Digital processing method and device based on bidirectional feedback tuning model

Abstract

The invention discloses a digital processing method and device based on a bidirectional feedback tuning model, relates to the technical field of digital processing, and mainly aims to improve the accuracy of digital processing and realize self-optimization of a digital processing model. The method mainly comprises the following steps: analyzing a digital processing task to obtain to-be-processed data, a target processing link and a processing mode; a digital processing model matched with the processing mode is recognized from a digital processing platform server, and the digital processing model is obtained based on historical manual correction information of a historical processing process and training data set optimization of historical model performance evaluation parameter updating; processing the to-be-processed data by using the digital processing model, and sending an obtained initial processing result to a manual correction and modification node; and obtaining manual correction information of the initial processing result, and generating a processing result of the to-be-processed data according to the manual correction information and the initial processing result. The method is mainly used for digital processing.

Description

Technical Field

[0001] The present invention relates to the field of digital processing technology, and in particular to a digital processing method and device based on a two-way feedback tuning model. Background Art

[0002] Digital processing is the process of converting paper or non-digital information materials, such as newspapers, books, magazines, atlases, forms, official documents, historical documents, cultural relics, etc., into digital format through computer technology. As large models gradually move towards verticalization and specialization, customized data processing services will occupy a larger proportion in the digital processing market. Customized data processing services have become the mainstream of current market demand, which also puts higher requirements on the adaptability of digital processing models.

[0003] In order to improve the adaptability of the digital processing model, it is necessary to continuously update the data set of the digital processing model so as to fine-tune the model based on the data set. The current data set mainly relies on manually labeled data samples to construct. Since a large amount of data will be generated during the digitization process, the data sets used for model training are mostly manually selected. There are problems such as a small number of samples of a certain type of data or uneven distribution, which leads to problems such as overfitting, bias, underfitting, weak generalization ability, and misclassification caused by imbalanced categories in model training. When faced with new data, the digital processing model cannot adapt to the data in time, resulting in unstable accuracy of the digital processing results. Summary of the invention

[0004] In view of this, the present invention provides a digital processing method and device based on a two-way feedback tuning model, the main purpose of which is to solve the problem of unstable accuracy of digital processing results caused by the difficulty in optimizing the current digital processing model and further improving the accuracy.

[0005] According to one aspect of the present invention, a digital processing method based on a bidirectional feedback tuning model is provided, comprising:

[0006] Analyze the digital processing task to obtain the data to be processed, the target processing link and the processing method, wherein the target processing link includes any processing link in the whole process of digital processing;

[0007] Identifying a digital processing model matching the processing method from a digital processing platform server, wherein the digital processing model is obtained by tuning a training data set based on historical manual correction information of historical processing processes and updated historical model performance evaluation parameters;

[0008] Processing the data to be processed using the digital processing model, and sending the obtained initial processing result to the manual correction node;

[0009] The manual correction information of the initial processing result is obtained, and the processing result of the to-be-processed data is generated according to the manual correction information and the initial processing result.

[0010] Furthermore, after generating the processing result of the data to be processed according to the manual correction information and the initial processing result, the method further includes:

[0011] updating the data set generation strategy according to the processing result and the manual correction information, and regenerating the training data set according to the updated data set generation strategy;

[0012] Using the training data set to train the current version of the digital processing model to obtain a new version of the digital processing model;

[0013] A model performance evaluation is performed on the new version of the digital processing model to obtain performance evaluation parameters of the new version of the model, and a data set generation strategy is updated according to the performance evaluation parameters of the new version of the model and the performance evaluation parameters of the current version of the model to obtain a further updated data set generation strategy.

[0014] Furthermore, after obtaining the updated data set generation strategy, the method further includes:

[0015] When the performance evaluation parameters of the new version model are better than the performance evaluation parameters of the current version model, publishing the new version of the digital processing model to the digital processing platform;

[0016] In the case where the performance evaluation parameters of the new version model are inferior to the performance evaluation parameters of the current version model,

[0017] Regenerate the training data set according to the updated data set generation strategy, return to the step of using the training data set to train the current version of the digital processing model to obtain a new version of the digital processing model, and iteratively update the data set generation strategy until the performance evaluation parameters of the new version of the model are better than the performance evaluation parameters of the current version of the model, or the number of cycles reaches a threshold.

[0018] Furthermore, the updating of the data set generation strategy based on the processing result and the manual correction information includes:

[0019] Quantifying the manual correction information to obtain reward values ​​corresponding to different data types and / or different processing items, wherein the manual correction information includes processing error prompts and / or data type missing information;

[0020] Reinforcement learning is performed on the parameters of the data set generation strategy based on the processing results and the reward value to update the data set generation strategy, wherein the parameters include the number of data sets, the amount of data of different data types, and the number of data labels.

[0021] Furthermore, the processing result includes positive data and negative data, and the regenerating the training data set according to the updated data set generation strategy includes:

[0022] Acquire an original training data set, and add, delete, or modify data types or data contents of the original training data set according to the updated data set generation strategy;

[0023] Filtering new training data from the positive data and negative data according to the updated data set generation strategy;

[0024] A training data set is generated according to the addition, deletion and modification results of the newly added training data and the original training data set.

[0025] Further, the updating of the data set generation strategy according to the performance evaluation parameters of the new version model to obtain a further updated data set generation strategy includes:

[0026] Obtain regenerated training data sets and processing result feedback information;

[0027] The performance evaluation parameters of the new version model, the regenerated training data set and the processing result feedback information are used as feedback states to perform reinforcement learning on the parameters of the data set generation strategy to obtain a further updated data set generation strategy.

[0028] Furthermore, the digital processing model includes an OCR recognition model, a layout document automatic indexing model, a table detection model, a table structuring model, a formula detection model and a formula structuring model.

[0029] According to another aspect of the present invention, a digital processing device based on a bidirectional feedback tuning model is provided, comprising:

[0030] An analysis module is used to analyze the digital processing task to obtain the data to be processed, the target processing link and the processing method, wherein the target processing link includes any processing link in the whole process of digital processing;

[0031] A determination module, configured to identify a digital processing model matching the processing method from a digital processing platform server, wherein the digital processing model is obtained by tuning a training data set based on historical manual correction information of a historical processing process and updated with historical model performance evaluation parameters;

[0032] A processing module, used to process the data to be processed using the digital processing model, and send the obtained initial processing result to the manual correction node;

[0033] The acquisition module is used to acquire the manual correction information of the initial processing result, and generate the processing result of the data to be processed according to the manual correction information and the initial processing result.

[0034] Furthermore, the device also includes:

[0035] A first updating module, used for updating a data set generation strategy according to the processing result and the manual correction information, and regenerating a training data set according to the updated data set generation strategy;

[0036] A first training module is used to train the current version of the digital processing model using the training data set to obtain a new version of the digital processing model;

[0037] The second updating module is used to perform model performance evaluation on the new version of the digital processing model to obtain performance evaluation parameters of the new version of the model, and update the data set generation strategy based on the performance evaluation parameters of the new version of the model and the performance evaluation parameters of the current version of the model to obtain a further updated data set generation strategy.

[0038] Furthermore, the device also includes:

[0039] A publishing module, used for publishing the new version of the digital processing model to the digital processing platform when the performance evaluation parameters of the new version of the model are better than the performance evaluation parameters of the current version of the model;

[0040] The second training module is used to regenerate the training data set according to the again updated data set generation strategy when the performance evaluation parameters of the new version model are inferior to the performance evaluation parameters of the current version model, and return to the step of using the training data set to train the current version of the digital processing model to obtain a new version of the digital processing model, and iteratively update the data set generation strategy until the performance evaluation parameters of the new version model are better than the performance evaluation parameters of the current version model, or the number of cycles reaches a threshold.

[0041] Furthermore, the first update module includes:

[0042] A processing unit, configured to quantify the manual correction information to obtain reward values ​​corresponding to different data types and / or different processing items, wherein the manual correction information includes processing error prompts and / or data type missing information;

[0043] The first reinforcement learning unit is used to perform reinforcement learning on the parameters of the data set generation strategy according to the processing result and the reward value to update the data set generation strategy, wherein the parameters include the number of data sets, the amount of data of different data types and the number of data labels.

[0044] Furthermore, the first update module further includes:

[0045] A first acquisition unit is used to acquire an original training data set, and add, delete, or modify data types or data contents of the original training data set according to the updated data set generation strategy;

[0046] A screening unit, configured to screen out newly added training data from the positive data and the negative data according to the updated data set generation strategy;

[0047] A generating unit is used to generate a training data set according to the addition, deletion and modification results of the newly added training data and the original training data set.

[0048] Furthermore, the second update module includes:

[0049] A second acquisition unit, used to acquire the regenerated training data set and processing result feedback information;

[0050] The second reinforcement learning unit is used to perform reinforcement learning on the parameters of the data set generation strategy by using the performance evaluation parameters of the new version model, the regenerated training data set and the processing result feedback information as feedback states to obtain a further updated data set generation strategy.

[0051] Furthermore, in a specific application scenario, the digital processing model in the processing module includes an OCR recognition model, a layout document automatic indexing model, a table detection model, a table structuring model, a formula detection model and a formula structuring model.

[0052] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored in the storage medium, and the executable instruction enables a processor to execute operations corresponding to the digital processing method based on the bidirectional feedback tuning model as described above.

[0053] According to another aspect of the present invention, there is provided a terminal, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0054] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the digital processing method based on the bidirectional feedback tuning model.

[0055] By means of the above technical solution, the technical solution provided by the embodiment of the present invention has at least the following advantages:

[0056] The present invention provides a digital processing method and device based on a two-way feedback tuning model. The embodiment of the present invention obtains the data to be processed, the target processing link and the processing method by analyzing the digital processing task, wherein the target processing link includes any processing link in the whole process of digital processing; identifies the digital processing model matching the processing method from the digital processing platform server, wherein the digital processing model is obtained by tuning the training data set updated based on the historical manual correction information of the historical processing process and the historical model performance evaluation parameters; uses the digital processing model to process the data to be processed, and sends the obtained initial processing result to the manual correction node; obtains the manual correction information of the initial processing result, and generates the processing result of the data to be processed based on the manual correction information and the initial processing result. Based on the historical manual correction information of the historical processing process and the historical model performance evaluation parameters, the generation strategy of the model training data set is optimized from two directions to realize the automatic generation of the training data set, which greatly reduces the time of digital processing model tuning. At the same time, it ensures the accuracy of digital processing model tuning, so that the digital processing model can timely and accurately adapt to the changes in digital processing needs, thereby improving the accuracy of digital processing.

[0057] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0059] Figure 1 A flow chart of a digital processing method based on a bidirectional feedback tuning model provided by an embodiment of the present invention is shown;

[0060] Figure 2 Another flow chart of a digital processing method based on a bidirectional feedback tuning model provided by an embodiment of the present invention is shown;

[0061] Figure 3 A schematic diagram of a reinforcement learning process provided by an embodiment of the present invention is shown;

[0062] Figure 4A schematic diagram of a process of digital processing and digital processing model optimization provided by an embodiment of the present invention is shown;

[0063] Figure 5 A block diagram of a digital processing device based on a bidirectional feedback tuning model provided by an embodiment of the present invention is shown;

[0064] Figure 6 A schematic diagram of the structure of a terminal provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0065] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0066] In view of the difficulty in optimizing the current digital processing model and further improving the accuracy, the accuracy of the digital processing results is unstable. The embodiment of the present invention provides a digital processing method based on a two-way feedback tuning model, such as Figure 1 As shown, the method includes:

[0067] 101. Analyze the digital processing task to obtain the data to be processed, the target processing link and the processing method.

[0068] In the embodiment of the present invention, the overall process of digital processing includes multiple processing links, which can also be called processing nodes. For example, the digital processing link includes multiple links such as data acquisition, image preprocessing, OCR recognition, data cleaning and proofreading, data structuring and annotation. The digital processing task is generated based on the data processing requirements submitted by the user through the digital processing platform and the data to be processed (data to be processed). According to the data processing needs, the target processing link and processing method can be parsed. For example, the data processing requirement is "annotate the metadata in journal A", and the data to be processed is the original data of the entire journal of journal A. Then, parsing the digital processing task can obtain the target processing link as "data structuring and annotation", and the processing method is "metadata field annotation of document data". Among them, the digital processing platform can be a platform open to users across the entire network, or it can be a customized platform within the enterprise, and the embodiment of the present invention does not make specific limitations.

[0069] 102. Identify a digital processing model matching the processing method from a digital processing platform server.

[0070] In an embodiment of the present invention, all digital processing models are stored in the digital processing platform server, and different digital processing models support corresponding data processing methods. Therefore, after determining the processing method, it is necessary to obtain the digital processing model used to implement the processing method. Among them, the processing method is the functional type of data processing. For example, a processing method for processing various original files such as images, paper or electronic files into high-quality PDFs with correct layout and correct fonts. A processing method for extracting key metadata information such as titles, author organizations, abstracts, keywords, etc. in layout documents. A processing method for fragmenting the full text into fragmented chapters, paragraphs, pictures, tables, formulas, etc.

[0071] It should be noted that the digital processing model is obtained by tuning the training data set based on the historical manual correction information of the historical processing process and the historical model performance evaluation parameter update. In other words, the model of the digital processing model is a model that is continuously trained based on the training data set to achieve parameter tuning, and the update basis of the training data set is feedback information from two directions of historical manual correction information and historical model performance evaluation parameters. Among them, the historical manual correction information is the manual correction information of the initial processing result of the previous processing data of the current digital processing model. Manual correction is to ensure the accuracy of the processing result. It is a necessary step before outputting the final processing result. It is not an additional manual intervention step for tuning the model. Therefore, the update process of the training data set does not require manual update marking of the data, does not cause an increase in labor costs, and greatly reduces the construction time of the training data set. In addition, updating the training data set based on the historical model performance evaluation parameters is a process of tuning the model from the performance dimension of the model itself, which can improve the processing accuracy and robustness of the model, thereby effectively improving the fine-tuning efficiency of the digital processing model and meeting the timeliness requirements of digital processing for model adaptability adjustment to a greater extent.

[0072] 103. Process the data to be processed using the digital processing model, and send the obtained initial processing result to a manual correction node.

[0073] In the embodiment of the present invention, since the amount of data to be processed is huge and there are constant additions and updates, such as, taking a journal as an example, the document format of different issues of the same journal is constantly updated, and new journals will also be published with new document formats. The training data set of the digital processing model cannot achieve full coverage of the data to be processed, nor can it ensure 100% accuracy of data processing. Therefore, after processing by the digital processing model, the initial processing result obtained by processing needs to be sent to the manual correction node to ensure the accuracy of data processing by manual intervention correction. Although manual correction is involved here, the workload of manual correction is much smaller than the data processing amount of the digital processing model, and with the continuous iteration of the processing model, the proportion of manual participation will be infinitely reduced, and the digital processing model still has its important value.

[0074] 104. Obtain manual correction information of the initial processing result, and generate a processing result of the to-be-processed data based on the manual correction information and the initial processing result.

[0075] In an embodiment of the present invention, the manual correction information is the content of the manual modification of the erroneous data in the initial processing result on the digital processing platform, the data type corresponding to the modified content, the cause of the processing error, etc. For example, in the metadata extraction result of the journal, the extraction error of the document title, after manual correction, the modified content includes the correct position coordinates of the document title, the data type is the publication type of the current journal, and the cause of the processing error is attributed to the layout recognition problem. Since the fine-tuning efficiency of the digital processing model is high, it can adapt to the high-frequency update of the data in time and continuously improve the processing coverage breadth of the model, so the accuracy of digital processing can be improved. In addition, introducing manual correction on the basis of processing based on the digital processing model can further improve the accuracy of digital processing.

[0076] In an embodiment of the present invention, for further explanation and limitation, as Figure 2 As shown, after the step of generating the processing result of the data to be processed according to the manual correction information and the initial processing result, the method further includes:

[0077] 201. Update a data set generation strategy according to the processing result and the manual correction information, and regenerate a training data set according to the updated data set generation strategy.

[0078] 202. Using the training data set to train the current version of the digital processing model to obtain a new version of the digital processing model.

[0079] 203. Perform model performance evaluation on the new version of the digital processing model to obtain performance evaluation parameters of the new version of the model, and update the data set generation strategy based on the performance evaluation parameters of the new version of the model and the performance evaluation parameters of the current version of the model to obtain a further updated data set generation strategy.

[0080] In the embodiment of the present invention, each time the digital processing model completes data processing, the data set generation strategy is updated according to the processing results and the manual teaching information, so as to regenerate the training data set for fine-tuning the digital processing model based on the updated data set generation strategy, so that the training data set is more targeted to the model training. For example, if it is detected that there are fewer samples of a certain type of data or the distribution is uneven, the data set generation strategy can be adjusted to increase the number or proportion of this type of data in the regenerated training data set. After fine-tuning the digital processing model based on the regenerated training data set, it is also necessary to evaluate the model performance of the fine-tuned digital processing model to ensure the accuracy and reliability of the model in practical applications from the model performance dimension. Among them, the performance evaluation parameters of the model include accuracy, recall rate, F1 score, etc. In the model evaluation process, K-fold cross validation or leave-one-out method can be used for cross validation, and error analysis can be performed based on error types and error causes. After determining the performance evaluation parameters of the model, the data set generation strategy is updated again based on the obtained performance evaluation parameters and the comparison results between the performance evaluation parameters and the performance evaluation parameters before the model fine-tuning. The updated dataset generation strategy will also be used to guide the generation of training datasets for the next round of model training of the digital processing model. By updating the dataset generation strategy twice, two-way feedback can be achieved on the fine-tuning of the model from both the processing result dimension and the model performance dimension, ensuring the quality of model tuning and maximizing the tuning effect, thereby ensuring the accuracy and reliability of digital processing model processing.

[0081] In one embodiment of the present invention, for further explanation and limitation, after obtaining the updated data set generation strategy in step 1, the method further includes:

[0082] When the performance evaluation parameters of the new version model are better than the performance evaluation parameters of the current version model, publishing the new version of the digital processing model to the digital processing platform;

[0083] When the performance evaluation parameters of the new version model are inferior to the performance evaluation parameters of the current version model, the training data set is regenerated according to the again updated data set generation strategy, and the step of using the training data set to train the current version of the digital processing model to obtain a new version of the digital processing model is returned to iteratively update the data set generation strategy until the performance evaluation parameters of the new version model are superior to the performance evaluation parameters of the current version model, or the number of cycles reaches a threshold.

[0084] In the embodiment of the present invention, the digital processing model will evaluate the performance parameters after each fine-tuning training. After obtaining the performance evaluation parameters of the new version of the current digital processing model, the performance evaluation parameters of the new version model are compared with the performance evaluation parameters of the digital processing model currently in use (the performance evaluation parameters of the current version model), and the performance parameters such as the accuracy, recall rate, and F1 score of the two are comprehensively compared. If the comparison result of the performance parameters indicates that the performance of the model after fine-tuning is better than the performance of the model before fine-tuning, the new version of the digital processing model is released to the digital processing platform to achieve the optimization update of the digital processing model currently used. If the comparison result of the performance parameters indicates that the performance of the new version model is not as good as the performance of the current version model, indicating that the performance of the new version model has decreased, the new version of the digital processing model is not released, and the new version of the digital processing model is used as the current version of the digital processing model, and the training data set regenerated based on the updated data set generation strategy is used to train it for fine-tuning training, and the data set generation strategy is updated according to the performance evaluation parameters of the new version model after fine-tuning training, until the performance evaluation parameters of the new version model are better than the performance evaluation parameters of the current version model, or the number of cycles reaches a threshold. The threshold of the number of cycles may be a smaller number such as 1 or 2, or may be customized according to actual application requirements, and is not specifically limited in the embodiments of the present invention.

[0085] In one embodiment of the present invention, for further explanation and limitation, the step of updating the data set generation strategy according to the processing result and the manual correction information includes:

[0086] Quantitatively process the manual correction information to obtain reward values ​​corresponding to different data types and / or different processing items;

[0087] Reinforcement learning is performed on the parameters of the data set generation strategy according to the processing result and the reward value to update the data set generation strategy.

[0088] In the embodiment of the present invention, the update of the data set generation strategy is based on reinforcement learning. The reinforcement learning system is an artificial intelligence system that learns strategies through interaction with the environment to achieve specific goals. The core idea of ​​the reinforcement learning system is that the intelligent agent interacts with the environment by performing actions and adjusts the strategy according to the rewards obtained to maximize the cumulative rewards. In the reinforcement learning process, the optimal parameters of the data set generation strategy are the goal of reinforcement learning, and the processing results and manual correction information are used as the rewards of reinforcement learning. Figure 3 As shown in the figure, the reward is the feedback obtained by the agent from the environment, which can characterize the processing ability of the digital processing model trained by the training data set generated by the parameters of the current data set generation strategy for a certain data type, or the performance in a certain processing operation. The reward can be a positive value (indicating good) or a negative value (indicating bad). The goal of introducing the reward is to maximize the cumulative reward through learning. The manual correction information (corresponding to the evaluation result in the figure) includes processing error prompts and data type missing information. The processing error prompt is one or more processing items with processing problems in digital processing, such as home page problems, layout recognition problems, metadata recognition problems, pages and types with errors, etc. The data type missing information is the data type not included in the original data set. For example, the missing data type is the type of a journal, the doctoral and master's thesis of a certain college, and the book published by a certain publishing house. The manual correction information is quantified by representing a missing data type in numerical form as a negative reward value, representing an identified error problem in numerical form as a negative reward value, and representing a correct processing item in numerical form as a positive reward value. The size of the specific reward value can be customized according to actual application requirements to establish a mapping relationship between different items and corresponding reward values, and the embodiment of the present invention does not make specific limitations.

[0089] It should be noted that the parameters of the data set generation strategy include the number of data sets, the amount of data of different data types, and the number of data labels. Based on the update of these parameters, the proportion of samples of different data types in the training data set can be adjusted. Under each layout, the number of labels that need to be indexed is the statistical information of the elements. For example, taking a journal as an example, the data set required by the automatic indexing model is the metadata of the entire article, including text blocks with reading order, attributes corresponding to the text blocks, font size of the text blocks, block spacing and other labels, so as to index the Chinese and English titles, Chinese and English abstracts, Chinese and English keywords, Chinese and English author institutions, references, notes, first, second, third and fourth level subheadings, Chinese and English figure captions, and Chinese and English table titles in the article during processing. For another example, the multimodal layout recognition model is used to identify the title, header, footer, chapter, paragraph, picture, table, formula, chart caption, column, footer annotation and other areas in the document page. The data set is the "title, header, footer, chapter, paragraph, picture, table, formula, chart caption, column and footer annotation" area information and text information in the document page and the corresponding page. The data set format can be any one of xml, json, and txt, and the embodiment of the present invention does not make specific limitations. The parameters of the data set generation strategy are updated by reinforcement learning. While realizing the automatic generation of the training data set, it ensures that the content of the training data set accurately adapts to the changes in the digital processing environment, improves the efficiency of the training data set generation, and ensures the timeliness of the adaptive adjustment of the digital processing model, thereby improving the adaptability of the digital processing model to data changes and the accuracy of data processing.

[0090] In one embodiment of the present invention, for further explanation and limitation, the step of regenerating the training data set according to the updated data set generation strategy includes:

[0091] Acquire an original training data set, and add, delete, or modify data types or data contents of the original training data set according to the updated data set generation strategy;

[0092] Filtering new training data from the positive data and negative data according to the updated data set generation strategy;

[0093] A training data set is generated according to the addition, deletion and modification results of the newly added training data and the original training data set.

[0094] In the embodiment of the present invention, Figure 3As shown, in the cycle of reinforcement learning, the data type or data content of the original training data set is added, deleted, and modified as the action of the reinforcement learning system on the digital processing model. The adjustment of the digital processing model is achieved by adjusting the training data set and fine-tuning the model. For example, if the reward value of a certain data type is detected to be high, some data samples under this data type can be deleted from the data set. If the reward value of another data type is detected to be negative and the data type is missing, it is necessary to add this data type to the data set. Similarly, the data content is also the logic of deletion and addition. If the cause of the processing error is the format update of a certain data content, it is necessary to modify the corresponding data content in the data set. Among them, the original training data set is the training data set used in the last fine-tuning training of the digital processing model. If the currently used digital processing model is a basic model that has not been fine-tuned, then this training data set is a data set based on manual marking. If the currently used digital processing model is a model that has undergone fine-tuning training, then this training data set is a data set automatically generated based on the data set generation strategy.

[0095] The processing results include positive data and negative data. Positive data is the correct processing result obtained by the digital processing model, and negative data is the wrong processing result obtained by the digital processing model, which is corrected by manual correction. Negative data is the content of processing errors, and the cause of the error needs to be analyzed to identify whether this data needs to be added as a sample to the training data set to train the processing ability of the model.

[0096] In one embodiment of the present invention, for further explanation and limitation, the step of updating the data set generation strategy according to the performance evaluation parameters of the new version model to obtain a further updated data set generation strategy includes:

[0097] Obtain regenerated training data sets and processing result feedback information;

[0098] The performance evaluation parameters of the new version model, the regenerated training data set and the processing result feedback information are used as feedback states to perform reinforcement learning on the parameters of the data set generation strategy to obtain a further updated data set generation strategy.

[0099] In the embodiment of the present invention, Figure 3 As shown, the performance evaluation parameters of the new version of the model, the regenerated training data set and the processing result feedback information are fed back to the reinforcement learning system (agent) as state, so as to conduct secondary reinforcement learning on the basis of reward-based reinforcement learning, and update the data set generation strategy again, thereby realizing the optimization of the fine-tuning training data set of the digital processing model based on two-way feedback reinforcement learning, making the fine-tuning training of the model more targeted, so as to obtain a more adaptable digital processing model.

[0100] In one embodiment of the present invention, for further explanation and limitation, the digital processing model includes an OCR recognition model, a layout document automatic indexing model, a table detection model, a table structuring model, a formula detection model and a formula structuring model.

[0101] In an embodiment of the present invention, the above-mentioned digital processing method is applicable to any processing method in any digital processing link. That is to say, the two-way feedback tuning method of the above-mentioned digital processing model is applicable to the digital processing model used in each processing method, specifically including OCR recognition model, layout document automatic indexing model, table detection model, formula detection model, table structuring model and formula structuring model at the fragmentation level. Among them, the layout document automatic indexing model further includes a multimodal layout recognition sub-model and an automatic indexing sub-model based on short text classification. Each digital processing link can be operated independently or as a whole, and the processing results of each independently operated digital processing link can be applied to specific application scenarios.

[0102] The present invention provides a digital processing method based on a two-way feedback tuning model. The embodiment of the present invention obtains the data to be processed, the target processing link and the processing method by analyzing the digital processing task, wherein the target processing link includes any processing link in the whole process of digital processing; the digital processing model matching the processing method is identified from the digital processing platform server, wherein the digital processing model is obtained by tuning the training data set updated based on the historical manual correction information of the historical processing process and the historical model performance evaluation parameters; the data to be processed is processed using the digital processing model, and the obtained initial processing result is sent to the manual correction node; the manual correction information of the initial processing result is obtained, and the processing result of the data to be processed is generated based on the manual correction information and the initial processing result. Based on the historical manual correction information of the historical processing process and the historical model performance evaluation parameters, the generation strategy of the model training data set is optimized from two directions to realize the automatic generation of the training data set, which greatly reduces the time of digital processing model tuning. At the same time, the accuracy of the digital processing model tuning is ensured, so that the digital processing model can timely and accurately adapt to the changes in digital processing needs, thereby improving the accuracy of digital processing.

[0103] In one example, if Figure 4As shown in the figure, the tuning of the digital processing model is a continuous process that accompanies digital processing. Each digital processing result and manual correction result will be used as reinforcement learning feedback and input into the reinforcement learning system, so that the reinforcement learning system can update the data set generation strategy. The training data set regenerated by the updated data set generation strategy can fine-tune the digital processing model. The evaluation parameters of the trained model are again input into the reinforcement learning system as reinforcement learning feedback, so that the reinforcement learning system can update the data set generation strategy again, thereby realizing the automatic tuning of the digital processing model.

[0104] Furthermore, as a response to the above Figure 1 The implementation of the method shown in the figure, the embodiment of the present invention provides a digital processing device based on a two-way feedback tuning model, such as Figure 5 As shown, the device comprises:

[0105] The analysis module 31 is used to analyze the digital processing task to obtain the data to be processed, the target processing link and the processing method, wherein the target processing link includes any processing link in the whole process of digital processing;

[0106] A determination module 32 is used to identify a digital processing model matching the processing method from a digital processing platform server, wherein the digital processing model is obtained by tuning a training data set based on historical manual correction information of a historical processing process and updated with historical model performance evaluation parameters;

[0107] A processing module 33, used to process the data to be processed using the digital processing model, and send the obtained initial processing result to the manual correction node;

[0108] The acquisition module 34 is used to acquire the manual correction information of the initial processing result, and generate the processing result of the data to be processed according to the manual correction information and the initial processing result.

[0109] Furthermore, the device also includes:

[0110] A first updating module, used for updating a data set generation strategy according to the processing result and the manual correction information, and regenerating a training data set according to the updated data set generation strategy;

[0111] A first training module is used to train the current version of the digital processing model using the training data set to obtain a new version of the digital processing model;

[0112] The second updating module is used to perform model performance evaluation on the new version of the digital processing model to obtain performance evaluation parameters of the new version of the model, and update the data set generation strategy based on the performance evaluation parameters of the new version of the model and the performance evaluation parameters of the current version of the model to obtain a further updated data set generation strategy.

[0113] Furthermore, the device also includes:

[0114] A publishing module, used for publishing the new version of the digital processing model to the digital processing platform when the performance evaluation parameters of the new version of the model are better than the performance evaluation parameters of the current version of the model;

[0115] The second training module is used to regenerate the training data set according to the again updated data set generation strategy when the performance evaluation parameters of the new version model are inferior to the performance evaluation parameters of the current version model, and return to the step of using the training data set to train the current version of the digital processing model to obtain a new version of the digital processing model, and iteratively update the data set generation strategy until the performance evaluation parameters of the new version model are better than the performance evaluation parameters of the current version model, or the number of cycles reaches a threshold.

[0116] Furthermore, the first update module includes:

[0117] A processing unit, configured to quantify the manual correction information to obtain reward values ​​corresponding to different data types and / or different processing items, wherein the manual correction information includes processing error prompts and / or data type missing information;

[0118] The first reinforcement learning unit is used to perform reinforcement learning on the parameters of the data set generation strategy according to the processing result and the reward value to update the data set generation strategy, wherein the parameters include the number of data sets, the amount of data of different data types and the number of data labels.

[0119] Furthermore, the first update module further includes:

[0120] A first acquisition unit is used to acquire an original training data set, and add, delete, or modify data types or data contents of the original training data set according to the updated data set generation strategy;

[0121] A screening unit, configured to screen out newly added training data from the positive data and the negative data according to the updated data set generation strategy;

[0122] A generating unit is used to generate a training data set according to the addition, deletion and modification results of the newly added training data and the original training data set.

[0123] Furthermore, the second update module includes:

[0124] A second acquisition unit, used to acquire the regenerated training data set and processing result feedback information;

[0125] The second reinforcement learning unit is used to perform reinforcement learning on the parameters of the data set generation strategy by using the performance evaluation parameters of the new version model, the regenerated training data set and the processing result feedback information as feedback states to obtain a further updated data set generation strategy.

[0126] Furthermore, in a specific application scenario, the digital processing model in the processing module includes an OCR recognition model, a layout document automatic indexing model, a table detection model, a table structuring model, a formula detection model and a formula structuring model.

[0127] The present invention provides a digital processing device based on a two-way feedback tuning model. The embodiment of the present invention obtains the data to be processed, the target processing link and the processing method by analyzing the digital processing task, wherein the target processing link includes any processing link in the whole process of digital processing; identifies the digital processing model matching the processing method from the digital processing platform server, wherein the digital processing model is obtained by tuning the training data set updated based on the historical manual correction information of the historical processing process and the historical model performance evaluation parameters; uses the digital processing model to process the data to be processed, and sends the obtained initial processing result to the manual correction node; obtains the manual correction information of the initial processing result, and generates the processing result of the data to be processed based on the manual correction information and the initial processing result. Based on the historical manual correction information of the historical processing process and the historical model performance evaluation parameters, the generation strategy of the model training data set is optimized from two directions to realize the automatic generation of the training data set, which greatly reduces the time of digital processing model tuning. At the same time, it ensures the accuracy of digital processing model tuning, so that the digital processing model can timely and accurately adapt to the changes in digital processing needs, thereby improving the accuracy of digital processing.

[0128] According to one embodiment of the present invention, a storage medium is provided, wherein the storage medium stores at least one executable instruction, and the computer executable instruction can execute the digital processing method based on the bidirectional feedback tuning model in any of the above method embodiments.

[0129] Figure 6 A schematic diagram of the structure of a terminal provided according to an embodiment of the present invention is shown. The specific embodiment of the present invention does not limit the specific implementation of the terminal.

[0130] like Figure 6As shown, the terminal may include: a processor (processor) 402 , a communication interface (Communications Interface) 404 , a memory (memory) 406 , and a communication bus 408 .

[0131] The processor 402 , the communication interface 404 , and the memory 406 communicate with each other via a communication bus 408 .

[0132] The communication interface 404 is used for network communication with other devices such as a client or other servers.

[0133] The processor 402 is used to execute the program 410, and specifically can execute the relevant steps in the above-mentioned embodiment of the digital processing method based on the two-way feedback tuning model.

[0134] Specifically, the program 410 may include program codes, which include computer operation instructions.

[0135] The processor 402 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiment of the present invention. The one or more processors included in the terminal may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0136] The memory 406 is used to store the program 410. The memory 406 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0137] The program 410 may be specifically configured to enable the processor 402 to perform the following operations:

[0138] Analyze the digital processing task to obtain the data to be processed, the target processing link and the processing method, wherein the target processing link includes any processing link in the whole process of digital processing;

[0139] Identifying a digital processing model matching the processing method from a digital processing platform server, wherein the digital processing model is obtained by tuning a training data set based on historical manual correction information of historical processing processes and updated historical model performance evaluation parameters;

[0140] Processing the data to be processed using the digital processing model, and sending the obtained initial processing result to the manual correction node;

[0141] The manual correction information of the initial processing result is obtained, and the processing result of the to-be-processed data is generated according to the manual correction information and the initial processing result.

[0142] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A digital processing method based on a two-way feedback tuning model, characterized in that: include: Analyze the digital processing task to obtain the data to be processed, the target processing link and the processing method, wherein the target processing link includes any processing link in the whole process of digital processing; Identifying a digital processing model matching the processing method from a digital processing platform server, wherein the digital processing model is obtained by tuning a training data set based on historical manual correction information of historical processing processes and updated historical model performance evaluation parameters; Processing the data to be processed using the digital processing model, and sending the obtained initial processing result to the manual correction node; The manual correction information of the initial processing result is obtained, and the processing result of the to-be-processed data is generated according to the manual correction information and the initial processing result.

2. The method according to claim 1, characterized in that After generating the processing result of the data to be processed according to the manual correction information and the initial processing result, the method further includes: updating the data set generation strategy according to the processing result and the manual correction information, and regenerating the training data set according to the updated data set generation strategy; Using the training data set to train the current version of the digital processing model to obtain a new version of the digital processing model; A model performance evaluation is performed on the new version of the digital processing model to obtain performance evaluation parameters of the new version of the model, and a data set generation strategy is updated according to the performance evaluation parameters of the new version of the model and the performance evaluation parameters of the current version of the model to obtain a further updated data set generation strategy.

3. The method according to claim 2, characterized in that After obtaining the updated data set generation strategy, the method further includes: When the performance evaluation parameters of the new version model are better than the performance evaluation parameters of the current version model, publishing the new version of the digital processing model to the digital processing platform; When the performance evaluation parameters of the new version model are inferior to the performance evaluation parameters of the current version model, the training data set is regenerated according to the updated data set generation strategy, and the step of using the training data set to train the current version of the digital processing model to obtain the new version of the digital processing model is returned to iteratively update the data set generation strategy until the performance evaluation parameters of the new version model are superior to the performance evaluation parameters of the current version model, or the number of cycles reaches a threshold.

4. The method according to claim 2, characterized in that: The updating of the data set generation strategy according to the processing result and the manual correction information includes: Quantifying the manual correction information to obtain reward values ​​corresponding to different data types and / or different processing items, wherein the manual correction information includes processing error prompts and / or data type missing information; Reinforcement learning is performed on the parameters of the data set generation strategy based on the processing results and the reward value to update the data set generation strategy, wherein the parameters include the number of data sets, the amount of data of different data types, and the number of data labels.

5. The method according to claim 4, characterized in that The processing result includes positive data and negative data, and the regenerating the training data set according to the updated data set generation strategy includes: Acquire an original training data set, and add, delete, or modify data types or data contents of the original training data set according to the updated data set generation strategy; Filtering new training data from the positive data and negative data according to the updated data set generation strategy; A training data set is generated according to the addition, deletion and modification results of the newly added training data and the original training data set.

6. The method according to claim 2, characterized in that The updating of the data set generation strategy according to the performance evaluation parameters of the new version model to obtain a further updated data set generation strategy includes: Obtain regenerated training data sets and processing result feedback information; The performance evaluation parameters of the new version model, the regenerated training data set and the processing result feedback information are used as feedback states to perform reinforcement learning on the parameters of the data set generation strategy to obtain a further updated data set generation strategy.

7. The method according to any one of claims 1 to 6, characterized in that The digital processing model includes an OCR recognition model, a layout document automatic indexing model, a table detection model, a table structuring model, a formula detection model and a formula structuring model.

8. A digital processing device based on a two-way feedback tuning model, characterized in that: include: An analysis module is used to analyze the digital processing task to obtain the data to be processed, the target processing link and the processing method, wherein the target processing link includes any processing link in the whole process of digital processing; A determination module, configured to identify a digital processing model matching the processing method from a digital processing platform server, wherein the digital processing model is obtained by tuning a training data set based on historical manual correction information of a historical processing process and updated with historical model performance evaluation parameters; A processing module, used to process the data to be processed using the digital processing model, and send the obtained initial processing result to the manual correction node; The acquisition module is used to acquire the manual correction information of the initial processing result, and generate the processing result of the data to be processed according to the manual correction information and the initial processing result.

9. A storage medium storing at least one executable instruction, wherein the executable instruction enables a processor to execute operations corresponding to the digital processing method based on the bidirectional feedback tuning model as described in any one of claims 1 to 7.

10. A terminal, comprising: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the digital processing method based on the bidirectional feedback tuning model as described in any one of claims 1 to 7.

Images