Information detection method and electronic equipment

Through the combination of multiple fine-tuned models and pre-trained classification models, the low-precision problem of the existing technology in complex information and long-context text recognition is solved, high-precision emotion detection is achieved, manual intervention is reduced, and it is suitable for massive information. analyze.

CN118964841APending Publication Date: 2024-11-15EVERSEC BEIJING TECH
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410979874.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

When processing complex information and long context-dependent text, natural language processing methods such as classification models combined with word2vec and LSTM or rule determination modes combined with dictionary are difficult to meet high-precision requirements and cannot effectively identify information. Emotional tendency.

Method used

Multiple fine-tuned models (such as Roberta-large, macBert-large, electra-large) are used to make emotional judgments on the information data, and combined with pre-trained classification model and sigmoid regression function to fusion results to automatically obtain the emotional labels and emotional scores of the information data. value, determine the detection result.

Benefits of technology

It improves the accuracy of information detection, reduces the need for manual analysis and judgment, and can more accurately identify the emotional tendency of information data, and is suitable for the analysis of massive information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964841A_ABST
    Figure CN118964841A_ABST
Patent Text Reader

Abstract

The invention provides an information detection method and electronic equipment, and the method comprises the steps: obtaining information data in response to determining that the information data meets a preset condition; performing information emotion judgment on the information data by adopting a preset database or a plurality of fine adjustment models to obtain an emotion label of the information data and an emotion score corresponding to the emotion label; and determining a detection result of the information data according to the emotion score. According to the method and the device, the emotion label of the information data and the emotion score corresponding to the emotion label are obtained by performing judgment of the multiple fine adjustment models on the information data, so that the detection result is obtained, the detection result obtained through the method and the device is more accurate, and the result is automatically obtained through the models without manual research and judgment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of deep learning technology, and in particular to an information detection method and electronic device. Background Art

[0002] In the age of universal Internet, information is always available and risks are everywhere. At present, the total number of Chinese netizens has exceeded 1 billion, and the Internet has become an important channel for people to express their personal opinions and views. Social public opinion has also shifted from offline to online. Internet information has become an important part of information, bringing all-round challenges to information work in the omnimedia era. AI information systems have improved the accuracy of data. Early information monitoring usually used "keywords" combined with "and, or, and not" judgment logic to retrieve data, which often required a lot of manual work to perform secondary processing of the data. Intelligent monitoring uses natural language processing technology to identify content in multiple dimensions, thereby improving the accuracy of data.

[0003] Existing solutions either use traditional natural language processing methods, such as word2vec and LSTM combined classification models to judge information, or use regular expressions combined with dictionaries to make rule-based judgments. The above models perform poorly on complex information and texts with long context dependencies, and are far from meeting high-precision requirements. Summary of the invention

[0004] The present disclosure provides an information detection method and an electronic device to at least solve the above technical problems existing in the prior art.

[0005] According to a first aspect of the present disclosure, there is provided an information detection method, the method comprising:

[0006] In response to determining that the information data satisfies a preset condition, acquiring the information data;

[0007] Using a preset database or multiple fine-tuning models to perform information sentiment determination on the information data, and obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels;

[0008] The detection result of the information data is determined according to the sentiment score.

[0009] In one possible implementation manner, the determining that the information data satisfies a preset condition includes:

[0010] Determining whether the information data is regional data or organizational data;

[0011] When the information data is regional data or organizational data, it is determined that the information data meets a preset condition.

[0012] In one possible implementation manner, when the information data is non-regional data or non-organizational data, it also includes:

[0013] The non-regional data or non-organizational data is detected by using a preset detection method to obtain the emotion labels of the non-regional data or non-organizational data and the emotion scores corresponding to the emotion labels.

[0014] In one possible implementation manner, the acquiring information data includes:

[0015] The information data is cleaned to obtain cleaned information data.

[0016] In one possible implementation, the using of a preset database or a plurality of fine-tuned models to perform information sentiment determination on the information data to obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels includes:

[0017] Determine whether the cleaned information data has been analyzed;

[0018] If it has been determined, the cleaned information data is matched with the labeled data in the preset database to obtain the emotion label of the information data and the emotion score corresponding to the emotion label;

[0019] Otherwise, multiple parallel fine-tuning models are used to perform information sentiment judgment on the information data to obtain sentiment results, and the sentiment results are input into a pre-trained classification model for classification to obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels.

[0020] In one possible implementation, the plurality of parallel fine-tuning models include: a first fine-tuning model, a second fine-tuning model, and a third fine-tuning model; the plurality of parallel fine-tuning models are used to perform information emotion judgment on the information data respectively to obtain an emotion result, including:

[0021] The first fine-tuning model performs information sentiment judgment on the information data to obtain a first result;

[0022] The second fine-tuning model performs information sentiment judgment on the information data to obtain a second result;

[0023] The third fine-tuning model performs information sentiment judgment on the information data to obtain a third result;

[0024] The first result, the second result and the third result are integrated to obtain a sentiment result.

[0025] In one possible implementation, a sigmoid regression function is used to fuse the first result, the second result, and the third result to obtain a sentiment result.

[0026] In one possible implementation manner, determining the detection result of the information data according to the emotion score includes:

[0027] According to the correspondence between the preset threshold range and the information detection result, the information detection result corresponding to the threshold range to which the emotion score belongs is determined as the detection result of the information data.

[0028] In one possible implementation, the first fine-tuning model, the second fine-tuning model, the third fine-tuning model, and the pre-trained classification model are updated every preset time period.

[0029] According to a second aspect of the present disclosure, there is provided an information detection device, the device comprising:

[0030] an acquisition module, configured to acquire the information data in response to determining that the information data satisfies a preset condition;

[0031] A determination module, used to perform information sentiment determination on the information data using a preset database or multiple fine-tuned models, to obtain a sentiment label of the information data and a sentiment score corresponding to the sentiment label;

[0032] A detection module is used to determine the detection result of the information data according to the emotion score.

[0033] According to a third aspect of the present disclosure, there is provided an electronic device, including:

[0034] at least one processor; and

[0035] a memory communicatively connected to the at least one processor; wherein,

[0036] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in the present disclosure.

[0037] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause the computer to execute the method described in the present disclosure.

[0038] The information detection method and electronic device disclosed in the present invention obtain the emotion labels of the information data and the emotion scores corresponding to the emotion labels by performing multiple fine-tuning models on the information data, thereby obtaining the detection results. The detection results obtained by the present application are more accurate and the results are automatically obtained through the model without manual analysis.

[0039] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0041] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts.

[0042] Figure 1 The schematic diagram shows the implementation process of the information detection method of the embodiment of the present disclosure. Figure 1 ;

[0043] Figure 2 The schematic diagram shows the implementation process of the information detection method of the embodiment of the present disclosure. Figure 2 ;

[0044] Figure 3 A schematic diagram showing the structure of an information detection device according to an embodiment of the present disclosure is shown;

[0045] Figure 4 A schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0046] In order to make the purpose, features, and advantages of the present disclosure more obvious and easy to understand, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present disclosure.

[0047] Intelligent information management is inevitable for the development of the industry. With the rapid development of self-media and mobile terminals, in less than 10 years, the total amount of daily information data has expanded from millions to tens of millions, and has now reached hundreds of millions. Such a huge amount of information far exceeds the limit of normal reading. To analyze the massive amount of information, it is necessary to use artificial intelligence technology to process and judge the information.

[0048] (1) A review of research methods on text sentiment analysis

[0049] Wang Ting, Yang Wenzhong and others proposed that text sentiment analysis is an important branch of natural language processing and is widely used in information analysis and content recommendation. It is also a hot topic in recent years. According to the different methods adopted, it is divided into sentiment analysis based on sentiment dictionary, sentiment analysis based on traditional machine learning, and sentiment analysis based on deep learning. By comparing these three methods, the research results were analyzed, the advantages and disadvantages of different methods were summarized, the relevant data sets and evaluation indicators and application scenarios were introduced, and the analysis of sentiment subtasks was briefly summarized. The future research trends and application fields of sentiment analysis problems were discovered. Provide certain help and guidance for researchers in related fields.

[0050] (2) Microblog sentiment analysis based on deep learning

[0051] The deep learning-based microblog sentiment analysis proposed by Liang Jun, Chai Yumei, Yuan Huibin and others aims to discover users' opinions and attitudes on hot events. Most existing studies use traditional algorithms such as SVM and CRF to analyze microblog sentiment based on manually annotated sentiment features. This paper mainly explores the feasibility of using deep learning to perform sentiment analysis on Chinese microblogs, using recursive neural networks to discover task-related features, avoiding artificial feature design that depends on specific tasks, and introducing a sentiment polarity transfer model based on the correlation between words in sentences to enhance the capture of text correlation. The method proposed in this paper is comparable in performance to the current method of manually annotating sentiment features, but saves a lot of manual annotation workload.

[0052] (3) Mining based on sentiment words and frequent patterns

[0053] Zhou Lizhu, He Yukai and others proposed that due to the rapid increase in Web texts, mining and analyzing these texts, especially the comments data actively posted by users, identifying their sentiment trends and evolution laws can better understand users' consumption habits, analyze hot information, and provide important decision-making basis for enterprises, governments and other institutions. First, the research object and goal of sentiment analysis are defined and explained, and the basic research ideas are given. Then, in the task of subjective sentence recognition, the main processing methods are reviewed and analyzed in detail; in the feature extraction of opinion classification, two mainstream processing ideas are introduced and discussed in detail - based on sentiment words and based on frequent pattern mining. Then, some other related sentiment analysis issues are briefly introduced. Finally, the existing achievements and shortcomings of sentiment analysis, as well as the challenges it faces, are summarized, and its development prospects are prospected.

[0054] The information detection method and electronic device provided in the embodiments of the present application are described below in conjunction with the accompanying drawings.

[0055] like Figure 1 As shown, the present application provides an information detection method, the method comprising:

[0056] S101, in response to determining that the information data meets a preset condition, acquiring the information data;

[0057] In some optional embodiments, the determining that the information data satisfies a preset condition includes:

[0058] Determining whether the information data is regional data or organizational data;

[0059] When the information data is regional data or organizational data, it is determined that the information data meets a preset condition.

[0060] Among them, the preset condition in this application is that the information data is regional data or organizational data, that is, when the information data is regional data or organizational data, it is considered that the information data meets the preset condition, and the information data is obtained at this time to perform subsequent steps S102 and S103.

[0061] In some optional embodiments, when the information data is non-regional data or non-organizational data, it also includes:

[0062] The non-regional data or non-organizational data is detected by using a preset detection method to obtain the emotion labels of the non-regional data or non-organizational data and the emotion scores corresponding to the emotion labels.

[0063] In some optional embodiments, the obtaining of information data includes:

[0064] The information data is cleaned to obtain cleaned information data.

[0065] Specifically, Figure 2 As shown, the present application first cleans the information data, for example, removing duplicate data, redundant data or invalid data in the information data. Then, it is determined whether the cleaned information data is regional data or organizational data. If not, the present application uses the suggested sentiment analysis template and the suggested label determination template to analyze and determine the non-regional data or non-organizational data, determine the sentiment label of the information data and the sentiment score corresponding to the sentiment label, and then send it to the prefecture-level city.

[0066] S102, using a preset database or multiple fine-tuning models to perform information sentiment determination on the information data, and obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels;

[0067] In some optional embodiments, the using of a preset database or a plurality of fine-tuned models to perform information sentiment determination on the information data to obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels includes:

[0068] Determine whether the cleaned information data has been analyzed;

[0069] If it has been determined, the cleaned information data is matched with the labeled data in the preset database to obtain the emotion label of the information data and the emotion score corresponding to the emotion label;

[0070] Otherwise, multiple parallel fine-tuning models are used to perform information sentiment judgment on the information data to obtain sentiment results, and the sentiment results are input into a pre-trained classification model for classification to obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels.

[0071] Specifically, for information data determined to be regional data or organizational data, the information data is first analyzed to determine whether the information data has been manually analyzed for the same data. If it has been analyzed, the information data is matched with the labeled data in the preset database, so that the sentiment label of the information data and the corresponding sentiment score of the sentiment label can be obtained according to the labeled data. The preset database in this application is a redis database. The redis database contains historical manually analyzed labeled data. When the information data is matched with the labeled data, if the match is successful, the manual labeling result is directly output. If it has not been manually labeled, the subsequent process is entered.

[0072] The subsequent process is to use multiple parallel fine-tuning models to perform information emotion judgment on the information data respectively to obtain emotion results, input the emotion results into the pre-trained classification model for classification, and obtain the emotion label of the information data and the emotion score corresponding to the emotion label, wherein the multiple parallel fine-tuning models include a first fine-tuning model, a second fine-tuning model and a third fine-tuning model, the first fine-tuning model performs information emotion judgment on the information data to obtain a first result; the second fine-tuning model performs information emotion judgment on the information data to obtain a second result; the third fine-tuning model performs information emotion judgment on the information data to obtain a third result; the first result, the second result and the third result are fused to obtain the emotion result, and then the first result, the second result and the third result are fused to obtain the emotion result. Among them, in this application, the first fine-tuning model adopts the Roberta-large fine-tuning model, the second fine-tuning model adopts the macBert-large fine-tuning model, and the third fine-tuning model adopts the electra-large fine-tuning model. The Roberta-large fine-tuning model, the macBert-large fine-tuning model and the electra-large fine-tuning model perform information sentiment analysis and judgment on the data. A sentiment analysis rule model based on rule recognition (containing approximately 20,000 regular expression rules) judges the information sentiment, and the final result is model-fused to output a comprehensive judgment result to determine the information sentiment tendency of the data, including positive, neutral and negative; then the comprehensive judgment result is input into the pre-trained classification model for classification to obtain the sentiment label and the corresponding sentiment score.

[0073] Specifically, in the present application, a sigmoid regression function is used to fuse the first result, the second result and the third result to obtain an emotional result.

[0074] It should be noted that the first fine-tuning model, the second fine-tuning model, the third fine-tuning model and the pre-trained classification model in the present application are all trained by the following method.

[0075] Acquire training data, clean and enhance the training data, and obtain training samples;

[0076] The fine-tuning network is trained using training samples to obtain the final fine-tuning model.

[0077] Specifically, the training data collected are open source and closed source manually annotated data sets, totaling about 980,000 text data; using real online manual analysis data, the full data sample set is further updated, and the model is repeatedly iterated and fine-tuned.

[0078] In this application, the collected training data needs to be preprocessed, including multi-dimensional data cleaning, removing expressions that are easy to cause model misrecognition, meaningless strings and other combinations; it also includes short text precision cleaning: for the two types of short text data, text and video data, English characters are retained after specific data cleaning respectively; video data analysis: using the author field of the video data to remove the author name contained in the text; text data analysis: separate data preprocessing is performed on the data with Weibo location and HTML tags in the Weibo data.

[0079] In addition, in order to solve the problem that there are many data sources, in addition to the news format standard, Weibo and other social media data may be messy, have typos, and have additional punctuation marks, there are four mainstream academic ways to enhance the data set: randomly swap the order of words, randomly insert punctuation marks, randomly delete words, replace synonyms, and use reverse translation, such as translating the Chinese data set into English first, and then translating from English back to Chinese. Data enhancement in this application can further improve the validation set performance of the overall model.

[0080] In order to solve the problem of limited labeled data and how to use limited labeled data to make high-precision and high-recall information label classification, the above 44 categories of fine-tuning model training set contain a total of 940,000 training data, of which the largest categories are sports and entertainment, containing 110,000 and 90,000 data sets respectively; the smallest categories are religious temples or rural revitalization, containing 40 and 92 data respectively. The reasons for using cost-sensitive learning are as follows:

[0081] Some categories have limited labels, and manual labeling is expensive, but it cannot be missed;

[0082] If some categories are misjudged, the cost is very high: for example, data in the current affairs category is judged as information, which is often complained by customers and seriously affects customer perception. The weighted cross entropy loss function is used for the pre-trained classification model in this application:

[0083] L=w1*y true *log(y pred )+w0*(1-y true )*log(1-y pred )

[0084] Among them, L is the loss function; w1 is the weight of the positive label. The larger the value, the greater the error in label prediction will multiply the loss function; y true is the true label; y pred is the predicted label; w0 is the reverse label weight. The larger the value, the greater the error in label prediction will be, which will multiply the loss function. The sum of w1 and w0 is 1.

[0085] In some optional embodiments, the first fine-tuning model, the second fine-tuning model, the third fine-tuning model and the pre-trained classification model are updated every preset time period.

[0086] In this application, machine learning methods are used to obtain sample data that are "difficult" to classify, and the data is manually confirmed and reviewed again. Then, the manually labeled data is retrained using a supervised learning model or a semi-supervised learning model to gradually improve the model effect and integrate human experience into the machine learning model. Active learning aims to select the most representative samples that are most difficult for machine learning models to determine from the misidentified samples for re-labeling. There are two types of active learning model classifications: the first is streaming active learning, and the second is offline batch learning. The streaming active learning method is combined on the model side, and the offline batch learning strategy is combined during the continuous observation and optimization period after the model goes online.

[0087] Specifically, the fine-tuning model in this application is as follows:

[0088] (I) Roberta pre-trained model,

[0089] The Roberta pre-trained model is a more finely tuned version of the BERT model. The RoBERTa model is an improved version of BERT. Compared with BERT, it has the following improvements in terms of model scale, computing power, and data:

[0090] Larger model parameters and larger training batch size. RoBERTa uses a larger batch size during training. We have tried batch sizes ranging from 256 to 4000.

[0091] More training data, including 160GB of plain text including CC-NEWS. The original BERT was trained using the 16GB BookCorpus dataset and English Wikipedia; in addition, RoBERTa has the following improvements in training methods:

[0092] Remove the next sentence prediction (NSP) task, dynamic masking. BERT relies on random masking and prediction tokens. The original BERT implementation performs masking once during data preprocessing to obtain a static mask. RoBERTa uses dynamic masking: a new masking pattern is generated each time a sequence is input to the model. In this way, as a large amount of data is continuously input, the model will gradually adapt to different masking strategies and learn different language representations. Text encoding. Byte-PairEncoding (BPE) is a mixture of character-level and word-level representations that supports processing many common words in natural language corpora. The original BERT implementation uses a character-level BPE vocabulary of 30K, which is learned after preprocessing the input using heuristic segmentation rules. Instead of taking this approach, Facebook researchers considered training BERT with a larger byte-level BPE vocabulary containing 50K sub-word units without any additional preprocessing or segmentation of the input.

[0093] (II) MacBERT pre-training model

[0094] Leveraging the previous model and proposing a simple modification, MacBERT shares the same pre-training task as BERT but with some modifications. For the MLM task, the following modifications are performed.

[0095] The whole word masked and Ngram masked strategies are used to select candidate tokens for masking, and the ratio of unigram to 4-gram at the word level is 40%, 30%, 20%, and 10%.

[0096] In this application, similar words are obtained by using a synonym toolkit based on word2vec similarity calculation. If an N-gram is selected for masking, similar words will be found separately. In rare cases, when there are no similar words, it is downgraded to use random word replacement. 15% of the input words are masked, of which 40% are replaced with similar words, 10% are replaced with random words, and the remaining 10% retain the original words. For tasks similar to NSP, the sentence order prediction (SOP) task introduced by ALBERT is performed, in which negative samples are created by switching the original order of two consecutive sentences.

[0097] (III) Electra pre-trained model

[0098] The overall structure of the ELECTRA model adopts the discriminator-generator architecture, but strictly speaking it is not a GAN network structure. In the GAN network structure, the gradient of the discriminator can be passed to the generator during model training, while in the ELECTRA model, since the input of the discriminator is discrete data, the gradient cannot be passed from the discriminator to the generator. Electra inputs real text, while the GAN network inputs random noise. The goals of Electra include: the generator learns the language model, and the discriminator is used to determine whether the sequence token is the original token, while the goals of the GAN model include: the generator deceives the discriminator as much as possible, and the discriminator must have the ability to distinguish true from false. In the pre-training stage of Electra, the gradient cannot be passed from the discriminator to the generator, but the GAN network can. When designing the loss of the generator, this application transforms from minimizing the MLM loss to maximizing the discriminator loss. The author makes this transformation based on the following assumption in the article: When the discriminator judges each token in the input sequence, it only relies on the information of the current token and other tokens in the sequence that have not been masked. Therefore, the author concludes that the tokens in the sequence that have not been masked are independent of the generator when the discriminator makes judgments.

[0099] Finally, the sigmoid regression method is used to combine the results of different models, and finally the information sentiment classification results after comprehensive judgment of different models are obtained.

[0100] S103: Determine a detection result of the information data according to the emotion score.

[0101] Specifically, in this application, based on the information sentiment analysis and label classification results, the data will be distributed into the following four intervals:

[0102] In the [-1, -0.60] range, the accuracy of the negative information data is greater than 95%;

[0103] [-0.59, -0.50] interval, suspected negative data, requires manual analysis, accuracy rate 40%+;

[0104] [-0.49,0.40] interval, high probability neutral data, can be quickly judged manually, accounting for 5% of the total data;

[0105] In the [0.41,1] interval, neutral or positive data account for 75% of the total, without the need for manual analysis, and the proportion of negative omissions is less than 0.5%.

[0106] As a sampling method in this application, the query method of uncertain row sampling is to extract sample data that is difficult to distinguish in the model and provide it to manual labeling, so as to achieve the ability to improve the algorithm effect at a faster speed. The key to the uncertainty sampling method is how to describe the uncertainty of samples or data, which usually consists of the following ideas:

[0107] Confidence sampling, for binary and multi-classification models, usually uses softmax for probability prediction. In a binary classification scenario, two data are predicted by the same classifier, and its prediction results for the two data are (0.9, 0.1) (0.51, 0.49) respectively. Therefore, the second data is obviously more difficult to distinguish and therefore has more value to be labeled.

[0108] Marginal sampling refers to selecting sample data that are very easy to be classified into two categories, or the probability of these data being classified into two categories is not much different. Marginal sampling is to select samples with the smallest difference between the model's largest and second largest probabilities. For binary classification, confidence sampling and marginal sampling are equivalent; for multi-classification problems, marginal sampling will select a set of samples for each category and the category that is most easily confused.

[0109] Entropy method, in binary classification and multi-classification scenarios, can select sample data with relatively large entropy as specific data to be labeled. Compared with the above two methods, entropy method takes into account the judgment results of any sample for all categories. Confidence sampling only considers the maximum probability, and edge sampling considers the top2 probability categories.

[0110] Based on the committee's query, in addition to considering the uncertainty sampling method of a single model, we can also consider the scenario of multiple models, which is a method similar to ensemble learning. Through the voting mode of multiple models, we can select those sample data that are more "difficult" to distinguish.

[0111]

[0112] Among them, y i represents all possible labels, V(y i ) represents the label y i The number of votes obtained, C represents the size of the committee.

[0113] Based on the query of the expected model change, the model expected change is the data that causes the largest overall gradient change during training. The specific method is to build a record object, record the ID of each batch, and continuously update the contribution of each batch data to the overall gradient change of the model during training. After the training is completed, the contribution of each batch data to the gradient change is arranged from large to small.

[0114] Based on the query of error reduction, during model training, for each batch, record the data that can best reduce the overall loss; the specific method is to build a record object, record the id of each batch, and continuously update the contribution of each batch data to the overall loss of the model during the training process. After the training is completed, the contribution of each batch data to the loss reduction is arranged from large to small.

[0115] Based on the variance reduction query, the entire data set is measured and the sample data that reduces the overall variance the most is selected.

[0116] Determined by minimizing future variance;

[0117] Minimize the Fisher information ratio, but the disadvantage is that the computational complexity is high.

[0118] In density-weighted queries, sometimes a data point may be an outlier or deviate greatly from the majority of data, making it unsuitable for sample selection or differentiation. Sometimes it is more valuable to consider dense, indistinguishable data. Therefore, when using uncertain sampling or committee queries, the dense rows of sample data can be taken into account.

[0119] like Figure 3 As shown, an embodiment of the present application provides an information detection device, the device comprising:

[0120] An acquisition module 301 is used to acquire information data in response to determining that the information data meets a preset condition;

[0121] A determination module 302 is used to perform information sentiment determination on the information data using a preset database or multiple fine-tuning models to obtain a sentiment label of the information data and a sentiment score corresponding to the sentiment label;

[0122] The detection module 303 is used to determine the detection result of the information data according to the emotion score.

[0123] In some optional embodiments, the acquisition module includes:

[0124] A judging unit, used to judge whether the information data is regional data or organizational data;

[0125] The determination unit is used to determine whether the information data meets a preset condition when the information data is regional data or organizational data.

[0126] In some optional embodiments, the acquisition module further includes:

[0127] The detection unit is used to detect non-regional data or non-organizational data using a preset detection method to obtain the emotion label of the non-regional data or non-organizational data and the emotion score corresponding to the emotion label.

[0128] In some optional embodiments, the acquisition module further includes:

[0129] The cleaning unit is used to clean the information data to obtain cleaned information data.

[0130] In some optional embodiments, the determination module includes:

[0131] A judgment unit, used to judge whether the cleaned information data has been analyzed;

[0132] If it has been determined, the cleaned information data is matched with the labeled data in the preset database to obtain the emotion label of the information data and the emotion score corresponding to the emotion label;

[0133] Otherwise, multiple parallel fine-tuning models are used to perform information sentiment judgment on the information data to obtain sentiment results, and the sentiment results are input into a pre-trained classification model for classification to obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels.

[0134] In some optional embodiments, the determination module further includes:

[0135] The first fine-tuning model performs information sentiment judgment on the information data to obtain a first result;

[0136] The second fine-tuning model performs information sentiment judgment on the information data to obtain a second result;

[0137] The third fine-tuning model performs information sentiment judgment on the information data to obtain a third result;

[0138] The first result, the second result and the third result are integrated to obtain a sentiment result.

[0139] In some optional embodiments, a sigmoid regression function is used to fuse the first result, the second result and the third result to obtain a sentiment result.

[0140] In some optional embodiments, the detection module includes:

[0141] According to the correspondence between the preset threshold range and the information detection result, the information detection result corresponding to the threshold range to which the emotion score belongs is determined as the detection result of the information data.

[0142] In some optional embodiments, the first fine-tuning model, the second fine-tuning model, the third fine-tuning model and the pre-trained classification model are updated every preset time period.

[0143] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device and a readable storage medium.

[0144] Figure 4 A schematic block diagram of an example electronic device 400 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0145] like Figure 4 As shown, the device 400 includes a computing unit 401, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. In the RAM 403, various programs and data required for the operation of the device 400 can also be stored. The computing unit 401, the ROM 402, and the RAM 403 are connected to each other via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0146] A number of components in the device 400 are connected to the I / O interface 405, including: an input unit 406, such as a keyboard, a mouse, etc.; an output unit 407, such as various types of displays, speakers, etc.; a storage unit 408, such as a disk, an optical disk, etc.; and a communication unit 409, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 409 allows the device 400 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0147] The computing unit 401 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 401 performs the various methods and processes described above, such as information detection methods. For example, in some embodiments, the information detection method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of the information detection method described above may be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to perform the information detection method in any other appropriate manner (e.g., by means of firmware).

[0148] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0149] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0150] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0151] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0152] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0153] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0154] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0155] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present disclosure, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0156] The above is only a specific embodiment of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present disclosure, which should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. An information detection method, characterized in that: The method comprises: In response to determining that the information data satisfies a preset condition, acquiring the information data; Using a preset database or multiple fine-tuning models to perform information sentiment determination on the information data, and obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels; The detection result of the information data is determined according to the sentiment score.

2. The method according to claim 1, characterized in that The determining information data satisfies a preset condition, including: Determining whether the information data is regional data or organizational data; When the information data is regional data or organizational data, it is determined that the information data meets a preset condition.

3. The method according to claim 2, characterized in that When the information data is non-regional data or non-organizational data, it also includes: The non-regional data or non-organizational data is detected by using a preset detection method to obtain the emotion labels of the non-regional data or non-organizational data and the emotion scores corresponding to the emotion labels.

4. The method according to claim 3, characterized in that The acquisition of information data includes: The information data is cleaned to obtain cleaned information data.

5. The method according to claim 4, characterized in that The step of using a preset database or a plurality of fine-tuned models to perform information sentiment determination on the information data to obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels includes: Determine whether the cleaned information data has been analyzed; If it has been determined, the cleaned information data is matched with the labeled data in the preset database to obtain the emotion label of the information data and the emotion score corresponding to the emotion label; Otherwise, multiple parallel fine-tuning models are used to perform information sentiment judgment on the information data to obtain sentiment results, and the sentiment results are input into a pre-trained classification model for classification to obtain sentiment labels of the information data and sentiment scores corresponding to the sentiment labels.

6. The method according to claim 5, characterized in that A plurality of parallel fine-tuning models, including: a first fine-tuning model, a second fine-tuning model and a third fine-tuning model; the plurality of parallel fine-tuning models are used to respectively perform information emotion judgment on the information data to obtain an emotion result, including: The first fine-tuning model performs information sentiment judgment on the information data to obtain a first result; The second fine-tuning model performs information sentiment judgment on the information data to obtain a second result; The third fine-tuning model performs information sentiment judgment on the information data to obtain a third result; The first result, the second result and the third result are integrated to obtain a sentiment result.

7. The method according to claim 6, characterized in that The first result, the second result and the third result are fused using a sigmoid regression function to obtain a sentiment result.

8. The method according to claim 1, characterized in that Determining the detection result of the information data according to the emotion score includes: According to the correspondence between the preset threshold range and the information detection result, the information detection result corresponding to the threshold range to which the emotion score belongs is determined as the detection result of the information data.

9. The method according to claim 6, characterized in that The first fine-tuning model, the second fine-tuning model, the third fine-tuning model and the pre-trained classification model are updated every preset time period.

10. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 9.

Citation Information

Patent Citations

  • Text-based emotion detection method and device, computer equipment and medium

    CN112597759A

  • Emotion analysis method, device and equipment and computer readable storage medium

    CN116976354A

  • Multi-label sentiment classification model training method, related device and medium

    CN118133148A