Mail batch processing method, system and equipment based on large model and storage medium

Abstract

The invention belongs to the technical field of artificial intelligence, and relates to a mail batch processing method, system and device based on a large model and a storage medium, and the method comprises the steps: 1) collecting mails needing batch processing, and carrying out the preprocessing; 2) performing supervision fine tuning and evaluation on the general basic large model by using the constructed mail fine tuning data set to obtain a mail large model; 3) utilizing the mail large model and a trained numerical entity identification dichotomy model to carry out association analysis on mails needing batch processing so as to obtain an associated mail cluster; and 4) by taking the associated mail cluster as a unit, carrying out batch abstract summarization on mails which need to be processed in batches by utilizing the general basic large model. Email processing efficiency and quality can be improved so as to adapt to high-speed development and complex requirements of the information era.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and relates to a method, system, device and storage medium for batch processing of emails, in particular to a method, system, device and storage medium for batch processing of emails based on a large model. Background Art

[0002] With the popularization of the Internet and the continuous expansion of enterprise business, emails have become an important way of information exchange, and the number of emails generated every day has increased explosively. These emails contain a large amount of key information, such as business progress, cooperation negotiation, problem feedback, etc. However, the manual method of processing emails is difficult to cope with such a large amount of information, resulting in low work efficiency and important information being easily ignored. At the same time, there are often correlation relationships between emails, such as a series of emails around the same project or activity. However, the existing email management systems lack effective correlation analysis means, making these correlation information unable to be fully utilized.

[0003] At present, some email processing tools provide simple email classification and search functions, but the classification is usually based on preset rules and cannot accurately identify the key content and types of emails. In terms of email summary generation, some tools use simple text extraction algorithms and select the text at the beginning or a specific position of the email as the summary. This method cannot accurately summarize the core information of the email. In the integration of related emails, most rely on manual screening and sorting, or simple keyword matching, and it is difficult to comprehensively and accurately identify related emails. For example, although some email management software can search for related emails according to keywords in the sender, recipient or subject, it cannot effectively associate emails with different keywords but related content. There are also some attempts in the prior art to process emails using machine learning techniques, but the models involved are not optimized for the scenario of batch processing of email associations, resulting in weak processing capabilities.

[0004] Therefore, in view of the above-mentioned defects existing in the prior art, it is necessary to develop a new type of method, system, device and storage medium for batch processing of emails. Summary of the Invention

[0005] In order to overcome the defects of the prior art, the present invention proposes a method, system, device and storage medium for batch processing of emails based on a large model, which can improve the efficiency and quality of email processing to adapt to the high-speed development and complex requirements of the information age.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for batch processing of emails based on a large model, characterized by comprising the following steps:

[0008] 1) Collect the emails to be processed in batches and perform preprocessing;

[0009] 2) Use the constructed email fine-tuning dataset to perform supervised fine-tuning and evaluation on the general basic large model to obtain the email large model;

[0010] 3) Use the email large model and the trained numerical entity recognition binary classification model to perform correlation analysis on the emails to be processed in batches to obtain correlated email clusters;

[0011] 4) Use the basic large model to perform batch summary and summarization on the emails to be processed in batches in units of correlated email clusters.

[0012] Preferably, step 1) specifically includes:

[0013] 11) Save the emails to be processed in batches as EML files through an automated script, and at the same time parse the email content to obtain the email sending and receiving time, sender and recipient information, body, signature information, and email attachments;

[0014] 12) Uniformly parse the data format of the obtained email content through data cleaning and remove outliers.

[0015] Preferably, step 2) specifically includes:

[0016] 21) Collect existing emails and perform preprocessing to obtain existing email data;

[0017] 22) Based on the existing email data, construct an email fine-tuning dataset. The data format in the email fine-tuning dataset is based on the instruction - input - output standard triple template. The instruction is used to set the prompt words. The input includes the email sending and receiving time, sender and recipient information, body, and attachment content. The output includes the email type, urgency level, body summary, key entity information, and questions that need to be replied to;

[0018] 23) Use the email fine-tuning dataset to perform supervised fine-tuning on the general basic large model to obtain the supervised fine-tuned basic large model;

[0019] 24) Based on the output of the supervised fine-tuned basic large model and the labeled answers, evaluate the supervised fine-tuned basic large model, and use the supervised fine-tuned basic large model that passes the evaluation as the email large model.

[0020] Preferably, step 3) specifically includes:

[0021] 31) Use the email large model to process each of the preprocessed emails to be batched to obtain the key entity information of the emails;

[0022] 32) Filter the key entity information through the trained binary classification model for numerical entity recognition to filter out the numerical key entity information and obtain non-numerical key entity information;

[0023] 33) Based on the non-numerical key entity information, perform string concatenation and convert the concatenated string into an embedding vector through an embedding technique;

[0024] 34) Based on the non-numerical key entity information and the embedding vector, perform correlation analysis on the emails to be batch processed to obtain correlated email clusters.

[0025] Preferably, the trained binary classification model in step 32) is obtained in the following manner:

[0026] 321) Collect entity data including numerical entity information and non-numerical entity information, and divide the entity data into a training set, a validation set, and a test set;

[0027] 322) Add a fully connected layer with a binary classification SoftMax structure on the basis of the BERT model to obtain a binary classification model for numerical entity recognition;

[0028] 323) Use the training set, the validation set, and the test set to train and update the model parameters of the binary classification model for numerical entity recognition to obtain a trained binary classification model for numerical entity recognition.

[0029] Preferably, when using the training set, the validation set, and the test set to train and update the model parameters of the binary classification model for numerical entity recognition in step 323), freeze the parameters of the BERT model, input the embedding vector of the entity data output by the BERT model into the fully connected layer with a binary classification SoftMax structure, calculate the binary cross entropy, use the binary cross entropy to measure the gap between the prediction result of the binary classification model for numerical entity recognition on the input entity data and the true category, and train and update the model parameters according to the binary cross entropy loss to obtain a trained binary classification model for numerical entity recognition.

[0030] Preferably, step 4) specifically includes:

[0031] 41) Concatenate the body abstracts of each email in the correlated email cluster to form a concatenated body abstract;

[0032] 42) The urgency level of each correlated email cluster takes the average of the urgency levels of each email in the correlated email cluster;

[0033] 43) Summarize the questions that need to be replied to in each email in the associated email clusters, and record the corresponding EML email addresses to form a list of email addresses that need to be replied to;

[0034] 44) Invoke the general basic large model to summarize the spliced text abstract once again, and output a batch of abstract titles and batch abstract contents;

[0035] 35) After all associated email clusters are processed, return a list, where each element in the list is a dictionary containing a batch of abstract titles, batch abstract contents, and a list of email addresses that need to be replied to, with the highest urgency ranked at the front.

[0036] In addition, the present invention also provides a large model-based email batch processing system, which is characterized by including:

[0037] An email collection and preprocessing module, which is used to collect emails that need to be batch processed and perform preprocessing;

[0038] An email large model acquisition module, which is used to perform supervised fine-tuning and evaluation on the general basic large model by using the constructed email dataset to obtain an email large model;

[0039] An email association analysis module, which is used to perform association analysis on the emails that need to be batch processed by using the email large model and the trained numerical entity recognition binary classification model to obtain associated email clusters;

[0040] A batch abstract summarization module, which is used to perform batch abstract summarization on the emails that need to be batch processed by using the basic large model in units of associated email clusters.

[0041] Moreover, the present invention also provides a large model-based email batch processing device, which is characterized by including:

[0042] One or more processors;

[0043] A memory for storing one or more programs;

[0044] When the one or more programs are executed by the one or more processors, the one or more processors implement the large model-based email batch processing method as described above.

[0045] Finally, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and is characterized in that when the program is executed by a processor, the steps of the large model-based email batch processing method as described above are implemented.

[0046] Compared with the prior art, the large model-based email batch processing method, system, device, and storage medium of the present invention have one or more of the following beneficial technical effects:

[0047] 1. The present invention can achieve efficient and accurate extraction of key information: by constructing a fine-tuning dataset and a new evaluation method that fits the email scenario, the general basic large model is fine-tuned and evaluated, so that the obtained email large model can more accurately extract key information in the email, such as email type, urgency, body summary, key entity information, and questions that need to be answered.

[0048] 2. The present invention can realize efficient email association analysis: the trained numerical entity recognition binary classification model can effectively filter numerical entity information, and retrieve related emails based on non-numerical key entity information and embedded vectors, so that the constructed related email clusters are more accurate, and the accuracy of email association analysis is comprehensively improved.

[0049] 3. The present invention can improve the batch mail processing capability: it can realize batch mail summary generation, process related mail clusters as a whole, improve mail processing efficiency, help users quickly understand the whole picture of events, and sort related mail clusters according to the degree of urgency, giving priority to important mails.

[0050] 4. The present invention has strong adaptability: the general basic large model is optimized for the mail processing scenario. Compared with directly applying the general basic large model, it can better meet the specific needs of mail processing and has better performance in processing different types of mails. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the mail batch processing method based on the large model of the present invention.

[0052] Figure 2 It is a schematic diagram of the composition of the mail batch processing system based on the large model of the present invention. DETAILED DESCRIPTION

[0053] Before describing in detail any embodiment of the present invention, it should be understood that the present invention is not limited in its application to the construction and arrangement details of the components described below or illustrated in the following figures. The present invention can have other embodiments and can be practiced or carried out in various ways. In addition, it should be understood that the words and terms used here are for descriptive purposes and should not be considered restrictive. The use of "including" or "having" and its variations herein is intended to cover the items and their equivalents and additional items displayed below. Unless otherwise specified or limited, the terms "install", "connect", "support" and "couple" and their variations are widely used and cover direct installation and indirect installation, connection, support and connection. In addition, "connect" and "couple" are not limited to physical or mechanical connections or connections.

[0054] Furthermore, on the first aspect, in the disclosure of the present invention, the orientation or positional relationship indicated by terms such as "longitudinal", "transverse", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside" and "outside" are based on the orientation or positional relationship shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore the above terms cannot be understood as limitations on the present invention; on the second aspect, the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element may be one, while in another embodiment, the number of the element may be multiple, and the term "one" cannot be understood as a limitation on the quantity.

[0055] In today's information age, companies and individuals receive massive amounts of emails every day. Traditional email processing methods rely on manual reading and summarization, which is inefficient and prone to missing important information. The emergence of big model technology can extract key entity information from email messages more accurately and efficiently compared to traditional natural language processing technology. At the same time, there are often associations between emails, such as a series of emails around the same project or activity. Only a single email message will prevent these related information from being fully utilized. The present invention aims to solve these problems and provide an efficient and accurate big model-based email batch processing method that can improve the efficiency and quality of email processing to adapt to the rapid development and complex needs of the information age.

[0056] Before introducing the specific content of the present invention, some technical terms used in the present invention are briefly introduced to help those skilled in the art to better understand the present invention.

[0057] 1. Large model: a pre-trained model with powerful language understanding and generation capabilities.

[0058] 2. Fine-tuning: Based on the general basic large model, the model is adapted to specific application scenarios by training on specific task datasets.

[0059] 3. EML file: standard email file format, used to save email content, send and receive information, etc.

[0060] 4. Embedding vector: Convert text into numerical vectors to facilitate computer understanding and calculation of similarities between texts.

[0061] 5. BERT model: A pre-trained language model that performs well in natural language processing tasks.

[0062] 6. SoftMax structure: An activation function used for multi-classification tasks that can convert model output into probability distribution.

[0063] 7. Binary Cross Entropy: A loss function used to measure the difference between two probability distributions, commonly used in binary classification tasks.

[0064] 8. Embedding technology: A technology that converts data such as text and images from the original high-dimensional sparse representation into a low-dimensional dense vector representation, which can retain the key features of the data and facilitate model learning and processing.

[0065] The content of the present invention will be described in detail below. Among them, Figure 1 shows a flowchart of the method for batch processing of emails based on a large model of the present invention. As Figure 1 shown, the method for batch processing of emails based on a large model of the present invention includes the following steps:

[0066] First, collect the emails that need to be batch processed and perform preprocessing.

[0067] To batch process a large number of emails that need to be batch processed, it is first necessary to collect the emails that need to be batch processed and perform preprocessing.

[0068] In the present invention, through an automated script, a large number of emails that need to be batch processed are saved as EML files, and at the same time, the email content is parsed to obtain the email sending and receiving time, sender and recipient information, body text, signature information, email attachments, etc. Then, existing data cleaning techniques are used to unify the data format of the parsed email content and remove outliers to improve data quality. Among them, when processing email attachments, corresponding file information such as PDF and Word files can be parsed through various third-party libraries to obtain the attachment content.

[0069] Second, use the constructed email fine-tuning dataset to perform supervised fine-tuning and evaluation on the general basic large model to obtain an email large model.

[0070] In the present invention, a large model is used to obtain information such as the email type, urgency level, body text summary, key entity information, and questions that need to be replied to for a large number of emails that need to be batch processed. Among them, the "key entity information" refers to the entity content with important significance and specific identification in the email text, such as people, organizations, locations, products, etc.

[0071] Considering that the general basic large model is not trained for the email scenario and the actual use effect is poor, therefore, in the present invention, it is necessary to use the constructed email fine-tuning dataset to perform supervised fine-tuning and evaluation on the general basic large model to obtain a dedicated email large model, which specifically includes:

[0072] 1. Collect existing emails and perform preprocessing to obtain existing email data.

[0073] To perform supervised fine-tuning on the general basic large model, it is first necessary to collect existing emails and perform preprocessing to obtain existing email data

[0074] Here, in the same way, the collected existing emails can be saved as EML files through an automated script. At the same time, the email content is parsed to obtain information such as the sending and receiving time of the email, the sender and recipient information, the body text, the signature information, and the email attachments. And through data cleaning, the data format of the parsed email content is unified, and outliers are removed to improve data quality. Among them, when processing email attachments, various third-party libraries can be used to parse the corresponding file information, such as PDF, Word and other files, to obtain the attachment content.

[0075] 2. Construct a mail fine-tuning dataset based on the existing email data.

[0076] In the present invention, the data format in the mail fine-tuning dataset is based on the instruction-input-output (i.e., instruction-input-output) standard triple template. The instruction is used to set the prompt words, requiring the mail large model after supervised fine-tuning to output content such as the mail type, urgency level, body text summary, key entity information, etc.; the input includes information such as the sending and receiving time of the email, the sender and recipient information, the body text, and the attachment content; the output is the content that the mail large model after supervised fine-tuning is to output.

[0077] Before supervised fine-tuning, it is necessary to manually annotate the output of each piece of data in the mail fine-tuning dataset, and divide the annotated data into two parts: one part is used for supervised fine-tuning, and the other part is used for evaluation. The output content of the output is as follows:

[0078] Mail type: Classify the mail, such as file transfer type, event invitation type, message subscription type, etc.;

[0079] Urgency level: The numerical value of the urgency level of the mail to be processed, from low to high is 1, 2, 3;

[0080] Body text summary: A summary of the mail body text and attachment content, requiring no more than 100 tokens;

[0081] Key entity information: Identify the key entity information in the mail body text and attachment content, and store it in the form of key-value pairs;

[0082] Questions that need to be replied: Identify the questions in the mail body text and attachment content that need to be replied by the recipient, and store them as a list.

[0083] Thus, taking the relevant data of each email (including information such as email sending and receiving time, sender and recipient information, body text, and attachment content) as input, manually annotating the corresponding output (including email type, urgency level, body text summary, key entity information, and questions to be replied), and adding the corresponding instruction, one can obtain a piece of data in the email fine-tuning dataset. Among them, the output is strictly in JSON format. The following is a simple example of the output:

[0084]

[0085] 3. Use the email fine-tuning dataset to perform supervised fine-tuning on the general basic large model to obtain the supervised fine-tuned basic large model.

[0086] After having the email fine-tuning dataset, one can use the email fine-tuning dataset to perform supervised fine-tuning on the general basic large model to obtain the supervised fine-tuned basic large model.

[0087] In the supervised fine-tuning of the present invention, the tool LLaMA-Factory can be used to perform fine-tuning on the general basic large model to better meet the requirements of the email scenario. The specific operation process is as follows:

[0088] First, it is necessary to load the parameter weight file of the general basic large model. These parameter weight files are the digital manifestation of the knowledge and features learned by the general basic large model during the previous large-scale training process. They record the connection strength and weight values between each neuron in the general basic large model. These parameter weight files generally come from public pre-trained model repositories or are obtained through self-built large-scale computing clusters and are the parameter weight files corresponding to the general basic large model obtained after long-term training.

[0089] Next, combine the loaded general - purpose large - model with the email fine - tuning dataset to perform LoRA fine - tuning training on the general - purpose large - model. LoRA fine - tuning is an efficient fine - tuning technique that adaptively adjusts the general - purpose large - model by adding a small number of trainable low - rank matrices on the basis of the general - purpose large - model, without retraining the parameters of the entire general - purpose large - model. After the LoRA fine - tuning training is completed, the general - purpose large - model will optimize and adjust its own parameters according to the characteristics and task requirements of the email fine - tuning dataset, resulting in a fine - tuned general - purpose large - model that contains a set of updated parameter matrices. The parameter matrix is a matrix structure composed of the weight parameters of each layer in the fine - tuned general - purpose large - model. They play a key role in the forward and backward propagation processes of the fine - tuned general - purpose large - model, determining the processing method and output result of the fine - tuned general - purpose large - model for the input data. Each parameter matrix corresponds to a specific layer in the fine - tuned general - purpose large - model, and the element values in the matrix are the connection weights between neurons in that layer.

[0090] Finally, merge these fine - tuned parameter matrices. The merged parameter matrix will contain the original knowledge of the general - purpose large - model and the specific knowledge learned from the email fine - tuning dataset, forming a new SFT large - model, that is, the general - purpose large - model after supervised fine - tuning.

[0091] 4. Evaluate the supervised - fine - tuned general - purpose large - model based on the output of the supervised - fine - tuned general - purpose large - model and the labeled answers, and use the supervised - fine - tuned general - purpose large - model that passes the evaluation as the email large - model.

[0092] After the supervised fine - tuning is completed, it is necessary to evaluate the effect of the supervised - fine - tuned general - purpose large - model. In the present invention, compare the embedding vectors of the email type, urgency level, body abstract, key entity information, and questions to be answered output by the supervised - fine - tuned general - purpose large - model with the embedding vectors of the email type, urgency level, body abstract, key entity information, and questions to be answered in the standard answers, and use the comparison result as the evaluation result of the supervised - fine - tuned general - purpose large - model.

[0093] Specifically, obtain the embedding vectors of the output text (including the email type, urgency level, body abstract, key entity information, and questions to be answered and the standard answers) of the general - purpose large - model after supervised questions and the embedding vectors of the labeled answers through the embedding model. Let the embedding vector of the output text of the general - purpose large - model after supervised questions be O, and the embedding vector of the labeled answer be S. The calculation of each score is as follows:

[0094] Email type score T: If O["email type"] == S["email type"], then T = 1; otherwise, T = 0.

[0095] Emergency level score E: If O["Emergency level"] == S["Emergency level"], then E = 1; otherwise E = 0.

[0096] Body abstract score A: A is equal to the cosine similarity between O["Body abstract"] and S["Body abstract"].

[0097] Key entity information score K: Let the number of entity items in O["Key entity information"] be n O , and the number of correctly identified entity items in S["Key entity information"] be n C , calculate the similarity between each entity item in O and each entity item in S. If there is a similarity > 0.85, then consider this entity item as one of the true labels and it is correctly identified, K = n C / n O .

[0098] Score Q for questions to be replied: Let the number of questions to be replied in O["Questions to be replied"] be m O , and the number of correctly identified questions in S["Questions to be replied"] be m C , calculate the similarity between each question in O and each question in S. If there is a similarity > 0.8, then consider this question as one of the true labels and it is correctly identified, Q = m C / m O .

[0099] Comprehensive score F: The default weight of each item in the comprehensive score is the same, F = (T + E + A + K + Q) / 5.

[0100] The final evaluation score is the average of all evaluation data.

[0101] In specific evaluations, the evaluation score threshold (0.7 - 0.8) can be set according to different data environments. At the same time, the average comprehensive score of each category can be viewed. Through data ratio adjustment, increase the proportion of email types with low scores, and re - fine - tune the large model after supervised fine - tuning until a satisfactory large model after supervised fine - tuning is obtained. For example, if the score of the type "Questions to be replied" is very low, then some email fine - tuning data containing "Questions to be replied" can be collected again, and the large model after supervised fine - tuning can be fine - tuned again using the re - collected email fine - tuning data until a satisfactory large model after supervised fine - tuning is obtained.

[0102] III. Use the email large model and the trained numerical entity recognition binary classification model to perform correlation analysis on the emails to be batch - processed to obtain correlated email clusters.

[0103] When performing email association, one of the main indicators is the key entity information. However, numerical key entity information such as time, amount, quantity, etc. is not suitable as association information because even if two emails are not associated, their numerical key entity information may be the same. Non-numerical key entity information such as activity location, project name, professional terms, etc. can be used as an important indicator for whether emails are associated. Therefore, when the present invention helps to perform email association using key entity information by training a binary classification model for numerical entity recognition, it filters out numerical key entity information and only retains non-numerical key entity information to enhance the relevance of emails.

[0104] Among them, the trained binary classification model for numerical entity recognition is obtained through the following steps:

[0105] 1. Collect entity data (including numerical entity information such as time, amount, quantity, etc. and non-numerical entity information such as activity location, project name, professional terms, etc.), and divide the collected entity data into a training set, a validation set, and a test set according to a ratio of 6:2:2.

[0106] 2. Add a fully connected layer with a binary classification SoftMax structure on the basis of the BERT model to obtain a binary classification model for numerical entity recognition.

[0107] 3. Use the training set, validation set, and test set to train the binary classification model for numerical entity recognition and update the model parameters to obtain a trained binary classification model for numerical entity recognition.

[0108] Among them, when using the training set, validation set, and test set to train the binary classification model for numerical entity recognition and update the model parameters, freeze the parameters of the BERT model to facilitate better training and updating of the parameters of the fully connected layer with a binary classification SoftMax structure. At the same time, input the embedding vector of the entity data output by the BERT model into the fully connected layer with a binary classification SoftMax structure and then calculate the binary cross-entropy. Use the binary cross-entropy to measure the gap between the prediction result and the true category of the binary classification model for numerical entity recognition for the input entity data. Then, the model parameters can be trained and updated according to the binary cross-entropy loss to obtain a trained binary classification model for numerical entity recognition.

[0109] After having a trained binary classification model for numerical entity recognition, the email large model and the trained binary classification model for numerical entity recognition can be used to perform association analysis on the emails to be batch processed to obtain associated email clusters. The specific process is as follows:

[0110] 1. Use the email large model to process each email in the batch of preprocessed emails to be processed to obtain the key entity information of the emails.

[0111] 2. Filter the key entity information through the trained binary classification model for numerical entity recognition to filter out the numerical key entity information, so as to obtain non-numerical key entity information.

[0112] 3. Based on the non-numerical key entity information, perform string splicing and convert the spliced string into an embedding vector through the embedding technology.

[0113] 4. Based on the non-numerical key entity information and the embedding vector, perform correlation analysis on the emails to be batch processed to obtain correlation email clusters.

[0114] Specifically, first, preliminarily classify the emails to be batch processed through the non-numerical key entity information, and preliminarily classify all emails with the same or similar non-numerical key entity information in the emails to be batch processed into one email cluster, thus forming multiple email clusters. Then, compare the embedding vectors of the emails within the same email cluster, and classify all emails with an embedding vector similarity greater than 0.6 into the same correlation email cluster. Thus, the emails to be batch processed are classified into multiple correlation email clusters, and all emails within the same correlation email cluster are emails with the same or similar non-numerical key entity information and an embedding vector similarity greater than 0.6.

[0115] IV. Use the basic large model to batch summarize the emails to be batch processed in units of correlation email clusters.

[0116] The large model after supervised fine-tuning has obtained the body summary, urgency level, questions to be replied, etc. of each email. Here, the body summaries of each email in the correlation email cluster can be spliced to form a spliced body summary. At the same time, take the average value of the urgency levels of each email in the correlation email cluster as the urgency level of the correlation email cluster. Moreover, summarize the questions to be replied to in each email in the correlation email cluster and record the corresponding EML email addresses to form a list of email addresses to be replied. Then, call the general basic large model to summarize the spliced body summary again to output the batch summary title and batch summary content. In this way, after all correlation email clusters are processed, a list is returned, and each element in the list is a dictionary containing the batch summary title, batch summary content, and a list of email addresses to be replied, and the one with the highest urgency level is ranked at the top.

[0117] Figure 2 The schematic diagram of the composition of the email batch processing system based on the large model of the present invention is shown.

[0118] As Figure 2 described, the email batch processing system based on the large model includes:

[0119] 1. Mail collection and preprocessing module.

[0120] The mail collection and preprocessing module is used to collect mails that need to be batch - processed and perform preprocessing.

[0121] 2. Mail large - model acquisition module.

[0122] The mail large - model acquisition module is used to supervise, fine - tune, and evaluate a general - purpose basic large - model using a constructed mail data set to obtain a mail large - model.

[0123] 3. Mail correlation analysis module.

[0124] The mail correlation analysis module is used to perform correlation analysis on mails that need to be batch - processed using the mail large - model and a trained binary classification model for numerical entity recognition to obtain correlated mail clusters.

[0125] 4. Batch summary module.

[0126] The batch summary module is used to perform batch summary on mails that need to be batch - processed using a basic large - model with correlated mail clusters as units.

[0127] In addition, the present invention also provides a mail batch - processing device based on a large - model, which includes: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above - mentioned mail batch - processing method based on a large - model.

[0128] Finally, the present invention provides a computer - readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the above - mentioned mail batch - processing method based on a large - model are implemented.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting the protection scope of the present invention. Those skilled in the art, based on the idea of the present invention, can modify or equivalently replace the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for batch processing of emails based on a large model, characterized in that, It includes the following steps: 1) Collect emails that need to be processed in batches and perform preprocessing; 2) Use the constructed email fine-tuning dataset to perform supervised fine-tuning and evaluation on the general basic large model to obtain an email large model; 3) Use the email large model and the trained numerical entity recognition binary classification model to perform correlation analysis on the emails that need to be processed in batches to obtain correlated email clusters; 4) Use the general basic large model to perform batch summary and summarization on the emails that need to be processed in batches in units of correlated email clusters.

2. The method for batch processing of emails based on a large model according to claim 1, wherein The specific steps of step 1) include: 11) Save the emails that need to be processed in batches as EML files through an automated script, and at the same time parse the email content to obtain the email sending and receiving time, sender and recipient information, body text, signature information, and email attachments; 12) Uniformly parse the data format of the obtained email content through data cleaning and remove outliers.

3. The method for batch processing of emails based on a large model according to claim 1, wherein, The specific steps of step 2) include: 21) Collect existing emails and perform preprocessing to obtain existing email data; 22) Based on the existing email data, construct an email fine-tuning dataset. The data format in the email fine-tuning dataset is based on the instruction-input-output standard triple template. Among them, the instruction is used to set the prompt words, the input includes the email sending and receiving time, sender and recipient information, body text, and attachment content, and the output includes the email type, urgency level, body text summary, key entity information, and questions that need to be replied; 23) Use the email fine-tuning dataset to perform supervised fine-tuning on the general basic large model to obtain a supervised fine-tuned basic large model; 24) Based on the output of the supervised fine-tuned basic large model and the labeled answers, evaluate the supervised fine-tuned basic large model, and use the supervised fine-tuned basic large model that passes the evaluation as the email large model.

4. The method for batch processing of emails based on a large model according to claim 1, wherein, The specific steps of step 3) include: 31) Use the email large model to process each email in the preprocessed batch emails to obtain the key entity information of the emails; 32) Filter the key entity information through the trained numerical entity recognition binary classification model, and filter out the numerical key entity information to obtain non-numerical key entity information; 33) Perform string concatenation based on the non-numerical key entity information and convert the concatenated string into an embedding vector through an embedding technique; 34) Based on the non-numerical key entity information and the embedding vector, perform correlation analysis on the emails that need to be processed in batches to obtain correlated email clusters.

5. The method for batch processing of emails based on a large model according to claim 4, wherein, The trained numerical entity recognition binary classification model in step 32) is obtained through the following method: 321) Collect entity data including numerical entity information and non-numerical entity information, and divide the entity data into a training set, a validation set, and a test set; 322) Add a fully connected layer with a binary classification SoftMax structure on the basis of the BERT model to obtain a numerical entity recognition binary classification model; 323) Use the training set, validation set, and test set to train the numerical entity recognition binary classification model and update the model parameters to obtain a trained numerical entity recognition binary classification model.

6. The method for batch processing of emails based on a large model according to claim 5, wherein In step 323), when training the binary classification model for numerical entity recognition and updating the model parameters using the training set, validation set, and test set, the parameters of the BERT model are frozen. The embedding vectors of the entity class data output by the BERT model are input into the fully connected layer with a binary classification SoftMax structure, and then the binary cross-entropy is calculated. The binary cross-entropy is used to measure the gap between the prediction result of the numerical entity recognition binary classification model for the input entity class data and the true class, and the model parameters are trained and updated according to the binary cross-entropy loss to obtain a trained numerical entity recognition binary classification model.

7. The method for batch processing of emails based on a large model according to claim 1, wherein Step 4) specifically includes: 41) Concatenate the body abstracts of each email in the associated email cluster to form a concatenated body abstract; 42) The urgency level of each associated email cluster is the average of the urgency levels of each email in the associated email cluster; 43) Summarize the questions that need to be replied to in each email in the associated email cluster, and record the corresponding EML email addresses to form a list of email addresses that need to be replied to; 44) Call the general basic large model to summarize the concatenated body abstract again to output a batch of abstract titles and batch of abstract contents; 35) After all associated email clusters are processed, return a list. Each element in the list is a dictionary containing the batch of abstract titles, batch of abstract contents, and a list of email addresses that need to be replied to, with the highest urgency level ranked at the front.

8. A batch email processing system based on a large model, characterized in that, Including: An email collection and preprocessing module, which is used to collect emails that need to be processed in batches and perform preprocessing; An email large model acquisition module, which is used to perform supervised fine-tuning and evaluation on the general basic large model using the constructed email dataset to obtain an email large model; An email association analysis module, which is used to perform association analysis on the emails that need to be processed in batches using the email large model and the trained binary classification model for numerical entity recognition to obtain associated email clusters; A batch abstract summarization module, which is used to perform batch abstract summarization on the emails that need to be processed in batches using the general basic large model with the associated email cluster as the unit.

9. An email batch processing device based on a large model, characterized in that, Including: One or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the large model-based email batch processing method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the large model-based email batch processing method according to any one of claims 1-7.

Images