Detection method and system for noise text information

Abstract

The invention relates to the technical field of public opinion monitoring, and discloses a noise text information-oriented detection method and system, and the method comprises the steps: obtaining a target data set; preprocessing the target data set to obtain preprocessed text features; a pre-training model of the Chinese text classification model is formed, and the pre-training model of the Chinese text classification model is used for the preprocessed text features to obtain representation vectors of the text; constructing two pre-training models which are different in structure and are based on Chinese text classification as a base classifier A and a base classifier B of the double-tower model; and adjusting the base classifier A and the base classifier B by using the same data set. According to the method, mutual verification can be carried out on the double-tower model, the output result of the double-tower model is comprehensively considered, and the final result is output, so that the overall specific content detection accuracy is effectively improved.

Description

Technical Field

[0001] The present application relates to the technical field of public opinion monitoring, for example, to a detection method and system for noise text information. Background Art

[0002] With the rapid development of the Internet and the explosive growth of the number of netizens, massive text information is widely disseminated and shared on various Internet platforms. While it brings convenience to people's lives and work, it also affects users' lives and work due to the existence of a large amount of specific content, such as false information and malicious advertisements. In addition, Internet text information also has high noise problems such as spelling errors, emoticons, and uneven text length. How to detect and filter specific content in massive Internet high-noise text information is a problem that must be considered.

[0003] At present, there are two main methods for detecting specific content of Internet text information: one is the traditional detection method, which is to match and filter massive text information through keywords, rule matching templates, etc. The other is the text detection method based on deep learning, which is to use deep neural network models, such as convolutional neural networks (CNN) and recurrent neural networks (RNN), to extract and represent text features and make classification judgments. The text detection method based on deep learning has achieved good results in the field of specific content detection. Its advantage is that it can automatically learn the features and patterns in the text and has strong adaptability to Internet text information with high noise, diversity and dynamics. For example, Purva Tendulkar et al. introduced a method for detecting online hate speech using a context-aware model. This method effectively captures the context information in the text by combining word embedding and sentence-level attention mechanism, and applies it to the task of detecting inappropriate online speech.

[0004] The specific content detection methods for Internet text information are mainly divided into two types: traditional template keyword matching methods and text classification methods based on deep learning. At present, traditional specific content detection methods mainly use keyword matching and rule template matching. Such methods have great limitations in specific scenarios. For example, in Internet text information with high noise, such as homophones and typos, the accuracy of keyword matching is often very low. The text classification method based on deep learning mainly uses deep neural network models to extract and represent text features and perform classification judgments. Such methods have problems such as low accuracy and poor robustness when facing specific content detection tasks with high-noise Internet text information, especially those with large word length spans. Summary of the invention

[0005] In order to provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. The summary is not an extensive review, nor is it intended to identify key / critical components or delineate the scope of protection of these embodiments, but rather serves as a prelude to the detailed description that follows.

[0006] The disclosed embodiments provide a method and system for detecting noisy text information. The deep learning text classification method mainly extracts and represents features of the text through a deep neural network model, and performs classification judgment. This type of method has the problems of low accuracy and poor robustness when facing the detection task of high-noise text information on the Internet, especially specific content with a large word length span.

[0007] In some embodiments, a method for detecting noisy text information is provided, the method comprising:

[0008] Get the target dataset;

[0009] Preprocess the target data set to obtain preprocessed text features;

[0010] Constructing a Chinese text classification model pre-training model, wherein the Chinese text classification model pre-training model is used for pre-processed text features to obtain a representation vector of the text;

[0011] Construct two pre-trained models based on Chinese text classification with different structures as base classifiers A and B of the dual-tower model;

[0012] The base classifier A and base classifier B are tuned using the same dataset.

[0013] Preferably, the base classifier A and the base classifier B are adjusted by a voting algorithm, which comprehensively considers the prediction results of the base classifier A and the base classifier B to obtain the final output result of the system.

[0014] Preferably, the process of the voting algorithm is that the label will be output only when the labels in base classifier A and base classifier B are consistent and the corresponding base classifier A and base classifier B have a confidence level greater than or equal to a preset ratio, otherwise it enters the automatic labeling and manual labeling process.

[0015] Preferably, the automatic labeling and manual labeling process is entered, including:

[0016] When the results of base classifier A and base classifier B are inconsistent, or the results of the two base classifiers are consistent but the confidence is not high, the automatic labeling algorithm is entered;

[0017] When any one of the base classifiers A and B determines that the label is the first specific content, the data content is determined to be the first specific content, the first specific content label is output, and manual labeling is performed;

[0018] When any one of base classifier A and base classifier B determines that the label is the second specific content, the system determines that the data content is the second specific content. When the confidence given by the base classifier that outputs the second specific content label is greater than the preset ratio, and the text content does not contain other specific content keywords but contains the second specific content keywords, the data is automatically labeled as the second specific content. If it does not contain the second specific content keywords but contains other specific content keywords, it is automatically labeled as other specific content. In other cases, manual labeling is performed;

[0019] When one of the base classifiers A and B determines that the label is the third specific content, the data content is determined to be the third specific content. When the confidence given by the base classifier that outputs the third specific content label is greater than the preset ratio, and the text content contains the third specific content keywords, it is automatically labeled as the third specific content. In other cases, manual labeling is performed.

[0020] Preferably, obtaining the target data set includes: obtaining the target data set containing noise from an Internet platform by crawling or manually collecting.

[0021] Preferably, preprocessing the target data set to obtain preprocessed text features includes:

[0022] The collected target data set is subjected to text cleaning, word segmentation and normalization. The text cleaning operation includes cleaning special symbols, format conversion, filtering stop words, removing stop words and meaningless characters based on rule-based template matching methods, removing non-critical information other than text content information, performing word segmentation and indexing, and converting the original data set into a format that can be processed by the model.

[0023] In some embodiments, a detection system for noisy text information is disclosed, including:

[0024] A dataset module, configured to obtain a target dataset;

[0025] A preprocessing module is configured to preprocess the target data set and obtain preprocessed text features;

[0026] A pre-training model module is configured to form a Chinese text classification model pre-training model, wherein the Chinese text classification model pre-training model is used for pre-processed text features to obtain a representation vector of the text;

[0027] The dual-tower model module is configured to construct two pre-trained models based on Chinese text classification with different structures as base classifiers A and B of the dual-tower model;

[0028] The adjustment module is configured to adjust the base classifier A and the base classifier B using the same data set.

[0029] Preferably, the system further includes: a voting algorithm module configured to comprehensively consider the prediction results of the base classifier A and the base classifier B through a voting algorithm to obtain the final output result of the system.

[0030] In some embodiments, an electronic device is disclosed, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the above-mentioned noise-oriented text information detection method is performed.

[0031] In some embodiments, a computer-readable storage medium is disclosed, on which a computer program is stored, and the program is executed by a processor as described above in the method for detecting noise-oriented text information.

[0032] The present disclosure provides a method and system for detecting noisy text information, which can achieve the following technical effects:

[0033] Using the dual-tower pre-trained model for feature extraction can capture the semantic and contextual information of the text and improve detection accuracy.

[0034] Through the idea of ​​ensemble learning, the prediction results of multiple models are combined to improve the performance and robustness of the overall model.

[0035] The voting learning algorithm is used to integrate the prediction results of multiple models to obtain a more accurate final judgment and improve detection accuracy.

[0036] Incremental learning methods are introduced to improve the robustness of the model and cope with changes in data distribution.

[0037] The above general description and the following description are exemplary and explanatory only and are not intended to limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] One or more embodiments are exemplarily described by corresponding drawings, which do not limit the embodiments. Elements with the same reference numerals in the drawings are shown as similar elements, and the drawings do not constitute a scale limitation, and wherein:

[0039] Figure 1 is a flow chart of a method for detecting noisy text information provided by an embodiment of the present disclosure;

[0040] Figure 2 It is a schematic diagram of a process of pre-processing a target data set provided by an embodiment of the present disclosure;

[0041] Figure 3 is a flowchart of a voting algorithm provided by an embodiment of the present disclosure;

[0042] Figure 4 is a schematic diagram of automatic labeling and manual labeling provided by an embodiment of the present disclosure;

[0043] Figure 5 It is a general task flow chart of detecting specific content of Internet high-noise text information based on a double-tower model provided by an embodiment of the present disclosure;

[0044] Figure 6 An electronic device for detecting noisy text information is provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0045] In order to be able to understand the features and technical contents of the embodiments of the present disclosure in more detail, the implementation of the embodiments of the present disclosure is described in detail below in conjunction with the accompanying drawings. The attached drawings are for reference only and are not used to limit the embodiments of the present disclosure. In the following technical description, for the convenience of explanation, a full understanding of the disclosed embodiments is provided through multiple details. However, one or more embodiments can still be implemented without these details. In other cases, to simplify the drawings, well-known structures and systems can be simplified for display.

[0046] The following description and accompanying drawings fully illustrate specific embodiments of the present invention so that those skilled in the art can practice them. Other embodiments may include structural, logical, electrical, process and other changes. The examples represent possible changes only. Unless explicitly required, separate components and functions are optional, and the order of operation can be changed. The parts and features of some embodiments may be included in or replace the parts and features of other embodiments. The scope of the embodiments of the present invention includes the entire scope of the claims, and all available equivalents of the claims. In this article, each embodiment may be represented individually or generally by the term invention, which is only for convenience, and if more than one invention is disclosed in fact, it is not to automatically limit the scope of the application to any single invention or inventive concept. In this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, without requiring or implying any actual relationship or order between these entities or operations. Moreover, the terms include, comprise or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method or device. In the absence of further restrictions, the elements defined by the statement include a ... do not exclude the presence of other identical elements in the process, method or device including the elements. Each embodiment is described in a progressive manner herein, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the embodiments can be referred to each other. For the methods, products, etc. disclosed in the embodiments, since they correspond to the method part disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0047] The specific content detection methods for Internet text information are mainly divided into two types: traditional template keyword matching methods and text classification methods based on deep learning. At present, traditional specific content detection methods mainly use keyword matching and rule template matching. Such methods have great limitations in specific scenarios. For example, in Internet text information with high noise, such as homophones and typos, the accuracy of keyword matching is often very low. The text classification method based on deep learning mainly uses deep neural network models to extract and represent text features and perform classification judgments. Such methods have problems such as low accuracy and poor robustness when facing specific content detection tasks with high-noise Internet text information, especially those with large word length spans.

[0048] In order to solve the problems existing in the related art, the embodiments of the present disclosure provide a method and system for detecting noisy text information.

[0049] Combination Figure 1 As shown, the present disclosure provides a flow chart of a method for detecting noisy text information, including:

[0050] S100, obtaining a target data set.

[0051] The target dataset containing noise is obtained from the Internet platform by crawling or manually collecting. The dataset contains text content such as posts, blogs, and comments posted by netizens on the Internet. The text content includes elements such as sentences, punctuation, and expressions, as well as noise such as typos, emoticons, and emojis caused by typos.

[0052] S200, preprocessing the target data set to obtain preprocessed text features.

[0053] The collected original target data is cleaned, segmented and normalized. The text cleaning operation includes cleaning special symbols, format conversion, filtering stop words, removing stop words and meaningless characters based on rule template matching, removing non-key information other than text content information, and then segmenting and indexing. Then, relevant models or algorithms are used to convert the original data set into a format that can be processed by the model, and finally output to downstream tasks. Figure 2 , which is a schematic diagram of a process of pre-processing a target data set according to an embodiment of the present disclosure.

[0054] S300, constructing a pre-training model for a Chinese text classification model, wherein the pre-training model for a Chinese text classification model is used for pre-processed text features to obtain a representation vector for the text.

[0055] The pre-trained model of the Chinese text classification model, which is based on the sum of word vectors, block vectors and position vectors, and consists of a 12-layer encoder, 12 attention heads, 768 hidden units, and 110MB parameters, is used to extract the pre-processed text features to obtain the text representation vector. The representation vector can effectively capture the semantic and contextual information of the text, convert the text information into the form of word vectors, and directly serve as the input for fine-tuning the twin tower pre-trained model.

[0056] The Chinese text classification model has been pre-trained on a large-scale Chinese corpus and has good semantic understanding and context modeling capabilities. The pre-processed high-noise Internet text information is input into the Chinese text classification model. Through the model's multi-layer attention mechanism and Transformer architecture, the text is converted into a high-dimensional representation vector. These representation vectors will be embedded through Embedding encoding, integrating the semantic information of each word in the text and its association with the context to form a global text representation.

[0057] The Chinese text classification model has the ability to understand semantics and context, and can better capture the semantic features of the text. By using the Chinese text classification model for feature extraction, we can convert the text into the form of word vectors, so that each word will be represented as a vector. These word vectors have a high dimension. The hidden layer of the Chinese text classification pre-training model is 768, so each word will be converted into a 768-dimensional digital vector. A sentence of length N is represented as a two-dimensional vector matrix of N*768. Through the feature extraction method based on the Chinese text classification pre-training model, the semantic information and contextual relationship of the word in the text can be expressed. This representation method helps the subsequent model fine-tuning process, allowing the model to better understand and distinguish different text information.

[0058] S400, construct two pre-trained models based on Chinese text classification with different structures as base classifier A and base classifier B of the dual-tower model.

[0059] Facing the high-noise text information on the Internet, when a single classification model inevitably produces systematic errors, we draw on the idea of ​​voting algorithms to build a dual-tower pre-training model, and use a multi-model mutual verification method to improve classification accuracy and reduce systematic errors. Specifically, two pre-training models based on Chinese text classification with different structures are constructed as the base classifiers A and B of the dual-tower model, and the base classifiers A and B are fine-tuned using the same data set after preprocessing and feature extraction. After the loss function of the model converges, the parameter files of the fine-tuned classifiers A and B are saved for subsequent judgment of specific content detection. The base classifier A includes a four-layer network structure of a text classifier output vector layer, a full-link layer, a full-link layer, and a network output layer. The base classifier B includes a six-layer network structure of a text classifier output vector layer, a pooling layer, a full-link layer, a full-link layer, a SoftMax layer, and a network output layer.

[0060] S500, adjusting base classifier A and base classifier B using the same data set.

[0061] The two base classifiers are fine-tuned separately using the same batch of data sets, and the actual fine-tuning results are used. The adjustment of base classifier A and base classifier B is done by a voting algorithm, which comprehensively considers the prediction results of base classifier A and base classifier B to obtain the final output of the system. It should be noted that the number of sources in the same data set is not specifically limited.

[0062] The voting algorithm is a kind of ensemble learning algorithm widely used in classification algorithms. There are two main types of voting: hard voting and soft voting. Hard voting is voting with the same weight for each base classifier, and its principle is the majority voting principle: if a classification result of a base classifier exceeds half, the ensemble algorithm selects the result; if there is no half result, there is no output. The principle of soft voting is also majority voting, but the weight of each base classifier's vote can be defined by yourself. When the classification effects of each base classifier are quite different, soft voting should be selected to give the base classifier a larger weight, so as to obtain better classification performance, thereby optimizing the classification results. In the present invention, a method similar to soft voting is adopted.

[0063] Both base classifier A and base classifier B are based on a text classification model consisting of 12 encoder layers, 12 attention heads, and 768 hidden units. The classification model uses large-scale unlabeled corpus for training to obtain a representation of the text that contains rich semantic information, and then fine-tunes the semantic representation of the text in a specific task and finally applies it to the task.

[0064] After the two base classifiers of the dual-tower pre-training model are fine-tuned on the same batch of data, the loss functions will converge. The fine-tuning process will automatically stop after the loss function value no longer decreases after more than 1200 rounds. The parameter value when the loss function is the smallest is retained as the model file to judge new Internet high-noise text information. The judgment result is the category of the pre-set classification task.

[0065] Specifically, Internet-specific content refers to topics that are currently very popular and attract a lot of attention from netizens. As time goes by and hot topics migrate and evolve, many new types of Internet-specific content will emerge. At the same time, the current types of hot topics of Internet-specific content will decline or even be eliminated. This phenomenon is called topic concept drift of Internet-specific content. The relevant processing methods will be elaborated in detail in the re-optimization mechanism after the model performance declines.

[0066] The task of detecting specific Internet content can be broadly defined as using natural language processing and machine learning techniques, utilizing text features or a series of different high-order information features in social media data, and realizing automatic detection of specific Internet content through supervised learning or semi-supervised learning methods. The solution of the present invention defines Internet specific content detection as a multi-classification problem. By fine-tuning the dual-tower pre-trained large model, it is able to predict a large number of new unlabeled data samples by learning a small number of labeled multi-classification samples and combining its cognitive ability accumulated in the pre-training stage. The two base classifiers of the dual-tower pre-trained large model respectively give the classification probability and related confidence for the multi-classification label for the same data. Both base classifiers convert the confidence into a value between 0 and 1 through the Softmax layer, and select the multi-probability and judgment results of which category a certain data sample belongs to in the multi-classification task according to the sorting.

[0067] Specifically, the judgment process of the two base classifiers of the dual-tower pre-training model is as follows:

[0068] F = softmax(WTv+b),

[0069] Where F is a two-dimensional array, F∈(L, P), L∈Rn, n is the number of categories equal to 5, representing five classification labels of specific content, P is the confidence corresponding to L; W is the original encoding of the text content of the corresponding data, WT represents the Embedding encoding of the original text information based on the text classification pre-training model, and converts it into a vector model, v is the final feature vector formed after encoding a single sentence, the latitude is (N*768), N is the character length of the sentence, 768 is the hidden layer size of the pre-training model; b is the representation of the noise part of the Internet information with high noise. For each piece of Internet text information to be predicted, the base classifier A and the base classifier B will simultaneously predict the label of the text information and its corresponding confidence. The dual-tower pre-training model system will use a voting algorithm to comprehensively consider the prediction results of the base classifier A and the base classifier B to obtain the final output result of the system. See Figure 3 , which is a flowchart of a voting algorithm according to an embodiment of the present disclosure.

[0070] In general, labels are output only when the labels of the two models are consistent and the confidence of the two models in the labels is greater than or equal to 95%. Otherwise, other situations will enter the automatic labeling and manual labeling process.

[0071] See also Figure 4 , which is a schematic diagram of automatic labeling and manual labeling in an embodiment of the present disclosure.

[0072] When the results of base classifier A and base classifier B are inconsistent, or the results of the two base classifiers are consistent but the confidence is not high, the automatic labeling algorithm is entered;

[0073] When any one of the base classifiers A and B determines that the label is the first specific content, the data content is determined to be the first specific content, the first specific content label is output, and manual labeling is performed;

[0074] When any one of base classifier A and base classifier B determines that the label is the second specific content, the system determines that the data content is the second specific content. When the confidence given by the base classifier that outputs the second specific content label is greater than the preset ratio, and the text content does not contain other specific content keywords but contains the second specific content keywords, the data is automatically labeled as the second specific content. If it does not contain the second specific content keywords but contains other specific content keywords, it is automatically labeled as other specific content. In other cases, manual labeling is performed;

[0075] When one of the base classifiers A and B determines that the label is the third specific content, the data content is determined to be the third specific content. When the confidence given by the base classifier that outputs the third specific content label is greater than the preset ratio, and the text content contains the third specific content keywords, it is automatically labeled as the third specific content. In other cases, manual labeling is performed.

[0076] The results of automatic annotation will be directly output as the results of the dual-tower pre-training model specific content discovery system. For manually annotated data, the system will temporarily not output the judgment results but store the relevant data. Manual judgment is required. After the judgment is completed, the results of manual annotation will be uploaded and fed back to the server, and the base classifiers A and B will be retrained and iterated to improve the accuracy of the dual-tower pre-training model specific content detection and achieve detection. Correspondingly, see Figure 4 , which is a general task flow chart of specific content detection of high-noise text information on the Internet based on the double-tower model in an embodiment of the present disclosure.

[0077] The disclosed embodiment also provides a re-optimization mechanism after the performance of the dual-tower pre-training model has declined.

[0078] Since the topics of high-noise text information on the Internet are not fixed, but will migrate all the time with factors such as hot events, breaking news, and public opinion orientation, the phenomenon of concept drift will occur, that is, the hot topics of the previous period may drop sharply or even disappear in the current period. At the same time, with the emergence of new hot topics, the detection of specific content also needs to be able to adapt to the continuous evolution of topics and cope with the phenomenon of concept drift, otherwise it will lead to a sharp decline in the accuracy of specific content detection. The dual-tower pre-training model specific content detection system proposed in the present invention uses an inherited learning voting algorithm, and the results of the base classifier A and the base classifier B will undergo system and manual secondary research and judgment, which can effectively cope with topic changes caused by concept drift. When the dual-tower pre-training model detects a significant change in the data distribution caused by the concept drift phenomenon or the evolution of the topic, resulting in a decrease in the accuracy of the system output, a re-optimization mechanism can be adopted, such as introducing incremental learning or domain adaptation methods, to cope with these changes and improve the robustness and performance of the model. Specifically, the re-optimization mechanism after the performance of the dual-tower pre-training model declines is divided into the following four steps:

[0079] S1100, concept drift detection.

[0080] The dual-tower pre-trained model specific content detection system proposed in the embodiment of the present disclosure will continuously detect the accuracy of the system's judgment. Specifically, when concept drift occurs or the topic changes suddenly, the results and confidence levels output by the two base classifiers of the dual-tower pre-trained model will be greatly affected, resulting in an increase in the amount of data that needs to be automatically and manually labeled. By setting a threshold, if the amount of manually labeled data in the same time period increases by more than the threshold, it is determined that concept drift is detected.

[0081] S1200, introduces an adaptive adjustment mechanism.

[0082] When concept drift is detected, all new topics or new types of specific content need to be specially labeled during manual labeling and the relevant data sets need to be stored. When the number of data sets for new topics is greater than 200, an adaptive adjustment mechanism can be introduced to iteratively update the two base classifiers of the dual-tower pre-training model specific content detection system, thereby improving the overall accuracy of the system.

[0083] S1300, incremental learning updates two base classifiers.

[0084] After detecting concept drift, the Twin Towers pre-trained model specific content detection system can use incremental learning methods to update the base classifier. Incremental learning refers to further training on the basis of existing models using new data to adapt to new topics or new types of specific content. The system can combine newly annotated data with existing data sets and use incremental learning algorithms to update the base classifier to improve the accuracy and generalization ability of the model. Specific methods of incremental learning can use online learning techniques, such as online gradient descent or online random forests. These methods can use new data to update models and adjust parameters while retaining existing knowledge.

[0085] S1400, re-predict data with low confidence.

[0086] After incremental learning, the system can re-predict data with low confidence during concept drift. This data may be due to inaccurate or uncertain judgments of the classifier caused by concept drift. By re-predicting this data, their classification accuracy can be improved, thereby improving the performance of the overall system. Data with low re-prediction confidence can be reclassified using the new model. By re-analyzing this data, their classification accuracy can be improved and the output results of the system can be corrected.

[0087] The above four steps combined can enable the dual-tower pre-trained model specific content detection system to maintain a high accuracy rate in the case of concept drift and topic changes, and be able to adapt to the ever-changing high-noise text information on the Internet.

[0088] Compared with the prior art, the disclosed embodiment utilizes a dual-tower pre-training model for feature extraction, which can capture the semantic and contextual information of the text and improve detection accuracy.

[0089] The introduction of the twin tower pre-training model is one of the key advantages of the present invention. Traditional specific content detection methods usually use models based on rules or shallow features, which often do not work well when faced with complex text semantics and contextual information. The twin tower pre-training model has powerful cognitive ability and attention mechanism. By using pre-training of large-scale corpus, it can better understand the text content and learn rich semantic and contextual information. The feature extraction ability of the pre-training model enables the system to effectively capture the potential features and patterns of specific content. Therefore, when new text data is input, the pre-training model can accurately determine whether there is specific content by comparing the matching degree of the learned semantic representation with the input text. The introduction of the twin tower pre-training model also helps to solve the problem of insufficient information caused by the large span of word length. Traditional methods are prone to losing key information or being restricted when processing long texts, while the pre-training model can effectively capture important fragments and key information in the text through the self-attention mechanism, thereby improving the correct detection and classification capabilities of specific content.

[0090] Through the idea of ​​ensemble learning, the prediction results of multiple models are combined to improve the performance and robustness of the overall model.

[0091] In the detection of specific content, a single model may have certain limitations and cannot fully meet the needs in various scenarios. In order to overcome this problem, the present invention adopts the idea of ​​ensemble learning, combining the prediction results of multiple models to improve the performance and robustness of the overall model. By integrating the advantages of multiple models, specific content can be modeled and detected from different angles and feature spaces. Each model has its own advantages and characteristics. By combining their prediction results, a more comprehensive and accurate judgment can be obtained. Ensemble learning can integrate the prediction results of multiple models through voting, weighted average or confidence-based methods to obtain the final judgment result. Ensemble learning can also use strategies such as stacking or boosting to further improve model performance. Stacking refers to taking the prediction results of multiple models as input and then making the final prediction through another model. This method can utilize the complementarity between different models to further improve the prediction ability of the overall model. Boosting refers to training multiple models in a certain order, and each model will be adjusted and optimized according to the prediction results of the previous model to gradually improve the performance of the model. Through the idea of ​​ensemble learning, the present invention can make full use of the advantages of different models and make up for the shortcomings of a single model. Ensemble learning can improve the generalization and robustness of the model, so that it can achieve better performance in different scenarios and data distributions. At the same time, through mutual verification and comprehensive consideration between models, the misjudgment rate can be reduced and the reliability of the overall system can be improved. In addition to combining the prediction results of multiple models, ensemble learning can also be combined with other technical means to further improve model performance. For example, methods such as sample weighting, feature selection, and feature combination can be introduced to optimize the input and feature representation of the model. These technical means can enhance the model's sensitivity to specific content and improve the recall and accuracy of detection.

[0092] The voting learning algorithm is used to integrate the prediction results of multiple models to obtain a more accurate final judgment and improve detection accuracy.

[0093] The voting learning algorithm is another important advantage of the present invention. In specific content detection, the judgment result of a single model may be affected by data noise or model uncertainty, resulting in low detection accuracy. In order to overcome this problem, the present invention adopts a voting learning algorithm to obtain a more accurate final judgment by integrating the prediction results of multiple models. When the same data is input into two fine-tuned dual-tower pre-training models, the system compares their output results and uses a voting algorithm to perform mutual verification between models. Through the consistency voting of multiple models, the misjudgment rate can be reduced and the detection accuracy can be improved. If the judgment results of the two models are consistent and the confidence is greater than the threshold, the information is directly used as the output result of the system. This method of comprehensively considering the judgment results of multiple models effectively makes up for the limitations of a single model and improves the performance of the overall system. The voting learning algorithm can also be further expanded to introduce strategies such as weighted voting and dynamic adjustment of thresholds to adjust the weight of the final judgment according to the performance and confidence of the model to further improve the detection accuracy. This method of integrating the prediction results of multiple models can not only improve the overall detection accuracy, but also increase the robustness and reliability of the system.

[0094] Introducing incremental learning methods to improve the robustness of the model and cope with changes in data distribution

[0095] In practical applications, changes in data distribution or concept drift are common situations. Traditional machine learning methods may suffer from performance degradation when facing these challenges because they usually need to retrain the model to adapt to the new data distribution. However, the introduction of incremental learning methods can effectively solve this problem. Incremental learning is an online learning method that can gradually learn and accumulate new knowledge without retraining the entire model. This feature enables the model to quickly adapt to new data while retaining old knowledge, thereby improving the robustness and adaptability of the model. Through incremental learning, the model can be updated according to the characteristics of new data without losing previously learned knowledge. This ability enables the model to continuously learn and improve in a changing environment, thereby better adapting to changes in data distribution and concept drift. Another benefit of incremental learning is that it can save computing resources and storage space. Compared with retraining the entire model, incremental learning only requires incremental updates to new data, so it can greatly reduce computing overhead and time costs. In addition, incremental learning can also reduce the demand for storage space because it only needs to store relevant information about new data without saving the data of the entire training set. In addition, the introduction of incremental learning methods can also effectively improve the flexibility and scalability of the model. Through incremental learning, the model can continuously learn from new data and gradually improve its performance and generalization ability. This iterative learning process enables the model to adapt to the ever-changing environment and needs, and has stronger adaptability and sustainability.

[0096] The present disclosure embodiment discloses a detection system for noisy text information, including:

[0097] A dataset module, configured to obtain a target dataset;

[0098] A preprocessing module is configured to preprocess the target data set and obtain preprocessed text features;

[0099] A pre-training model module is configured to form a Chinese text classification model pre-training model, wherein the Chinese text classification model pre-training model is used for pre-processed text features to obtain a representation vector of the text;

[0100] The dual-tower model module is configured to construct two pre-trained models based on Chinese text classification with different structures as base classifiers A and B of the dual-tower model;

[0101] The adjustment module is configured to adjust the base classifier A and the base classifier B using the same data set.

[0102] Preferably, the system further includes: a voting algorithm module configured to comprehensively consider the prediction results of the base classifier A and the base classifier B through a voting algorithm to obtain the final output result of the system.

[0103] Combination Figure 6 As shown, the embodiment of the present disclosure provides an electronic device for detecting noisy text information, including a processor and a memory. Optionally, the electronic device may also include a communication interface and a bus. The processor, the communication interface, and the memory may communicate with each other through the bus. The communication interface may be used for information transmission. The processor may call the logic instructions in the memory to execute the detection method for noisy text information of the above embodiment.

[0104] An embodiment of the present disclosure provides a storage medium storing computer executable instructions, wherein the computer executable instructions are configured to execute the above-mentioned detection method for noise-oriented text information.

[0105] The above description and the accompanying drawings fully illustrate the embodiments of the present disclosure so that those skilled in the art can practice them. Other embodiments may include structural, logical, electrical, process and other changes. The embodiments represent only possible changes. Unless explicitly required, individual components and functions are optional, and the order of operations may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the words used in this application are only used to describe the embodiments and are not used to limit the claims. As used in the description of the embodiments and the claims, unless the context clearly indicates otherwise, the singular forms of one (a), one (an) and said (the) are intended to include plural forms as well. Similarly, the terms and / or as used in this application refer to any and all possible combinations of one or more associated listings. In addition, when used in this application, the terms include (comprise) and its variants include (comprises) and / or include (comprising) refer to the existence of stated features, wholes, steps, operations, elements, and / or components, but do not exclude the existence or addition of one or more other features, wholes, steps, operations, elements, components and / or these groups. In the absence of further restrictions, an element defined by a statement including an ... does not exclude the existence of other identical elements in the process, method or device including the element. In this article, each embodiment may focus on the differences from other embodiments, and the same or similar parts between the embodiments may refer to each other. For the methods, products, etc. disclosed in the embodiments, if they correspond to the method part disclosed in the embodiments, then the relevant parts can refer to the description of the method part.

[0106] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software may depend on the specific application and design constraints of the technical solution. The technicians may use different methods for each specific application to implement the described functions, but such implementations should not be considered to exceed the scope of the embodiments of the present disclosure. The technicians may clearly understand that, for the convenience and simplicity of description, the specific working processes of the systems, systems and units described above may refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here.

[0107] The flowchart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to the embodiment of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. In some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, which can depend on the functions involved. In the description corresponding to the flowchart and the block diagram in the accompanying drawings, the operations or steps corresponding to different boxes can also occur in a different order from the order disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, which can depend on the functions involved. Each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system that performs the specified functions or actions, or may be implemented by a combination of dedicated hardware and computer instructions.

Claims

1. A method for detecting noisy text information, characterized in that: The method comprises: Get the target dataset; Preprocess the target data set to obtain preprocessed text features; Constructing a Chinese text classification model pre-training model, wherein the Chinese text classification model pre-training model is used for pre-processed text features to obtain a representation vector of the text; Construct two pre-trained models based on Chinese text classification with different structures as base classifiers A and B of the dual-tower model; The base classifier A and base classifier B are tuned using the same dataset.

2. The method for detecting noisy text information according to claim 1, characterized in that: The adjustment of base classifier A and base classifier B is carried out through a voting algorithm, which comprehensively considers the prediction results of base classifier A and base classifier B to obtain the final output result of the system.

3. The method for detecting noisy text information according to claim 2, characterized in that: The process of the voting algorithm is that the label will be output only when the labels in base classifier A and base classifier B are consistent and the corresponding base classifier A and base classifier B have a confidence level greater than or equal to the preset ratio. Otherwise, the automatic labeling and manual labeling process will be entered.

4. The method for detecting noisy text information according to claim 3, characterized in that: Enter the automatic and manual annotation process, including: When the results of base classifier A and base classifier B are inconsistent, or the results of the two base classifiers are consistent but the confidence is not high, the automatic labeling algorithm is entered; When any one of the base classifiers A and B determines that the label is the first specific content, the data content is determined to be the first specific content, the first specific content label is output, and manual labeling is performed; When any one of base classifier A and base classifier B determines that the label is the second specific content, the system determines that the data content is the second specific content. When the confidence given by the base classifier that outputs the second specific content label is greater than the preset ratio, and the text content does not contain other specific content keywords but contains the second specific content keywords, the data is automatically labeled as the second specific content. If it does not contain the second specific content keywords but contains other specific content keywords, it is automatically labeled as other specific content. In other cases, manual labeling is performed; When one of the base classifiers A and B determines that the label is the third specific content, the data content is determined to be the third specific content. When the confidence given by the base classifier that outputs the third specific content label is greater than the preset ratio, and the text content contains the third specific content keywords, it is automatically labeled as the third specific content. In other cases, manual labeling is performed.

5. The method for detecting noisy text information according to claim 1, characterized in that: Obtaining a target data set includes: obtaining a target data set containing noise from an Internet platform by crawling or manually collecting.

6. The method for detecting noisy text information according to claim 1, characterized in that: Preprocess the target data set to obtain the preprocessed text features, including: The collected target data set is subjected to text cleaning, word segmentation and normalization. The text cleaning operation includes cleaning special symbols, format conversion, filtering stop words, removing stop words and meaningless characters based on rule-based template matching methods, removing non-critical information other than text content information, performing word segmentation and indexing, and converting the original data set into a format that can be processed by the model.

7. A detection system for noisy text information, characterized in that: include: A dataset module, configured to obtain a target dataset; A preprocessing module is configured to preprocess the target data set and obtain preprocessed text features; A pre-training model module is configured to form a Chinese text classification model pre-training model, wherein the Chinese text classification model pre-training model is used for pre-processed text features to obtain a representation vector of the text; The dual-tower model module is configured to construct two pre-trained models based on Chinese text classification with different structures as base classifiers A and B of the dual-tower model; The adjustment module is configured to adjust the base classifier A and the base classifier B using the same data set.

8. The noise text information detection system according to claim 1, characterized in that: The system further comprises: The voting algorithm module is configured to comprehensively consider the prediction results of the base classifier A and the base classifier B through a voting algorithm to obtain the final output result of the system.

9. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images