Cross-language information processing method based on digital modeling

Abstract

The invention provides a cross-language information processing method based on digital modeling, and relates to the technical field of data processing, and the method comprises the steps: collecting an ancient language training corpus set with a word number label based on a digital humanity database; pre-training the RoBERTa model to generate a language information processing pre-training model; in order to meet the cross-language requirement, the method comprises the following steps: converting text understanding into images or videos, and constructing a cross-language training data set; finally, a cross-language information processing model is obtained by finely adjusting the pre-training model, and cross-language demand processing is achieved. The method can solve the problems that in the prior art, due to the fact that the performance is poor during ancient language processing, it is difficult to generate images and videos in combination with the specific context of a whole paragraph or a long sentence, the efficiency of executing cross-language tasks is low, the universality and adaptability of a model during processing diversified languages and tasks are further affected, and the processing efficiency is poor. Efficient processing of ancient texts and cross-language tasks is realized, and the understanding and processing efficiency of the model on the ancient texts is remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a cross-language information processing method based on digital modeling. Background Art

[0002] Cross-language information processing plays a key role in various natural language processing tasks such as machine translation, sentiment analysis, text summarization, semantic role labeling, etc. For example, in the prediction of background advertising exposure parameters for sports events, it is necessary to model and analyze advertising data in multiple languages.

[0003] Currently, existing technologies usually only mask one character, which is suitable for understanding the context of short sentences. However, in cross-language processing, it is necessary to generate images and videos in the specific context of entire paragraphs or long sentences, and traditional models are difficult to apply.

[0004] In summary, the existing technology performs poorly when processing ancient texts, and it is difficult to generate images and videos based on the specific context of entire paragraphs or long sentences, resulting in low efficiency in performing cross-language tasks, which further affects the versatility and adaptability of the model when processing diverse languages ​​and tasks. Summary of the invention

[0005] The purpose of this application is to provide a cross-language information processing method based on digital modeling to solve the problem that the existing technology performs poorly when processing ancient texts and has difficulty in generating images and videos in the specific context of entire paragraphs or long sentences, resulting in low efficiency in performing cross-language tasks, further affecting the versatility and adaptability of the model when processing diverse languages ​​and tasks.

[0006] In view of the above problems, the present application provides a cross-language information processing method based on digital modeling.

[0007] The present application provides a cross-language information processing method based on digital modeling, wherein the cross-language information processing method based on digital modeling includes: based on the digital humanities database corpus, collecting an ancient Chinese training corpus set, wherein the ancient Chinese training corpus set has a word count label set; based on the word count label set, configuring a noise quantity label set for the ancient Chinese training corpus set; based on the ancient Chinese training corpus set and the noise quantity label set, pre-training the RoBERTa model structure to generate a language information processing pre-training model; based on cross-language demand information, constructing a cross-language training data set, wherein the cross-language demand information refers to the process of generating images and / or videos based on text understanding; based on the cross-language training data set, fine-tuning the language information processing pre-training model to obtain a cross-language information processing model, and performing cross-language demand processing.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] Based on the digital humanities database corpus, an ancient Chinese training corpus is collected, wherein the ancient Chinese training corpus has a word count label set; based on the word count label set, a noise quantity label set is configured for the ancient Chinese training corpus; according to the ancient Chinese training corpus and the noise quantity label set, a RoBERTa model structure is pre-trained to generate a language information processing pre-training model; according to cross-language demand information, a cross-language training data set is constructed, wherein the cross-language demand information refers to the process of generating images and / or videos based on text understanding; according to the cross-language training data set, the language information processing The pre-trained model is fine-tuned to obtain a cross-language information processing model for cross-language demand processing, which effectively solves the problem that the existing technology performs poorly in processing ancient texts and is difficult to generate images and videos in the specific context of a whole paragraph or long sentence, resulting in low efficiency in performing cross-language tasks, further affecting the versatility and adaptability of the model in processing diverse languages ​​and tasks. The RoBERTa model is pre-trained by collecting ancient text training corpus based on the digital humanities database corpus and configuring word count label sets and noise number label sets for it, which significantly improves the model's understanding and processing capabilities for ancient texts. At the same time, a cross-language training dataset is constructed based on cross-language demand information, and the pre-trained model is fine-tuned, so that the model has the ability to perform cross-language tasks (such as generating images or videos based on text understanding). The versatility and adaptability of the model are enhanced, and its processing performance in the presence of noise is improved, achieving efficient processing of ancient texts and cross-language tasks.

[0010] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically cited below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0012] Figure 1 A flowchart of a cross-language information processing method based on digital modeling in this application;

[0013] Figure 2 This is a schematic diagram of the process of generating a cross-language information processing model in the cross-language information processing method based on digital modeling in this application. DETAILED DESCRIPTION

[0014] This application provides a cross-language information processing method based on digital modeling, which solves the problem that the prior art performs poorly when processing ancient Chinese and has difficulty in generating images and videos in the specific context of entire paragraphs or long sentences, resulting in low efficiency in performing cross-language tasks, further affecting the versatility and adaptability of the model when processing diverse languages ​​and tasks. It achieves efficient processing of ancient Chinese and cross-language tasks, and significantly improves the model's understanding and processing efficiency of ancient Chinese.

[0015] Below, the technical solutions in the present application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments of the present application. It should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application. It should also be noted that, for the convenience of description, only the parts related to the present application are shown in the accompanying drawings, rather than all of them.

[0016] Please refer to the attached Figure 1 The present application provides a cross-language information processing method based on digital modeling, wherein the cross-language information processing method based on digital modeling specifically includes the following steps:

[0017] S1: Based on the digital humanities database corpus, collect an ancient Chinese text training corpus set, wherein the ancient Chinese text training corpus set has a word count label set.

[0018] Specifically, the digital humanities database corpus is one or more digital humanities databases containing rich ancient Chinese resources. Determine the types of ancient Chinese that need to be collected, such as pre-Qin literature, history books, poems, songs, etc. and the time range. Use the search tools provided by the database to enter keywords or select categories for retrieval according to the collection target. Clean the downloaded ancient Chinese documents and remove irrelevant information, such as headers, footers, and comments. Use text processing tools, such as Python's jieba word library and regular expressions, to count the number of words in the cleaned ancient Chinese documents, count the total number of words in each document, and count the number of words in paragraphs or sentences as needed. Generate a word count label for each document based on the word count results. The label can be a specific number, such as within 100 words, 100-200 words, etc., or a continuous range of numbers. Associate the word count label with the corresponding ancient Chinese document to form an ancient Chinese training corpus with word count labels.

[0019] S2: Based on the word count tag set, configure a noise quantity tag set for the ancient Chinese training corpus.

[0020] Specifically, a reasonable noise interference range is set for ancient Chinese texts with different word counts. Noise is the randomness, error or uncertainty that may be introduced during text processing, such as character replacement, deletion, insertion, etc. One or more noise ratios are defined, which will be used to determine how much noise can be introduced into a text with a specific word count. For example, a basic noise ratio can be set to 5%, that is, out of every 100 characters, a maximum of 5 characters may be affected. For each ancient Chinese text, the amount of noise that can be introduced is determined based on its word count label. For example, if the word count label of a text is 100-200 words, then based on the basic noise ratio of 5%, a noise quantity label can be configured for it, such as 5-10 noise characters. Traverse the entire ancient text training corpus and generate a corresponding noise quantity label for each text. Associate these labels with the original ancient text to form a training corpus with word count labels and noise quantity labels.

[0021] S3: Pre-train the RoBERTa model structure based on the ancient Chinese training corpus and the noise quantity label set to generate a language information processing pre-training model.

[0022] Specifically, the ancient text training corpus is integrated with the corresponding noise quantity label set to ensure that each ancient text has its corresponding noise quantity label. The ancient text is preprocessed as necessary, such as word segmentation and stop word removal. According to the noise quantity label, a version with the corresponding amount of noise is generated for each ancient text, which can be achieved by randomly replacing, deleting or inserting characters. Using the existing RoBERTa model as the basis for pre-training, the model input layer is modified so that it can accept noisy ancient text as input. According to the task requirements, a suitable output layer is set. For example, if the task is to perform text classification, the output layer should be a classifier. Set the parameters in the training process, such as learning rate, batch size, training rounds, etc. The RoBERTa model is pre-trained using the integrated training dataset, including the original ancient text and the noisy ancient text. During the training process, the model will learn the ability to recover the original text from the noisy input, thereby enhancing the robustness to noise. The performance of the model on the training set and the validation set is regularly evaluated to ensure that the model is not overfitting or underfitting. After the training is completed, the pre-trained model is saved for subsequent fine-tuning or use. Test the performance of the pre-trained model on the test set to ensure that it has good generalization ability.

[0023] S4: constructing a cross-language training dataset based on the cross-language demand information, wherein the cross-language demand information refers to a process of generating images and / or videos based on text understanding.

[0024] Specifically, cross-language requirement information includes whether images or videos are to be generated, or both; and which language the input text description is in and whether it is necessary to support multiple languages. Analyze the differences in grammar, vocabulary, and expressions between different languages ​​to determine whether it is necessary to deal with ambiguity or cultural differences between languages. Collect text descriptions in different languages ​​and ensure that the text descriptions cover a wide range of topics and scenarios. Collect corresponding images or videos based on the text descriptions. If possible, collect images or videos that completely match the text descriptions. Otherwise, collect images or videos that are related to the text descriptions or can trigger the same emotional response. Remove images or videos that are irrelevant to the task or of poor quality, clean the text descriptions, remove irrelevant information, or perform necessary translations. Pair each text description with its corresponding image or video. If a text description corresponds to multiple images or videos, they can be regarded as different training samples. Divide the dataset into training, validation, and test sets. Ensure that the data in each set is evenly distributed and covers different topics and scenarios. If necessary, annotate the data, such as annotating objects, scenes, or emotions in the image. Transform images or videos, such as rotation, scaling, and cropping, to increase data diversity, and perform synonym replacement and sentence transformation on text descriptions to generate more training samples. If the amount of data for certain languages ​​or scenarios is small, data augmentation or sampling strategies can be used to balance the dataset.

[0025] S5: Fine-tune the language information processing pre-training model according to the cross-language training data set to obtain a cross-language information processing model and perform cross-language demand processing.

[0026] Specifically, during the fine-tuning process, you can choose to freeze some layers of the pre-trained model and only adjust the layers related to the task. This can speed up the training process and prevent overfitting. According to the size of the cross-language training dataset and the complexity of the task, set appropriate training parameters such as learning rate, batch size, training rounds, etc. Use the cross-language training dataset to fine-tune the pre-trained model. During the training process, the model will learn how to generate corresponding images or videos based on text descriptions in different languages. Regularly evaluate the performance of the model on the training set and validation set to ensure that the model is not overfitting or underfitting. Adjust the training parameters or model structure as needed. After the fine-tuning is completed, save the trained model to obtain a cross-language information processing model. This model can generate corresponding images or videos based on text descriptions in different languages, meeting the task requirements of cross-language demand processing. Use the obtained cross-language information processing model for actual cross-language demand processing. For example, you can enter a text description in French, and the model will generate the corresponding image or video. In this way, the model can generate images or videos between different languages ​​and realize cross-language information processing.

[0027] Furthermore, step S2 of the present application also includes:

[0028] Configure a benchmark word count; extract a first word count label of a first classical Chinese training corpus according to the classical Chinese training corpus and the word count label set; when the first word count label is less than or equal to the benchmark word count, configure a first noise quantity label to 1; when the first word count label is greater than the benchmark word count, configure the first noise quantity label to be a ratio of the first classical Chinese training corpus to the benchmark word count; add the first noise quantity label to the noise quantity label set.

[0029] Specifically, a reference number of words is set by a person skilled in the art, and this number of words will be used as a reference for judging the number of words in the ancient text training corpus. For example, the reference number of words can be set to 100 words. Next, select the first ancient text from the ancient text training corpus and extract its corresponding word count label. Assume that the word count label of this ancient text is 80 words. Now, configure a noise quantity label for it according to the word count label and the reference number of words of this ancient text. If the word count label of this ancient text, 80 words, is less than or equal to the reference number of words, 100 words, the first noise quantity label is configured as 1. It means that 1 noise character will be introduced on the basis of this ancient text. If the word count label of this ancient text is greater than the reference number of words, the first noise quantity label is configured as the ratio of this ancient text to the reference number of words. For example, if this ancient text has 120 words, then the ratio is 120 / 100=1.2, and this ratio can be rounded or rounded as a noise quantity label, such as 1 or 2 noise characters. Finally, the configured first noise quantity label is added to the noise quantity label set.

[0030] Furthermore, step S3 of the present application also includes:

[0031] Construct a noise adding plug-in, wherein the noise adding plug-in is used to add noise to the ancient Chinese training corpus according to the noise quantity label set, and adding noise is to randomly replace the content of the ancient Chinese training corpus with a preset identifier, replacing it with a noise quantity at a time; according to the noise adding plug-in, based on the noise quantity label set, the ancient Chinese training corpus is retrieved, the RoBERTa model structure is unsupervised pre-trained, and the language information processing pre-training model is generated.

[0032] Specifically, the noise adding plug-in can receive the ancient text training corpus and the noise quantity label set as input. The plug-in contains a preset label library for randomly replacing the content in the ancient text training corpus. For each ancient text in the ancient text training corpus, the plug-in determines the amount of noise to be added based on its corresponding noise quantity label. Use the labels in the preset label library to randomly replace the ancient text content, and one replacement counts as one noise quantity. During the replacement process, it should be ensured that the replaced labels are similar in length to the original content to avoid causing excessive interference to the model training. The plug-in outputs the ancient text training corpus with noise for subsequent model pre-training. Make sure that the ancient text training corpus and the noise quantity label set are prepared and in the correct format. Run the noise adding plug-in, take the ancient text training corpus and the noise quantity label set as input, and the plug-in will output the ancient text training corpus with noise. Use the existing RoBERTa model as the basis for pre-training. Set the parameters in the training process, such as learning rate, batch size, training rounds, etc., and use the ancient text training corpus with noise to perform unsupervised pre-training on the RoBERTa model. During the training process, the model will learn to recover the original text from the noisy input, thereby enhancing the robustness to noise. Regularly evaluate the performance of the model on the training set to ensure that the model is not overfitting. After the training is completed, save the pre-trained model, that is, the language information processing pre-trained model.

[0033] Furthermore, step S3 of the present application also includes:

[0034] Construct a multi-noise training loss function: Among them, LOSS1 represents the loss value of any training, SIM0 represents the cosine similarity threshold, and a i Characterizes the predicted semantics of the i-th noise, a i0 Characterizes the truth semantics of the i-th noise, SIM(a i ,a i0 ) represents the cosine similarity between the predicted semantic vector and the true semantic vector, N represents the number of noises in any training, and c is a small constant to avoid =0; according to the noise adding plug-in, based on the noise quantity label set, combined with the multi-noise training loss function, the ancient Chinese training corpus is retrieved, the RoBERTa model structure is unsupervised pre-trained, and the language information processing pre-training model is generated.

[0035] Specifically, this loss function calculates the cosine similarity between the predicted semantics and the true semantics of each noise and compares it with the threshold. If the similarity is lower than the threshold, a loss is incurred, and the size of the loss is proportional to the difference between the similarity and the threshold. Run the noise addition plug-in, take the ancient text training corpus and the noise quantity label set as input, and the plug-in will output the ancient text training corpus with noise. Use the existing RoBERTa model as the basis for pre-training. Set the parameters in the training process, such as learning rate, batch size, training rounds, etc. Use the ancient text training corpus with noise for unsupervised pre-training of the RoBERTa model. During the training process, use the multi-noise training loss function to calculate the loss, and update the model parameters to minimize the loss. Regularly evaluate the performance of the model on the training set to ensure that the model is not overfitting. After the training is completed, save the pre-trained model, that is, the language information processing pre-training model.

[0036] Further, if Figure 2 As shown, step S5 of the present application also includes:

[0037] The cross-language training dataset includes a classical Chinese sentence training dataset, a vernacular Chinese sentence training dataset and a cross-language target dataset; the language information processing pre-training model is fine-tuned with the vernacular Chinese sentence training dataset as supervision and the classical Chinese sentence training dataset as input to obtain a vernacular Chinese sentence parsing model; a convolutional neural network is configured with the vernacular Chinese sentence training dataset as input and the cross-language target dataset as supervision to obtain a cross-language generation model; the vernacular Chinese sentence parsing model and the cross-language generation model are sequentially connected in series to generate the cross-language information processing model.

[0038] Specifically, the cross-language training dataset includes a classical Chinese sentence training dataset, a vernacular Chinese sentence training dataset, and a cross-language target dataset. Load the language information processing pre-training model obtained through unsupervised pre-training. Fine-tune the pre-training model with the vernacular Chinese sentence training dataset as supervision and the classical Chinese sentence training dataset as input. During the training process, the model will learn how to convert classical Chinese sentences into vernacular Chinese sentences. After the fine-tuning training is completed, save the trained model to obtain the vernacular Chinese sentence parsing model. Configure the convolutional neural network structure for cross-language generation tasks. Train the convolutional neural network with the vernacular Chinese sentence training dataset as input and the cross-language target dataset as supervision. During the training process, the model will learn how to convert vernacular Chinese sentences into expressions in other languages. After the training is completed, save the trained convolutional neural network model to obtain the cross-language generation model, and connect the vernacular Chinese sentence parsing model and the cross-language generation model in sequence. Ensure that the input and output formats of the two models are compatible. Test the connected cross-language information processing model to ensure that it functions properly. Make necessary optimization and adjustments to the model based on the test results. After testing and optimization are completed, the concatenated model is saved to generate a cross-language information processing model.

[0039] Furthermore, the present application also includes:

[0040] A first evaluation function is configured, wherein the first evaluation function is used to calculate the mean of semantic similarity after a preset number of trainings; a second evaluation function is configured, wherein the second evaluation function is used to calculate the variance of semantic similarity after a preset number of trainings; the language information processing pre-training model is fine-tuned using the vernacular sentence training data set as supervision and the classical Chinese sentence training data set as input, and when the first evaluation function is greater than or equal to a first evaluation threshold and the second evaluation function is less than or equal to a second evaluation threshold, the vernacular sentence parsing model is output.

[0041] Specifically, the first evaluation function is used to calculate the mean of the semantic similarity of the preset number of trainings, receives a series of semantic similarities as input, and calculates their mean. The second evaluation function is used to calculate the variance of the semantic similarity of the preset number of trainings, receives a series of semantic similarities as input, and calculates their variance. Load the language information processing pre-training model obtained by unsupervised pre-training. Set the parameters in the fine-tuning training process, such as learning rate, batch size, training rounds, etc. The pre-training model is fine-tuned with the vernacular sentence training data set as supervision and the ancient Chinese sentence training data set as input. During the training process, the semantic similarity is calculated regularly, and the first evaluation function and the second evaluation function are applied for evaluation. When the mean of the semantic similarity calculated by the first evaluation function is greater than or equal to the first evaluation threshold, and the variance of the semantic similarity calculated by the second evaluation function is less than or equal to the second evaluation threshold, it is considered that the model performance meets the requirements. In this case, the fine-tuned training model is output, that is, the vernacular sentence parsing model. The output vernacular sentence parsing model is saved, and necessary tests are performed to ensure that it functions properly.

[0042] Furthermore, the present application also includes:

[0043] A cross-language generation model training plug-in is configured, wherein the cross-language generation model training plug-in is used to process image or video data for semantic recognition and generate semantic parsing sentences; when the vernacular sentence training data set is input into the convolutional neural network to generate cross-language fitting data: the cross-language fitting data is processed according to the cross-language generation model training plug-in to obtain a first semantic parsing sentence; the cross-language target data is processed according to the cross-language generation model training plug-in to obtain a second semantic parsing sentence; the semantic similarity between the first semantic parsing sentence and the second semantic parsing sentence is used as a training evaluation function; based on the training evaluation function and the cross-language generation model training plug-in, the vernacular sentence training data set is used as input, the cross-language target data set is used as supervision, a convolutional neural network is configured, and the cross-language generation model is generated.

[0044] Specifically, the cross-language generative model training plug-in can process image or video data, perform semantic recognition, and generate semantic parsing sentences. Use a programming language such as Python and a deep learning framework such as TensorFlow or PyTorch to implement the plug-in. Ensure that the plug-in can receive image or video data as input and output semantic parsing sentences. Provide the vernacular sentence training dataset as input to the convolutional neural network. The convolutional neural network should process the input data and generate cross-language fitting data corresponding to the cross-language target dataset. Use the cross-language generative model training plug-in to process the cross-language fitting data and the cross-language target data. Obtain a first semantic parsing sentence from the cross-language fitting data, and obtain a second semantic parsing sentence from the cross-language target data. Define a training evaluation function for calculating the semantic similarity between the first semantic parsing sentence and the second semantic parsing sentence. Cosine similarity, BLEU score, or other suitable semantic similarity measurement methods can be used. Determine the architecture, learning rate, batch size, training rounds, and other parameters of the convolutional neural network. Train the convolutional neural network using the vernacular sentence training dataset as input and the cross-language target dataset as supervision. When the model performance reaches a satisfactory level, save the trained convolutional neural network model to obtain a cross-language generation model. Test the generated cross-language generation model to ensure that it functions properly. Perform necessary optimization and adjustments to the model based on the test results.

[0045] Furthermore, the present application also includes:

[0046] Based on the training evaluation function and the cross-language generation model training plug-in, the vernacular sentence training data set is used as input and the cross-language target data set is used as supervision to configure a convolutional neural network. When the proportion of the number of times the output value of the training evaluation function for a preset number of consecutive trainings is greater than or equal to the output threshold of the evaluation function is greater than or equal to the number proportion threshold, the cross-language generation model is generated and the cross-language generation model training plug-in is deleted.

[0047] Specifically, load the implemented cross-language generation model training plug-in. Determine parameters such as learning rate, batch size, and training rounds. Take the vernacular sentence training dataset as input and the cross-language target dataset as supervision for training. During the training process, use the training evaluation function to regularly evaluate the performance of the model. Start the training process and record the output value of the evaluation function for each training. When the output value of the evaluation function for a preset number of consecutive trainings is greater than or equal to the output threshold of the evaluation function, the model performance is considered to be stable and good. When the above conditions are met, save the trained convolutional neural network model, that is, generate a cross-language generation model. At the same time, delete the cross-language generation model training plug-in. Test the generated cross-language generation model to ensure that it functions properly. Make necessary optimizations and adjustments to the model based on the test results.

[0048] In summary, the cross-language information processing method based on digital modeling provided by this application has the following technical effects:

[0049] Based on the digital humanities database corpus, an ancient Chinese training corpus is collected, wherein the ancient Chinese training corpus has a word count label set; based on the word count label set, a noise quantity label set is configured for the ancient Chinese training corpus; according to the ancient Chinese training corpus and the noise quantity label set, a RoBERTa model structure is pre-trained to generate a language information processing pre-training model; according to cross-language demand information, a cross-language training data set is constructed, wherein the cross-language demand information refers to the process of generating images and / or videos based on text understanding; according to the cross-language training data set, the language information processing The pre-trained model is fine-tuned to obtain a cross-language information processing model for cross-language demand processing, which effectively solves the problem that the existing technology performs poorly in processing ancient texts and is difficult to generate images and videos in the specific context of a whole paragraph or long sentence, resulting in low efficiency in performing cross-language tasks, further affecting the versatility and adaptability of the model in processing diverse languages ​​and tasks. The RoBERTa model is pre-trained by collecting ancient text training corpus based on the digital humanities database corpus and configuring word count label sets and noise number label sets for it, which significantly improves the model's understanding and processing capabilities for ancient texts. At the same time, a cross-language training dataset is constructed based on cross-language demand information, and the pre-trained model is fine-tuned, so that the model has the ability to perform cross-language tasks (such as generating images or videos based on text understanding). The versatility and adaptability of the model are enhanced, and its processing performance in the presence of noise is improved, achieving efficient processing of ancient texts and cross-language tasks.

[0050] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0051] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is also intended to include these modifications and variations.

Claims

1. A cross-language information processing method based on digital modeling, characterized in that: include: Based on the digital humanities database corpus, an ancient Chinese text training corpus set is collected, wherein the ancient Chinese text training corpus set has a word count label set; Based on the word count tag set, configuring a noise quantity tag set for the ancient Chinese text training corpus; Pre-training the RoBERTa model structure according to the ancient Chinese training corpus and the noise quantity label set to generate a language information processing pre-training model; Constructing a cross-language training data set according to the cross-language demand information, wherein the cross-language demand information refers to the process of generating images and / or videos according to text understanding; According to the cross-language training data set, the language information processing pre-training model is fine-tuned to obtain a cross-language information processing model to perform cross-language demand processing.

2. The cross-language information processing method based on digital modeling according to claim 1, characterized in that: Based on the word count tag set, configuring a noise quantity tag set for the ancient text training corpus includes: Configure the baseline word count; Extracting a first word count label of a first ancient Chinese training corpus according to the ancient Chinese training corpus set and the word count label set; When the first word count label is less than or equal to the reference word count, configuring the first noise quantity label to be 1; When the first word count label is greater than the reference word count, configuring the first noise quantity label as the ratio of the first classical Chinese training corpus to the reference word count; Adding the first noise quantity label to the noise quantity label set.

3. The cross-language information processing method based on digital modeling according to claim 1, characterized in that: According to the ancient text training corpus and the noise quantity label set, the RoBERTa model structure is pre-trained to generate a language information processing pre-training model, including: Construct a noise adding plug-in, wherein the noise adding plug-in is used to add noise to the ancient Chinese text training corpus according to the noise quantity label set, and adding noise is to randomly replace the content of the ancient Chinese text training corpus with a preset identifier, replacing it with a noise quantity at a time; According to the noise adding plug-in, based on the noise quantity label set, the ancient Chinese training corpus is retrieved, the RoBERTa model structure is unsupervised pre-trained, and the language information processing pre-training model is generated.

4. The cross-language information processing method based on digital modeling according to claim 3, characterized in that: According to the noise adding plug-in, based on the noise quantity label set, the ancient Chinese text training corpus is retrieved, the RoBERTa model structure is pre-trained, and the language information processing pre-training model is generated, including: Construct a multi-noise training loss function: Among them, LOSS1 represents the loss value of any training, SIM0 represents the cosine similarity threshold, and a i Characterizes the predicted semantics of the i-th noise, a i0 Characterizes the truth semantics of the i-th noise, SIM(a i ,a i0 ) represents the cosine similarity between the predicted semantic vector and the true semantic vector, N represents the number of noises in any training, and c is a small constant to avoid is equal to 0; According to the noise adding plug-in, based on the noise quantity label set, combined with the multi-noise training loss function, the ancient Chinese training corpus is retrieved, the RoBERTa model structure is unsupervised pre-trained, and the language information processing pre-training model is generated.

5. The cross-language information processing method based on digital modeling according to claim 1, characterized in that: Fine-tuning the language information processing pre-training model according to the cross-language training dataset to obtain a cross-language information processing model includes: The cross-language training dataset includes a classical Chinese sentence training dataset, a vernacular sentence training dataset and a cross-language target dataset; Using the vernacular sentence training data set as supervision and the classical Chinese sentence training data set as input, fine-tuning the language information processing pre-training model to obtain a vernacular sentence parsing model; Using the vernacular sentence training dataset as input and the cross-language target dataset as supervision, configuring a convolutional neural network to obtain a cross-language generation model; The vernacular sentence parsing model and the cross-language generation model are sequentially connected in series to generate the cross-language information processing model.

6. The cross-language information processing method based on digital modeling according to claim 5, characterized in that: Using the vernacular sentence training data set as supervision and the classical Chinese sentence training data set as input, fine-tuning the language information processing pre-training model to obtain a vernacular sentence parsing model, including: Configuring a first evaluation function, wherein the first evaluation function is used to calculate the mean value of semantic similarity of a preset number of trainings; Configuring a second evaluation function, wherein the second evaluation function is used to calculate the variance of semantic similarity after a preset number of trainings; The vernacular sentence training data set is used as supervision and the classical Chinese sentence training data set is used as input to perform fine-tuning training on the language information processing pre-training model. When the first evaluation function is greater than or equal to a first evaluation threshold and the second evaluation function is less than or equal to a second evaluation threshold, the vernacular sentence parsing model is output.

7. The cross-language information processing method based on digital modeling according to claim 5, characterized in that: Taking the vernacular sentence training dataset as input and the cross-language target dataset as supervision, configuring a convolutional neural network to obtain a cross-language generation model includes: Configuring a cross-language generative model training plug-in, wherein the cross-language generative model training plug-in is used to process image or video data for semantic recognition and generate semantic parsing sentences; When the vernacular sentence training data set is input into the convolutional neural network to generate cross-language fitting data: Processing the cross-language fitting data according to the cross-language generative model training plug-in to obtain a first semantic parsing sentence; Processing the cross-language target data according to the cross-language generation model training plug-in to obtain a second semantic parsing sentence; Using the semantic similarity between the first semantic analysis sentence and the second semantic analysis sentence as a training evaluation function; Based on the training evaluation function and the cross-language generation model training plug-in, the vernacular sentence training dataset is used as input, the cross-language target dataset is used as supervision, a convolutional neural network is configured, and the cross-language generation model is generated.

8. The cross-language information processing method based on digital modeling according to claim 7, characterized in that: Based on the training evaluation function and the cross-language generation model training plug-in, taking the vernacular sentence training dataset as input and the cross-language target dataset as supervision, configuring a convolutional neural network to generate the cross-language generation model, including: Based on the training evaluation function and the cross-language generation model training plug-in, the vernacular sentence training data set is used as input and the cross-language target data set is used as supervision to configure a convolutional neural network. When the proportion of the number of times the output value of the training evaluation function for a preset number of consecutive trainings is greater than or equal to the output threshold of the evaluation function is greater than or equal to the number proportion threshold, the cross-language generation model is generated and the cross-language generation model training plug-in is deleted.

Images