciliation based on neural network in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected using the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0128] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned neural network-based financial reconciliation intelligent processing system 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0129] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned intelligent processing method of financial reconciliation based on neural network is implemented.

[0130] It should be noted that in order to simplify the description disclosed in the present application and thereby facilitate the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, various features are sometimes grouped into one embodiment, drawing or description thereof.

Claims

1. A method for intelligent processing of financial reconciliation based on neural network, characterized in that: The method comprises: Determine the reconciliation task description data corresponding to the template financial reconciliation session flow; Using the feature extraction network in the neural network model, feature extraction is performed on the reconciliation task description data to generate a reconciliation task description vector; using the feature extraction network, feature extraction is performed on the financial reconciliation guidance data to be learned to generate sample financial reconciliation guidance features, wherein the financial reconciliation guidance data to be learned includes the financial reconciliation guidance knowledge points to be learned; Using the graph convolution unit in the neural network model, extracting features from the template financial reconciliation conversation flow to generate graph convolution features of the template conversation flow; Determine the feature distance based on the sample financial reconciliation guide feature and the candidate matching feature, and determine the target training cost based on the feature distance result; the candidate matching feature is determined based on the reconciliation task description vector and the template conversation flow graph convolution feature; The financial reconciliation guidance knowledge points in the to-be-learned financial reconciliation guidance data are optimized according to the target training cost.

2. The intelligent processing method for financial reconciliation based on neural network according to claim 1 is characterized in that: The to-be-learned financial reconciliation guidance data includes first financial reconciliation guidance data, wherein the first financial reconciliation guidance data includes first financial reconciliation guidance knowledge points to be learned and each reference financial reconciliation label information in a financial reconciliation branch; the sample financial reconciliation guidance feature includes a first sample financial reconciliation guidance feature; The determining of the feature distance based on the sample financial reconciliation guide feature and the candidate matching feature, and determining the target training cost based on the feature distance result, includes: Using a feature interaction network, perform feature interaction aggregation on the account reconciliation task description vector and the template conversation flow graph convolutional features to generate a first candidate matching feature; Determine the feature distance based on the first sample financial reconciliation guidance feature and the first candidate matching feature to generate a first feature distance; the first feature distance reflects the deviation between the template financial reconciliation conversation flow and each reference financial reconciliation tag information in the financial reconciliation branch; A first target training cost is determined based on the first feature distance and a reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch.

3. The intelligent processing method for financial reconciliation based on neural network according to claim 2 is characterized in that: The utilizing the feature interaction network to perform feature interaction aggregation on the reconciliation task description vector and the template conversation flow graph convolution feature to generate a first candidate matching feature includes: Determine financial reconciliation docking features by using the interactive attention unit in the feature interaction network based on the reconciliation task description vector as a search vector and the template conversation flow graph convolution features as indexes and contents; The self-attention unit in the feature interaction network is used to perform self-attention processing on the financial reconciliation docking feature to generate the first candidate matching feature.

4. The intelligent processing method for financial reconciliation based on neural network according to claim 2 or 3, characterized in that: The optimizing the financial reconciliation guidance knowledge points in the to-be-learned financial reconciliation guidance data according to the target training cost includes: optimizing the first financial reconciliation guidance knowledge points in the first financial reconciliation guidance data according to the first target training cost; The method also includes: optimizing parameter information of the feature interaction network according to the first target training cost.

5. The neural network-based intelligent processing method for financial reconciliation according to any one of claims 1 to 3, characterized in that: The financial reconciliation guidance data to be learned includes second financial reconciliation guidance data and third financial reconciliation guidance data, wherein the second financial reconciliation guidance data includes second financial reconciliation guidance knowledge points to be learned and reconciliation task description knowledge points extracted from the reconciliation task description data, and the third financial reconciliation guidance data includes the second financial reconciliation guidance knowledge points and each reference financial reconciliation label information in the financial reconciliation branch; The method of using the feature extraction network to extract features from the financial reconciliation guidance data to be learned and generating sample financial reconciliation guidance features includes: The feature extraction network is used to extract features from the second financial reconciliation guidance data and the third financial reconciliation guidance data, respectively, to generate second sample financial reconciliation guidance features corresponding to the second financial reconciliation guidance data and third sample financial reconciliation guidance features corresponding to the third financial reconciliation guidance data.

6. The intelligent processing method for financial reconciliation based on neural network according to claim 5 is characterized in that: The determining of the feature distance based on the sample financial reconciliation guide feature and the candidate matching feature, and determining the target training cost based on the feature distance result, includes: Extracting a member vector reflecting the overall semantic feature from the reconciliation task description vector as the second candidate matching feature; extracting a member vector reflecting the overall financial reconciliation session semantic feature from the template conversation flow graph convolution feature as the third candidate matching feature; Determine a feature distance based on the second sample financial reconciliation guide feature and the second candidate matching feature to generate a second feature distance, and determine a first unit training cost based on the second feature distance; Determine a feature distance based on the third sample financial reconciliation guiding feature and the second candidate matching feature to generate a third feature distance, and determine a second unit training cost based on the third feature distance and a reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch; Determine a feature distance based on the third sample financial reconciliation guidance feature and the third candidate matching feature to generate a fourth feature distance, and determine a third unit training cost based on the fourth feature distance and a reference financial reconciliation label of the template financial reconciliation conversation flow in the financial reconciliation branch; A second target training cost is determined based on the first unit training cost, the second unit training cost, and the third unit training cost.

7. The neural network-based intelligent processing method for financial reconciliation according to claim 6 is characterized in that: The optimizing the financial reconciliation guidance knowledge points in the to-be-learned financial reconciliation guidance data according to the target training cost includes: The second financial reconciliation guidance knowledge points in the second financial reconciliation guidance data and the third financial reconciliation guidance data are optimized according to the second target training cost.

8. The neural network-based intelligent financial reconciliation processing method according to claim 1 is characterized in that: The method further comprises: Get the input target financial reconciliation session flow; Using the graph convolution unit in the neural network model, extracting features of the target financial reconciliation conversation flow to generate graph convolution features of the target conversation flow; The feature distance is determined based on the target conversation flow graph convolution features and the target financial reconciliation guidance features, and based on the feature distance results, the financial reconciliation label data corresponding to the target financial reconciliation conversation flow is determined, wherein the target financial reconciliation guidance features are generated by extracting features from the target financial reconciliation guidance data using the feature extraction network in the neural network model, and the target financial reconciliation guidance data includes the optimized financial reconciliation guidance knowledge points and each reference financial reconciliation label information in the financial reconciliation branch.

9. The neural network-based intelligent processing method for financial reconciliation according to claim 8, characterized in that: The target financial reconciliation guidance feature includes a first target financial reconciliation guidance feature and a second target financial reconciliation guidance feature, the first target financial reconciliation guidance feature corresponds to first target financial reconciliation guidance data including a first financial reconciliation guidance knowledge point, the second target financial reconciliation guidance feature corresponds to second target financial reconciliation guidance data including a second financial reconciliation guidance knowledge point, and the first financial reconciliation guidance knowledge point and the second financial reconciliation guidance knowledge point are obtained in different ways; The determining of the feature distance based on the target conversation flow graph convolution feature and the target financial reconciliation guidance feature, and determining the financial reconciliation label data corresponding to the target financial reconciliation conversation flow based on the feature distance result, includes: Determine a feature distance based on the target conversation flow graph convolution feature and the first target financial reconciliation guidance feature to generate a fifth feature distance result; Determine a feature distance based on the target conversation flow graph convolution feature and the second target financial reconciliation guidance feature to generate a sixth feature distance result; Determining the financial reconciliation tag data based on the fifth characteristic distance result and the sixth characteristic distance result; Among them, the first financial reconciliation guidance knowledge point is generated by learning based on the feature interaction fusion features of the reconciliation task description vector with mapping relationship and the template conversation flow graph convolution feature, and the second financial reconciliation guidance knowledge point is generated by deriving the first financial reconciliation guidance knowledge point using the first derivation strategy.

10. A financial reconciliation intelligent processing system based on neural network, characterized in that: The neural network-based intelligent processing system for financial reconciliation includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the neural network-based intelligent processing method for financial reconciliation as described in any one of claims 1 to 9.

Images