Logistics text information automatic classification method and system based on self-supervised learning

Abstract

The invention relates to the technical field of logistics informatization, in particular to a logistics text information automatic classification method and system based on self-supervised learning, and the method comprises the steps: obtaining a text data set in the logistics field; preprocessing the text data set to obtain preprocessed text data; based on the preprocessed text data, constructing a multi-task self-supervised learning model comprising a mask word prediction task, a text fragment sorting task and a log information classification task; training the multi-task self-supervised learning model to obtain a pre-trained model; extracting text features from the preprocessed text data based on a pre-training model; a classification model training step: training a logistics text classification model based on the text features and predefined logistics text categories; receiving a logistics text to be classified; based on the logistics text classification model, classifying the logistics texts to be classified to obtain a classification result; and the classification result is output, so that the dependence on manual annotation data is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the field of logistics information technology, and more specifically, to a logistics text information automatic classification method and system based on self-supervised learning. Background Art

[0002] With the rapid development of e-commerce and logistics industry, the amount of logistics text information has exploded. These text information include logistics orders, delivery records, customer feedback and other forms, which contain rich business information and customer needs. However, how to effectively and automatically classify and analyze these massive amounts of logistics text has become a major challenge for logistics companies.

[0003] Traditional logistics text classification methods mainly rely on manually formulated rules or machine learning algorithms based on supervised learning. These methods have many limitations in practical applications. First, the method of manually formulating rules requires a lot of manpower and is difficult to deal with complex and changeable logistics texts. Secondly, although the method based on supervised learning has improved the degree of automation of classification to a certain extent, it still requires a large amount of labeled data, which is often difficult to obtain or too costly in practical applications.

[0004] In addition, existing logistics text classification methods often fail to fully utilize the specific knowledge and language characteristics of the logistics field. For example, logistics texts often contain professional terms, abbreviations, and specific expressions, which make general text classification methods perform poorly in the logistics field. At the same time, the structure and content of logistics texts also have their own particularities, such as the format of order information, the description of logistics status, etc. These characteristics are often not fully considered in existing methods.

[0005] Another key issue is that existing logistics text classification methods are often difficult to adapt to the rapid changes in business and the emergence of new categories. The business model and service types of the logistics industry are constantly evolving, and new logistics concepts and terms are constantly emerging. This requires classification methods to have good adaptability and scalability to quickly respond to these changes, but existing methods have obvious shortcomings in this regard. Summary of the invention

[0006] The present invention aims to solve the above technical problems and proposes a method and system for automatic classification of logistics text information based on self-supervised learning. This method innovatively combines self-supervised learning technology with logistics domain knowledge, effectively overcoming the reliance on a large amount of labeled data in the prior art, while making full use of the characteristics of logistics text and domain knowledge.

[0007] The present invention provides a method for automatic classification of logistics text information based on self-supervised learning, comprising:

[0008] Data acquisition steps include:

[0009] Get text datasets in the logistics field;

[0010] Preprocessing the text data set to obtain preprocessed text data;

[0011] Self-supervised learning steps include:

[0012] Based on the preprocessed text data, a multi-task self-supervised learning model including a mask word prediction task, a text segment sorting task, and a log information classification task is constructed;

[0013] Training the multi-task self-supervised learning model to obtain a pre-trained model;

[0014] The feature extraction steps include:

[0015] Based on the pre-trained model, extracting text features from the pre-processed text data;

[0016] Classification model training steps include:

[0017] Based on the text features and predefined logistics text categories, training a logistics text classification model;

[0018] Classification application steps include:

[0019] Receive logistics text to be classified;

[0020] Based on the logistics text classification model, classify the logistics text to be classified to obtain a classification result;

[0021] The classification result is output.

[0022] Preferably, the preprocessing in the data acquisition step specifically includes:

[0023] Performing data cleaning on the text data set and deleting irrelevant data;

[0024] Perform word segmentation on the cleaned data;

[0025] The text dataset is divided into a training set and a test set, where the initial ratio of the training set to the test set is 8:2.

[0026] Preferably, the mask word prediction task in the self-supervised learning step specifically includes:

[0027] In the preprocessed text data, a certain proportion of words are randomly selected for masking;

[0028] Based on context information, predict the masked words;

[0029] Calculate the cross entropy loss between the predicted results and the true values.

[0030] Preferably, the text segment sorting task in the self-supervised learning step specifically includes:

[0031] Randomly selecting two text segments from the preprocessed text data;

[0032] swapping the positions of the two text segments;

[0033] Determine whether the order after the swap is consistent with the original order;

[0034] The sorting loss is calculated based on the judgment results.

[0035] Preferably, the log information classification task in the self-supervised learning step specifically includes:

[0036] Classifying the preprocessed text data based on predefined logistics log categories;

[0037] Calculate the cross entropy loss between the classification result and the true category.

[0038] Preferably, the feature extraction step specifically comprises:

[0039] Inputting the preprocessed text data into a word embedding layer to obtain a word vector representation;

[0040] Input the word vector representation into a bidirectional gated recurrent unit network to capture contextual information;

[0041] A multi-scale attention mechanism is applied to the output of the bidirectional gated recurrent unit network to obtain text features.

[0042] Preferably, the classification model training step specifically includes:

[0043] Based on the text features, a fully connected neural network is constructed as a classifier;

[0044] Optimizing the classifier using a cross entropy loss function;

[0045] An iterative training strategy is adopted, and the results of each round of classification are used as the input of the next round of training to continuously optimize the classification model.

[0046] Preferably, the iterative training strategy specifically includes:

[0047] After completing 4 rounds of training, the ratio of training set to test set was adjusted to 6:4;

[0048] Continue model training based on the adjusted dataset ratio;

[0049] Record the classification accuracy of each round of training, and stop training when the accuracy no longer improves significantly.

[0050] Preferably, the model evaluation step is also included:

[0051] Based on the test set data, the performance of the logistics text classification model is evaluated;

[0052] Calculate evaluation indicators such as classification accuracy, precision, recall, and F1 score;

[0053] Based on the evaluation results, adjust the model parameters or retrain the model.

[0054] The automatic classification system of logistics text information based on self-supervised learning includes:

[0055] A data acquisition module is used to acquire a text data set in the field of logistics, and preprocess the text data set to obtain preprocessed text data;

[0056] A self-supervised learning module, for constructing a multi-task self-supervised learning model including a mask word prediction task, a text segment sorting task, and a log information classification task based on the preprocessed text data, and training the multi-task self-supervised learning model to obtain a pre-trained model;

[0057] A feature extraction module, used for extracting text features from the preprocessed text data based on the pretrained model;

[0058] A classification model training module, used for training a logistics text classification model based on the text features and predefined logistics text categories;

[0059] The classification application module is used to receive the logistics text to be classified, classify the logistics text to be classified based on the logistics text classification model, obtain the classification result, and output the classification result.

[0060] The present invention effectively solves many problems faced by existing technologies by innovatively combining advanced technologies such as multi-scale time-frequency analysis, topological feature extraction, nonlinear dynamic modeling, multimodal feature fusion, quantum computing acceleration, and environmental factor compensation.

[0061] Specifically, the present invention has the following significant beneficial effects:

[0062] The method of the present invention has significant technical effects and advantages. First, by introducing a multi-task self-supervised learning framework, the method can make full use of a large amount of unlabeled logistics text data, effectively reducing the reliance on manually labeled data and greatly reducing the cost and difficulty of data preparation. This not only improves the practicality of the method, but also enables the model to learn from a wider range of data, enhancing its generalization ability.

[0063] Secondly, the method of the present invention enables the model to comprehensively learn the semantics, structure and domain knowledge of logistics text through carefully designed tasks such as mask word prediction, text segment sorting and log information classification. This multi-angle learning method enables the model to have a deeper and more comprehensive understanding of logistics text, so that it can classify more accurately.

[0064] In addition, the method of the present invention has good adaptability and scalability. By adopting iterative training strategies and dynamic data set adjustment, the method can continuously absorb new data and knowledge and adapt to changes in logistics business and emerging categories. This adaptive ability makes the method have continuous effectiveness in practical applications.

[0065] Finally, the system design proposed in this invention adopts a modular structure, and each functional module works together, which not only ensures the clarity of the processing flow, but also provides good scalability. This design enables the system to flexibly respond to different logistics text classification requirements and is easy to maintain and upgrade.

[0066] In summary, the automatic classification method and system of logistics text information based on self-supervised learning proposed in the present invention not only solves the problems existing in the prior art, but also has made significant progress in classification accuracy, efficiency, adaptability and scalability, etc., providing logistics companies with an efficient and intelligent text classification solution, which is expected to play an important role in improving logistics management efficiency, optimizing customer service, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 The present invention is a flow chart of the method.

[0068] Figure 2 It is a logic block diagram of the data acquisition module of the present invention.

[0069] Figure 3 It is a logical block diagram of the self-supervised learning module of the present invention.

[0070] Figure 4 It is a logic block diagram of the feature extraction module of the present invention.

[0071] Figure 5 It is a logic block diagram of the classification model training module of the present invention.

[0072] Figure 6 It is a logic block diagram of the classification application module of the present invention. DETAILED DESCRIPTION

[0073] Please refer to Figure 1-5The present invention provides a method and system for automatic classification of logistics text information based on self-supervised learning. The method makes full use of a large amount of unlabeled text data in the logistics field, and effectively extracts the semantic features of logistics text through an innovative multi-task self-supervised learning framework, thereby achieving efficient and accurate automatic classification.

[0074] Specifically, the method of the present invention comprises the following steps:

[0075] First, in the data acquisition step, this method acquires text datasets in the field of logistics. These data can come from multiple sources such as the business system of the logistics company, customer feedback, logistics documents, etc. After acquiring the original data, this method preprocesses the text dataset to obtain preprocessed text data. The preprocessing process may include operations such as noise removal, format unification, and word segmentation, laying the foundation for subsequent processing.

[0076] Secondly, in the self-supervised learning step, this method builds a multi-task self-supervised learning model based on the preprocessed text data, including mask word prediction task, text segment sorting task and log information classification task. The design of these three tasks fully considers the characteristics of logistics text and can capture the semantic information of the text from different angles. Subsequently, this method trains the multi-task self-supervised learning model to obtain a pre-trained model. This pre-trained model has learned a wealth of knowledge in the logistics field, laying a solid foundation for subsequent feature extraction and classification tasks.

[0077] Next, in the feature extraction step, this method extracts text features from the preprocessed text data based on the pretrained model. These features contain the semantic information of the text and can effectively represent the content and intent of the logistics text.

[0078] Then, in the classification model training step, the method trains a logistics text classification model based on the extracted text features and predefined logistics text categories. This classification model learns how to map text features to specific logistics categories, thereby achieving the goal of automatic classification.

[0079] Finally, in the classification application step, this method receives the logistics text to be classified, classifies it using the trained logistics text classification model, obtains the classification result, and outputs the result.

[0080] Preferably, in one embodiment of the present invention, the preprocessing in the data acquisition step specifically includes the following operations: First, the text data set is cleaned and irrelevant data is deleted. This may include removing special characters, correcting obvious spelling errors, etc. Secondly, the cleaned data is segmented. Considering that the logistics text may contain professional terms, a field-specific segmentation dictionary can be used to improve the accuracy of segmentation. Finally, the text data set is divided into a training set and a test set, wherein the initial ratio of the training set to the test set is 8:2. This division ratio can ensure sufficient training data while leaving enough test data to evaluate the model performance.

[0081] In another embodiment of the present invention, the masked word prediction task in the self-supervised learning step specifically includes the following process: First, randomly select a certain proportion of words in the preprocessed text data for masking. The mask ratio is an important hyperparameter, which is usually set between 15% and 20%. Too high a mask ratio may make the task too difficult, while too low a mask ratio may cause insufficient model learning. Secondly, based on contextual information, predict the masked words. This process usually uses a bidirectional transformer structure to capture contextual information. Finally, calculate the cross entropy loss between the predicted result and the true value. The cross entropy loss function can be expressed by the following LaTeX formula:

[0082]

[0083] Where L is the loss value, N is the vocabulary size, and y i is the true label (one-hot encoding), is the probability distribution predicted by the model.

[0084] This masked word prediction task enables the model to learn the semantic relationship between words, which is crucial for understanding the context of logistics text. For example, when predicting the masked word in "The goods have arrived [MASK]", the model needs to understand the logistics process and may predict related words such as "warehouse" and "destination".

[0085] Through these steps, the method of the present invention can effectively utilize a large amount of unlabeled logistics text data, learn rich domain knowledge and language representation, and thus provide strong support for subsequent classification tasks. This method based on self-supervised learning not only reduces the dependence on labeled data, but also can capture subtle semantic differences in logistics texts, improving the accuracy and robustness of classification.

[0086] In a preferred embodiment of the present invention, the text segment sorting task in the self-supervised learning step further includes the following specific process: First, two text segments are randomly selected from the preprocessed text data. The length of these segments can be fixed or dynamically changed. Preferably, the segment length can be set to 10 to 20 words, which can ensure sufficient context information without making the task too simple.

[0087] Next, the method swaps the positions of the two text segments. This operation simulates text order errors that may occur in real scenarios, such as the confusion of the order of information entry in logistics documents. Then, the method determines whether the swapped order is consistent with the original order. This judgment process can be implemented through a binary classification task, in which the model needs to predict whether the two segments are in the correct relative position.

[0088] Finally, the ranking loss is calculated based on the judgment results. This loss function can use binary cross entropy, and its LaTeX expression is as follows:

[0089] L=-[ylog(p)+(1-y)log(1-p)],

[0090] Among them, L is the loss value, y is the true label (1 means the order is correct, 0 means the order is wrong), and p is the probability predicted by the model.

[0091] This text segment sorting task can help the model understand the overall structure and logical order of logistics text. For example, when processing logistics delivery information, the model can learn that the shipping address usually appears before the delivery address, thereby better understanding the semantic structure of the text.

[0092] In another embodiment of the present invention, the log information classification task in the self-supervised learning step specifically includes the following process: First, the pre-processed text data is classified based on predefined logistics log categories. These predefined categories may include key nodes in the logistics process such as "order creation", "warehouse outbound", "in transit", and "delivered".

[0093] Next, this method calculates the cross entropy loss between the classification result and the true category. The cross entropy loss function for multi-category classification problems can be expressed as the following LaTeX formula:

[0094]

[0095] Among them, L is the loss value, C is the number of categories, and y i is the true label (one-hot encoding), is the probability distribution predicted by the model.

[0096] The introduction of this log information classification task enables the model to learn professional knowledge and business processes in the logistics field. For example, the model can learn business rules such as "signing for" usually occurs after "delivery" and "return" may occur after "signing for". The injection of this domain knowledge greatly improves the model's ability to understand logistics texts.

[0097] Preferably, the feature extraction step of the present invention further includes the following specific operations: First, the preprocessed text data is input into the word embedding layer to obtain a word vector representation. In the field of logistics, pre-trained word vectors such as Word2Vec or GloVe can be used and fine-tuned on this basis to better capture domain-specific semantics.

[0098] Secondly, the word vector representation is input into the Bidirectional Gated Recurrent Unit (Bi-GRU) network to capture contextual information. The use of Bi-GRU allows the model to consider the contextual information of the text at the same time, which is particularly important for understanding the temporal relationship in logistics text. For example, when processing logistics tracking information, the state changes of the context can help the model understand the current state more accurately.

[0099] Finally, a multi-scale attention mechanism is applied to the output of Bi-GRU to obtain text features. The multi-scale attention mechanism allows the model to pay attention to contextual information of different lengths, and its mathematical expression can be simplified as:

[0100] α i =softmax(W a tanh(W h h i +b a )),

[0101] c=∑ i α i h i ,

[0102] Among them, α i is the attention weight, h i is the hidden state of Bi-GRU, W a , W h and b a is a learnable parameter and c is the final context vector.

[0103] This multi-level feature extraction process can effectively capture the semantic information of logistics texts, from word level to sentence level and then to document level, to fully understand the text content. For example, when processing logistics complaint information, the model can focus on specific problem description words (such as "breakage" and "delay") and the overall tone and intonation, so as to more accurately judge the type and urgency of the complaint.

[0104] Through these carefully designed steps, the method of the present invention can fully and deeply understand the semantic content of logistics texts, laying a solid foundation for subsequent classification tasks. This multi-task, multi-level learning method not only improves the generalization ability of the model, but also enhances its understanding of specific knowledge in the logistics field, so that text classification can be performed more accurately. In a preferred embodiment of the present invention, the classification model training step further includes the following specific operations: First, based on the extracted text features, the method constructs a fully connected neural network as a classifier. This fully connected neural network usually contains multiple hidden layers, and each layer uses a nonlinear activation function, such as ReLU (Rectified Linear Unit). Preferably, the number of hidden layers can be set to 2 to 3 layers, and the number of neurons in each layer can be reduced layer by layer, for example from 512 to 256 and then to 128. Such a structure can gradually extract more advanced features.

[0105] Next, this method uses the cross entropy loss function to optimize the classifier. For multi-class classification problems, the mathematical expression of the cross entropy loss function is as follows:

[0106]

[0107] Among them, L represents the loss value, C is the number of categories, and y i is the true label (one-hot encoding), is the probability distribution predicted by the model.

[0108] In particular, this method adopts an iterative training strategy, taking the classification results of each round as the input of the next round of training to continuously optimize the classification model. This strategy is similar to the self-training method in semi-supervised learning, which can make full use of unlabeled data and gradually improve model performance.

[0109] In another embodiment of the present invention, the iterative training strategy specifically includes the following steps: First, after completing 4 rounds of training, the method adjusts the ratio of the training set to the test set to 6:4. The purpose of this dynamic adjustment strategy is to increase the proportion of test data while ensuring sufficient training data to more comprehensively evaluate the model performance.

[0110] Subsequently, the method continues to train the model based on the adjusted dataset ratio. In each round of training, the model predicts the unlabeled data and adds the prediction results with high confidence to the training data for the next round. The confidence threshold can be set to 0.95, that is, only when the model's classification confidence for a sample exceeds 95%, it will be included in the next round of training.

[0111] Finally, this method records the classification accuracy of each round of training, and stops training when the accuracy no longer improves significantly. "Significant improvement" can be defined as the improvement in accuracy is less than 0.1% in three consecutive rounds of training. This early stopping strategy can effectively prevent model overfitting and save computing resources.

[0112] It is worth noting that the method of the present invention also includes a model evaluation step. In this step, the method first evaluates the performance of the logistics text classification model X based on the test set data. The evaluation indicators include but are not limited to classification accuracy, precision, recall and F1 score. The calculation formulas of these indicators are as follows:

[0113] Accuracy:

[0114] Accuracy:

[0115] Recall:

[0116] F1 score:

[0117] Among them, TP represents true positive examples, TN represents true negative examples, FP represents false positive examples, and FN represents false negative examples. Then, this method adjusts the model parameters or retrains the model based on the evaluation results. For example, if it is found that the recall rate of some categories is particularly low, you can consider increasing the proportion of these categories in the training set, or adjusting the category weights to balance the importance of different categories.

[0118] Next, this method adjusts the model parameters or retrains the model based on the evaluation results. For example, if the recall rate of some categories is found to be particularly low, you can consider increasing the proportion of these categories in the training set or adjusting the category weights to balance the importance of different categories.

[0119] Finally, the present invention also provides a logistics text information automatic classification system based on self-supervised learning. The system includes the following modules:

[0120] The data acquisition module 1 is used to acquire a text data set in the field of logistics, and preprocess the text data set to obtain preprocessed text data.

[0121] The self-supervised learning module 2 is used to construct a multi-task self-supervised learning model including a mask word prediction task, a text segment sorting task and a log information classification task based on the preprocessed text data, and train the multi-task self-supervised learning model to obtain a pre-trained model.

[0122] The feature extraction module 3 is used to extract text features from the preprocessed text data based on the pre-trained model.

[0123] The classification model training module 4 is used to train a logistics text classification model based on text features and predefined logistics text categories.

[0124] The classification application module 5 is used to receive the logistics text to be classified, classify the logistics text to be classified based on the logistics text classification model, obtain the classification result, and output the classification result.

[0125] Preferably, the data acquisition module 1 further comprises a data cleaning submodule 11 and a data partitioning submodule 12. The data cleaning submodule 11 is responsible for removing noise and irrelevant information in the text, while the data partitioning submodule 12 is responsible for partitioning the data set into a training set and a test set.

[0126] The self-supervised learning module 2 may further include a mask prediction submodule 21, a sorting task submodule 22 and a classification task submodule 23, which correspond to three pre-training tasks respectively.

[0127] The feature extraction module 3 may include a word embedding submodule 31, a sequence modeling submodule 32 and an attention mechanism submodule 33, which are respectively responsible for word vector representation, context information capture and important information extraction.

[0128] The classification model training module 4 may include a neural network construction submodule 41 and an optimization strategy submodule 42, wherein the former is responsible for constructing the network structure of the classifier and the latter is responsible for implementing the iterative training strategy.

[0129] The classification application module 5 may include a text preprocessing submodule 51 , a feature extraction submodule 52 and a classification prediction submodule 53 , which are respectively responsible for preprocessing the input text, extracting features and performing final classification prediction.

[0130] Through this modular design, the system of the present invention can flexibly respond to different logistics text classification requirements while ensuring the clarity and scalability of the processing flow. For example, if you need to adapt to a new logistics business scenario, you only need to make adjustments in the corresponding module without changing the structure of the entire system. This design concept enables the system to not only efficiently and accurately complete the automatic classification task of logistics text, but also has good adaptability and maintainability.

[0131] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, replacement, and improvement made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. Automatic classification method of logistics text information based on self-supervised learning, characterized by ,include: Data acquisition steps include: Get text datasets in the logistics field; Preprocessing the text data set to obtain preprocessed text data; Self-supervised learning steps include: Based on the preprocessed text data, a multi-task self-supervised learning model including a mask word prediction task, a text segment sorting task, and a log information classification task is constructed; Training the multi-task self-supervised learning model to obtain a pre-trained model; The feature extraction steps include: Based on the pre-trained model, extracting text features from the pre-processed text data; Classification model training steps include: Based on the text features and predefined logistics text categories, training a logistics text classification model; Classification application steps include: Receive logistics text to be classified; Based on the logistics text classification model, classify the logistics text to be classified to obtain a classification result; The classification result is output.

2. The method according to claim 1, characterized in that , the preprocessing in the data acquisition step specifically includes: Performing data cleaning on the text data set and deleting irrelevant data; Perform word segmentation on the cleaned data; The text dataset is divided into a training set and a test set, where the initial ratio of the training set to the test set is 8:

2.

3. The method according to claim 1, characterized in that ,The mask word prediction task in the self-supervised learning step specifically includes: In the preprocessed text data, a certain proportion of words are randomly selected for masking; Based on context information, predict the masked words; Calculate the cross entropy loss between the predicted results and the true values.

4. The method according to claim 1, characterized in that ,The text segment sorting task in the self-supervised learning step specifically includes: Randomly selecting two text segments from the preprocessed text data; swapping the positions of the two text segments; Determine whether the order after the swap is consistent with the original order; The sorting loss is calculated based on the judgment results.

5. The method according to claim 1, characterized in that ,The log information classification task in the self-supervised learning step specifically includes: Classifying the preprocessed text data based on predefined logistics log categories; Calculate the cross entropy loss between the classification result and the true category.

6. The method according to claim 1, characterized in that , the feature extraction step specifically includes: Inputting the preprocessed text data into a word embedding layer to obtain a word vector representation; Input the word vector representation into a bidirectional gated recurrent unit network to capture contextual information; A multi-scale attention mechanism is applied to the output of the bidirectional gated recurrent unit network to obtain text features.

7. The method according to claim 1, characterized in that , the classification model training step specifically includes: Based on the text features, a fully connected neural network is constructed as a classifier; Optimizing the classifier using a cross entropy loss function; An iterative training strategy is adopted, and the results of each round of classification are used as the input of the next round of training to continuously optimize the classification model.

8. The method according to claim 7, characterized in that , the iterative training strategy specifically includes: After completing 4 rounds of training, the ratio of training set to test set was adjusted to 6:4; Continue model training based on the adjusted dataset ratio; Record the classification accuracy of each round of training, and stop training when the accuracy no longer improves significantly.

9. The method according to claim 1, characterized in that , also includes the model evaluation step: Based on the test set data, the performance of the logistics text classification model is evaluated; Calculate evaluation indicators such as classification accuracy, precision, recall, and F1 score; Based on the evaluation results, adjust the model parameters or retrain the model.

10. Automatic classification system of logistics text information based on self-supervised learning, characterized by ,include: A data acquisition module is used to acquire a text data set in the field of logistics, and preprocess the text data set to obtain preprocessed text data; A self-supervised learning module, for constructing a multi-task self-supervised learning model including a mask word prediction task, a text segment sorting task, and a log information classification task based on the preprocessed text data, and training the multi-task self-supervised learning model to obtain a pre-trained model; A feature extraction module, used for extracting text features from the preprocessed text data based on the pretrained model; A classification model training module, used for training a logistics text classification model based on the text features and predefined logistics text categories; The classification application module is used to receive the logistics text to be classified, classify the logistics text to be classified based on the logistics text classification model, obtain the classification result, and output the classification result.

Images