Internet-oriented negative emotion emergency topic detection system and method

Abstract

The invention relates to the field of intelligent detection of negative emotional emergency topics, and particularly discloses an internet-oriented negative emotional emergency topic detection system and method, which are characterized in that text information published by a microblog user, picture data published by the microblog user and video data published by the microblog user are used as input data; and feature extraction and analysis are respectively carried out on the input data through a convolutional neural network model to obtain a classification result. Wherein the convolutional neural network model has excellent performance in the aspect of feature processing of image data, text data and the like. That is, feature vectors with more feature representations can be obtained through the convolutional neural network model. In this way, whether the microblog content to be detected is the negative emotion topic or not can be judged more accurately. Therefore, proper countermeasures can be taken in time, and negative effects caused by the countermeasures can be avoided as much as possible.

Description

Technical Field

[0001] The present application relates to the field of intelligent detection of sudden negative emotion topics, and more specifically, to an Internet-oriented system and method for detecting sudden negative emotion topics. Background Art

[0002] The rapid development of Internet technology has provided a faster way for people to communicate with each other. On the one hand, the amount of information in cyberspace is growing; on the other hand, the spread of information has also accelerated with the rise of various network platforms. Weibo is one of the largest social media platforms in China, with a large user base. Negative emotional topics that occur on Weibo may quickly arouse public attention and negative emotions, and have a negative impact on individuals, organizations or brands. Therefore, detecting negative topics on Weibo can help users understand and respond to possible crises in a timely manner. Moreover, Weibo is a platform with high real-time performance, and negative emotional topics often spread rapidly in a short period of time. Therefore, timely detection of negative emotional topics is very important for early warning and response. However, current detection methods may have a certain time lag and cannot capture the latest negative emotional topics in real time.

[0003] Therefore, we are looking forward to a negative emotional sudden topic detection system and method for the Internet, which can extract and analyze the features of these data through a convolutional neural network model, so as to determine whether the microblog content to be detected is a negative emotional topic, and take appropriate countermeasures in time to avoid its negative impact as much as possible. Summary of the invention

[0004] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides an Internet-oriented negative emotion sudden topic detection system and method, which takes the text information posted by Weibo users, the picture data posted by Weibo users, and the video data posted by Weibo users as input data, and respectively extracts and analyzes the features of these input data through a convolutional neural network model to obtain classification results. Among them, the convolutional neural network model has excellent performance in feature processing of image data, text data, etc. That is, a feature vector with more characteristic representation can be obtained through the convolutional neural network model. In this way, it is possible to more accurately judge whether the Weibo content to be detected is a negative emotion topic. In this way, appropriate countermeasures can be taken in time to avoid the negative impact it brings as much as possible.

[0005] According to a first aspect of the present application, a negative emotion sudden topic detection system for the Internet is provided, which includes:

[0006] The microblog data collection module is used to collect text information posted by microblog users, image data posted by microblog users, and video data posted by microblog users;

[0007] A microblog data feature extraction module, used to obtain feature information of text information posted by the microblog user, feature information of picture data posted by the microblog user, and feature information of video data posted by the microblog user to obtain a microblog text feature vector, a microblog picture feature vector, and a microblog video feature vector;

[0008] The microblog classification result generating module is used to obtain a classification result based on the microblog text feature vector, the microblog image feature vector and the microblog video feature vector.

[0009] According to a second aspect of the present application, a method for detecting sudden negative emotional topics on the Internet is provided, which comprises:

[0010] Collect text information posted by Weibo users, picture data posted by Weibo users, and video data posted by Weibo users;

[0011] Acquire feature information of text information posted by the microblog user, feature information of picture data posted by the microblog user, and feature information of video data posted by the microblog user to obtain a microblog text feature vector, a microblog picture feature vector, and a microblog video feature vector;

[0012] A classification result is obtained based on the microblog text feature vector, the microblog image feature vector and the microblog video feature vector.

[0013] Compared with the prior art, the present application provides an Internet-oriented negative emotion sudden topic detection system and method, which uses text information posted by Weibo users, image data posted by Weibo users, and video data posted by Weibo users as input data, and respectively extracts and analyzes these input data through a convolutional neural network model to obtain classification results. Among them, the convolutional neural network model has excellent performance in feature processing of image data, text data, etc. That is, a feature vector with more characteristic representation can be obtained through the convolutional neural network model. In this way, it is possible to more accurately judge whether the Weibo content to be detected is a negative emotion topic. In this way, appropriate countermeasures can be taken in a timely manner to avoid the negative impact it brings as much as possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0015] Figure 1 The figure illustrates a schematic block diagram of an Internet-oriented negative emotion sudden topic detection system according to an embodiment of the present application.

[0016] Figure 2 The figure illustrates a schematic block diagram of a microblog data feature extraction module in an Internet-oriented negative emotion sudden topic detection system according to an embodiment of the present application.

[0017] Figure 3 The figure illustrates a schematic block diagram of a microblog image feature acquisition unit in a microblog data feature extraction module in an Internet-oriented negative emotion sudden topic detection system according to an embodiment of the present application.

[0018] Figure 4 The figure illustrates a schematic block diagram of a microblog video feature acquisition unit in a microblog data feature extraction module in an Internet-oriented negative emotion sudden topic detection system according to an embodiment of the present application.

[0019] Figure 5 The diagram illustrates a schematic block diagram of a video feature extraction subunit in a microblog video feature acquisition unit in a microblog data feature extraction module in a negative emotion sudden topic detection system for the Internet according to an embodiment of the present application.

[0020] Figure 6 The figure illustrates a schematic block diagram of a microblog classification result generation module in an Internet-oriented negative emotion sudden topic detection system according to an embodiment of the present application.

[0021] Figure 7 The figure illustrates a flow chart of a method for detecting sudden negative emotional topics on the Internet according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0023] Figure 1 FIG. 1 is a schematic block diagram of a negative emotion burst topic detection system for the Internet according to an embodiment of the present application. Figure 1As shown, according to the embodiment of the present application, the Internet-oriented negative emotion sudden topic detection system 100 includes: a Weibo data collection module 110, which is used to collect text information posted by Weibo users, image data posted by Weibo users, and video data posted by Weibo users; a Weibo data feature extraction module 120, which is used to obtain feature information of the text information posted by the Weibo users, feature information of the image data posted by the Weibo users, and feature information of the video data posted by the Weibo users to obtain Weibo text feature vectors, Weibo image feature vectors, and Weibo video feature vectors; a Weibo classification result generation module 130, which is used to obtain classification results based on the Weibo text feature vectors, the Weibo image feature vectors, and the Weibo video feature vectors.

[0024] The Internet is widely used in emotional topics, which has affected people's emotional expression, social interaction and emotional support. There are many online forums and communities for emotional topics on the Internet, such as Weibo, Reddit, Quora and special emotional support forums. These platforms provide a safe environment for people to share their emotional distress, problems and experiences anonymously or publicly, and get support and help from other people's experiences and advice.

[0025] As described in the above background technology, Weibo is one of the largest social media platforms in China, with a huge user base. Negative emotional sudden topics occurring on Weibo may quickly arouse public attention and negative emotions, and have a negative impact on individuals, organizations or brands. Therefore, detecting negative topics on Weibo can help users understand and respond to possible crises in a timely manner. Moreover, Weibo is a platform with high real-time performance, and negative emotional sudden topics often spread rapidly in a short period of time. Therefore, timely detection of negative emotional sudden topics is very important for early warning and response. However, the current detection method may have a certain time lag and cannot capture the latest negative emotional topics in real time. Therefore, a negative emotional sudden topic detection system and method for the Internet is expected, which extracts and analyzes these data through a convolutional neural network model, thereby determining whether the microblog content to be detected is a negative emotional topic, and taking appropriate countermeasures in a timely manner to avoid its negative impact as much as possible.

[0026] In the embodiment of the present application, the microblog data collection module 110 is used to collect text information posted by microblog users, picture data posted by microblog users, and video data posted by microblog users. It should be understood that by collecting text information posted by microblog users, the user's emotional expression on a specific topic or event can be obtained. The text information contains the user's views, comments, emotions, etc., which can help understand the user's attitude and emotional tendency towards a certain topic. In this way, the user's emotional expression can be fully understood. In addition, microblog users not only publish text information, but also share multimedia content such as pictures and videos. These contents are also emotional, and the user's emotional expression can be understood by analyzing the visual elements, expressions, scenes, etc. in the images and videos. That is, collecting text information, picture data, and video data posted by microblog users can provide comprehensive and diverse emotional expression content, help accurate emotional analysis and early warning, and timely response to negative emotional topics. Therefore, in the embodiment of the present application, the text information posted by microblog users, the picture data posted by microblog users, and the video data posted by microblog users are collected as input data.

[0027] Furthermore, when collecting text information posted by Weibo users, considering that Weibo provides an open API interface, the API can be used to obtain text information posted by Weibo users. Specifically, appropriate API interfaces can be selected as needed, such as user Weibo timeline API, user-posted original Weibo API, etc. By calling the API, the user's Weibo content, comments, and forwarding information can be obtained. When collecting image data posted by Weibo users, crawler technology can be used to extract user-posted image data from Weibo web pages. The image data posted by users can be obtained by simulating user login, browsing user homepages, and grabbing images in Weibo. It should be noted that when using crawler technology, it is necessary to comply with the usage specifications and laws and regulations of relevant websites to ensure that data collection is carried out legally and compliantly. In addition, when collecting video data posted by Weibo users, video streaming APIs can be used. These APIs can help process and analyze video data, such as extracting video key frames, identifying objects in videos, emotions, and other information.

[0028] In the embodiment of the present application, the microblog data feature extraction module 120 is used to obtain feature information of text information posted by the microblog user, feature information of picture data posted by the microblog user, and feature information of video data posted by the microblog user to obtain microblog text feature vectors, microblog picture feature vectors, and microblog video feature vectors. It should be understood that after collecting these input data, feature extraction and analysis tasks are further performed on these input data.

[0029] Specifically, Figure 2FIG. 1 is a schematic block diagram of a microblog data feature extraction module in a negative emotion burst topic detection system for the Internet according to an embodiment of the present application. Figure 2 As shown, the microblog data feature extraction module 120 includes: a microblog text feature acquisition unit 121, which is used to extract features of the text information posted by the microblog user through a sequence encoder based on the Clip model to obtain the microblog text feature vector; a microblog image feature acquisition unit 122, which is used to perform convolution encoding on the image data posted by the microblog user to obtain the microblog image feature vector;

[0030] The microblog video feature acquisition unit 123 is used to perform convolution encoding on the video data posted by the microblog user to obtain the microblog video feature vector.

[0031] First, feature extraction and analysis are performed on the text information posted by the Weibo user. It should be understood that Clip (Contrastive Language-Image Pretraining) is a model proposed by OpenAI for jointly training the representation of images and text. The Clip model has been pre-trained on a large-scale data set and can learn rich semantic representation capabilities. Therefore, feature extraction and analysis of Weibo text information is performed using a sequence encoder based on the Clip model. That is, by using a sequence encoder based on the Clip model, Weibo text can be converted into a high-dimensional feature vector, which contains rich semantic information, and can also efficiently extract features from the text, providing faster processing speed and response time. In this way, it helps the system to quickly detect negative emotional topics and take countermeasures in real-time data streams.

[0032] In a specific embodiment of the present application, the microblog text feature acquisition unit 121 is used to: perform word segmentation processing on the text information posted by the microblog user to obtain a word sequence; use the embedding layer of the sequence encoder of the Clip model to map each word in the word sequence to a word embedding vector to obtain a sequence of word embedding vectors; use the converter-based Bert model of the sequence encoder of the Clip model to perform global context semantic encoding on the sequence of embedding vectors to obtain multiple feature vectors; and cascade the multiple feature vectors to obtain the microblog text feature vector.

[0033] Then, feature extraction and analysis are performed on the image data posted by the microblog user. Specifically, Figure 3 The figure shows a schematic block diagram of a microblog image feature acquisition unit in a microblog data feature extraction module in a negative emotion burst topic detection system for the Internet according to an embodiment of the present application. Figure 3As shown, the microblog image feature acquisition unit 122 includes: a picture feature extraction subunit 122-1, which is used to obtain a microblog image feature map by passing the picture data posted by the microblog user through a picture feature extraction module based on a three-dimensional convolutional neural network model; and a picture feature pooling subunit 122-2, which is used to perform a global mean pooling operation based on a feature matrix on the microblog image feature map to obtain the microblog image feature vector.

[0034] It should be understood that the picture information posted by Weibo users can provide rich visual content and contextual information. Pictures can often convey emotions and information intuitively, which is crucial for accurately judging the emotional tendency of Weibo content. By extracting features from pictures, pictures can be converted into numerical feature representations for subsequent sentiment analysis and topic detection. Considering that the picture data posted by Weibo users can be regarded as time series data, because pictures are usually posted in chronological order. The three-dimensional convolutional neural network is a neural network model specifically used to process video and time series data. Based on this, a picture feature extraction module based on a three-dimensional convolutional neural network is used to extract and analyze the picture information of Weibo. In this way, by using a picture feature extraction module based on a three-dimensional convolutional neural network model, the spatiotemporal features in the picture can be effectively captured, and a feature map with rich semantic information can be extracted.

[0035] In a specific embodiment of the present application, the image feature extraction subunit 122-1 is used to: use each layer of the three-dimensional convolutional neural network model to perform the following on the input data in the forward pass of the layer: use the convolution units of each layer of the three-dimensional convolutional neural network model to perform convolution processing based on a three-dimensional convolution kernel on the input data to obtain a convolution feature map; use the pooling units of each layer of the three-dimensional convolutional neural network model to perform pooling processing based on a local feature matrix on the convolution feature map to obtain a pooling feature map; and use the activation units of each layer of the three-dimensional convolutional neural network model to perform nonlinear activation on the feature values ​​of each position in the pooling feature map to obtain an activation feature map; wherein the output of the last layer of the three-dimensional convolutional neural network model is the microblog image feature map.

[0036] Next, considering that the microblog image feature map usually has a high dimension and contains rich local and global information. However, when performing sentiment analysis and topic detection, people often pay more attention to the overall sentiment tendency and topic characteristics rather than local details. Therefore, in order to obtain the global feature representation of the microblog image, the microblog image feature map is further pooled. By performing a global mean pooling operation, the feature map can be reduced in dimension and converted into a feature vector of fixed length.

[0037] In a specific embodiment of the present application, the image feature pooling subunit 122-2 has a specific pooling operation process as follows: 1. Input: The input is a two-dimensional matrix representing a feature map of a Weibo image. The feature map is usually extracted by a convolutional neural network model and has multiple channels and different spatial sizes. 2. Calculate the global mean: For each channel, calculate the average value of the eigenvalues ​​of all positions in the feature map. This can be achieved by summing each channel of the feature matrix and then dividing it by the total number of pixels of the feature matrix. The result is a vector containing the mean of each channel. 3. Feature vector generation: Combine the mean of each channel into a feature vector. The mean of each channel can be arranged in sequence to form a one-dimensional vector. This vector is the feature vector of the Weibo image. 4. Output: The output is a one-dimensional vector representing the feature vector of the Weibo image. The feature vector has a fixed length and contains the global feature information of the Weibo image.

[0038] Through the above steps, the microblog image feature map can be converted into a feature vector of the microblog image. This feature vector can be used for subsequent tasks such as sentiment analysis, topic detection, and multimodal feature fusion.

[0039] Then, feature extraction and analysis are performed on the video data posted by the microblog user. Specifically, Figure 4 The figure shows a schematic block diagram of a microblog video feature acquisition unit in a microblog data feature extraction module in a negative emotion burst topic detection system for the Internet according to an embodiment of the present application. Figure 4 As shown, the Weibo video feature acquisition unit 123 includes: a video feature extraction subunit 123-1, which is used to obtain a Weibo video feature map by passing the video data posted by the Weibo user through a video feature extraction module based on a spatial attention mechanism; a video feature pooling subunit 123-2, which is used to perform a global mean pooling operation based on a feature matrix on the Weibo video feature map to obtain the Weibo video feature vector.

[0040] Considering that a video usually consists of a series of continuous frames, and in sentiment analysis and topic detection, it is not necessary to fully analyze the entire video. By obtaining multiple key frames in the video, a summary of the video can be formed, that is, a set of representative frames that represent the content of the entire video. Therefore, the key frames in the video are first obtained. In this way, the amount of video data processing can be reduced and the efficiency of the system can be improved. In addition, it should be understood that each frame in the video data contains important visual content, but not all frames are equally important for sentiment analysis and topic detection. By introducing the spatial attention mechanism, the video frames can be weighted to highlight the frames that contribute more to sentiment analysis and topic detection. This can improve the efficiency and accuracy of the system and avoid excessive processing of irrelevant frames.

[0041] Specifically, Figure 5 The figure shows a schematic block diagram of a video feature extraction subunit in a microblog video feature acquisition unit in a microblog data feature extraction module in a negative emotion burst topic detection system for the Internet according to an embodiment of the present application. Figure 5 As shown, the video feature extraction subunit 123-1 includes: a key frame acquisition secondary subunit 123-11, which is used to obtain multiple video key frames in the video data posted by the Weibo user; a video key feature acquisition secondary subunit 123-12, which is used to arrange the multiple video key frames into a three-dimensional video input tensor and then obtain a Weibo video feature map through a video feature extraction module based on a spatial attention mechanism.

[0042] Among them, the video key feature acquisition secondary subunit 123-12 is used to: use the convolution coding part of the video feature extraction module to perform deep convolution coding on the video data posted by the Weibo user to obtain an initial convolution feature map; input the initial convolution feature map into the spatial attention part of the video feature extraction module to obtain a spatial attention map; pass the spatial attention map through the Softmax activation function to obtain a spatial attention feature map; calculate the position point multiplication of the spatial attention feature map and the initial convolution feature map to obtain the Weibo video feature map.

[0043] Next, it should be understood that a video feature map is usually a three-dimensional tensor that contains information about the video in both time and space dimensions. However, when performing sentiment analysis and topic detection, it is often necessary to convert video data into a feature vector of fixed length for subsequent classification or judgment. By performing a global mean pooling operation, the spatial dimension of the video feature map can be reduced to a vector of fixed length. Therefore, a pooling operation is further performed on the microblog video feature map.

[0044] In a specific embodiment of the present application, the video feature pooling subunit 123-2 has a specific pooling process as follows: 1. Input the Weibo video feature map, the dimension is (H, W, C), where H represents the height, W represents the width, and C represents the number of channels. 2. Feature extraction: Use a convolutional neural network model to extract features from the Weibo video feature map. This can include multiple convolutional layers and pooling layers to capture spatial and temporal features in the video. 3. Feature matrix: The feature map obtained after feature extraction is represented as a feature matrix with a dimension of (H1, W1, C1), where H1 represents the new height, W1 represents the new width, and C1 represents the new number of channels. 4. Global mean pooling: Perform a global mean pooling operation on the feature matrix, that is, calculate the mean of each channel in the feature matrix. This can be obtained by summing the feature values ​​of each channel and dividing it by the size of the channel dimension. 5. Weibo video feature vector: The mean obtained after global mean pooling is used as the feature vector of the Weibo video. The dimension of the feature vector is C1, where each dimension represents the mean of the corresponding channel.

[0045] Through the above steps, the microblog video feature map can be subjected to a global mean pooling operation based on the feature matrix to obtain a microblog video feature vector. This feature vector can be used to determine whether the microblog content to be detected is a negative emotional topic, and take corresponding countermeasures to avoid its negative impact as much as possible.

[0046] In the embodiment of the present application, the server classification result generating module 130 is used to obtain a classification result based on the microblog text feature vector, the microblog image feature vector and the microblog video feature vector.

[0047] Specifically, Figure 6 FIG. 1 is a schematic block diagram of a microblog classification result generation module in a negative emotion burst topic detection system for the Internet according to an embodiment of the present application. Figure 6 As shown, the server classification result generation module 130 includes: a multi-source information fusion unit 131, which is used to cascade the microblog text feature vector, the microblog image feature vector and the microblog video feature vector to obtain a microblog content classification feature vector; a classification unit 132, which is used to pass the microblog content classification feature vector through a classifier to obtain the classification result. The classification result is used to indicate whether the microblog content to be detected is a negative emotional topic.

[0048] More specifically, in an embodiment of the present application, the multi-source information fusion unit includes: a microblog text-image feature fusion subunit, used to fuse the microblog text feature vector and the microblog image feature vector to obtain a microblog text-image feature vector; a microblog content classification feature generation subunit, used to perform target domain adaptation based on feature space conversion on the microblog text-image feature vector and the microblog video feature vector to obtain the microblog content classification feature vector.

[0049] In an embodiment of the present application, the microblog text-image feature fusion subunit is used to fuse the microblog text feature vector and the microblog image feature vector to obtain a microblog text-image feature vector. It should be understood that microblog content usually contains multiple types of information such as text and images. By fusing text and image feature vectors, different types of features can be comprehensively utilized to more comprehensively represent microblog content. This can improve the accuracy of sentiment classification and topic detection because different types of features can provide information at different levels.

[0050] In an embodiment of the present application, the microblog content classification feature generation subunit is used to perform target domain adaptation based on feature space conversion on the microblog text-image feature vector and the microblog video feature vector to obtain the microblog content classification feature vector. In particular, considering that when the microblog text-image feature vector and the microblog video feature vector are fused to obtain the microblog content classification feature vector, the use of traditional feature matching methods will ignore the more complex semantic associations between features. Traditional methods usually combine features of different modalities based on simple mathematical operations, such as weighted summation or splicing, which can achieve basic feature fusion, but cannot capture the deep interactive relationship and semantic consistency between different modalities. For video data, its content changes over time, and traditional static feature matching methods are difficult to adapt to this dynamic change and cannot effectively capture the continuity and evolution process between video frames. In addition, the impact of local feature distribution on overall expression has not been particularly considered by traditional methods, and in multimodal data analysis, local features (such as a certain area in an image or certain words in a sentence) are crucial to understanding the entire content. In order to solve these problems, in the technical solution of the present application, the microblog text-image feature vector and the microblog video feature vector are adapted to the target domain based on feature space conversion to obtain a microblog content classification feature vector.

[0051] Specifically, in an embodiment of the present application, the microblog text-picture feature vector and the microblog video feature vector are subjected to target domain adaptation based on feature space conversion to obtain a microblog content classification feature vector, including: mapping the microblog text-picture feature vector and the microblog video feature vector to an intrinsic decomposition space to obtain a microblog text-picture intrinsic feature vector and a microblog video intrinsic feature vector; performing inter-domain feature calibration on the microblog text-picture intrinsic feature vector and the microblog video intrinsic feature vector to obtain a microblog text-picture modulated intrinsic feature vector and a microblog video modulated intrinsic feature vector; constructing a microblog content classification fine-grained semantic information field between the microblog text-picture modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector; and fusing the microblog text-picture modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector based on the microblog content classification fine-grained semantic information field to obtain the microblog content classification feature vector.

[0052] More specifically, in an embodiment of the present application, a microblog content classification fine-grained semantic information field is constructed between the microblog text-picture modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector, including: multiplying the transposed vectors of the microblog text-picture modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector, and then dividing by the scale of the microblog video modulated intrinsic feature vector to obtain a microblog content classification fine-grained interaction matrix; and performing convolution encoding on the microblog content classification fine-grained matrix to obtain the microblog content classification fine-grained semantic information field.

[0053] More specifically, in an embodiment of the present application, the microblog text-picture modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector are fused based on the microblog content classification fine-grained semantic information field to obtain the microblog content classification feature vector, including: multiplying the microblog content classification fine-grained semantic information field by the microblog text-picture modulated intrinsic feature vector to obtain the microblog text-picture feature vector after information field mapping; multiplying the microblog content classification fine-grained semantic information field by the microblog video modulated intrinsic feature vector to obtain the microblog video feature vector after information field mapping; and weightedly fusing the microblog text-picture feature vector after information field mapping and the microblog video feature vector after information field mapping to obtain the microblog content classification feature vector.

[0054] In the embodiment of the present application, specifically, the target domain adaptation based on feature space conversion is performed on the microblog text-image feature vector and the microblog video feature vector to obtain the microblog content classification feature vector, including: processing the microblog text-image feature vector and the microblog video feature vector according to the following formula to obtain the microblog content classification feature vector; wherein the formula is:

[0055]

[0056] v c =αv kt +βv st

[0057] Among them, V1 represents the microblog text-image feature vector, V2 represents the microblog video feature vector, PCA(·) represents principal component analysis, U1 represents the sequence of microblog text-image eigendecomposition vectors, Λ1 represents the microblog text-image diagonal matrix, λ 11 and λ 1m denote the eigenvalues ​​of the first and mth positions of the microblog text-image diagonal matrix, respectively. T denotes the transpose of the vector. concat{·;·} denotes vector concatenation. v 11 、v 12 、v 1m Respectively represent the first, second and mth eigenvectors of the sequence of microblog text-image eigendecomposition vectors, V k represents the microblog text-image intrinsic feature vector, U2 represents the sequence of microblog video intrinsic decomposition vectors, Λ2 represents the microblog text-image diagonal matrix, λ 21 and λ 2m They represent the eigenvalues ​​of the first and mth positions of the Weibo video diagonal matrix, respectively, 21 、v 22 、v 2m Respectively represent the first, second and mth eigenvectors of the sequence of microblog video eigendecomposition vectors, V s represents the microblog video intrinsic feature vector, v i represents the i-th eigenvalue of the microblog text-image intrinsic eigenvector, v j represents the jth eigenvalue of the microblog video intrinsic feature vector, represents the square of the vector norm, exp represents the natural exponential function, V′ k represents the intrinsic feature vector after microblog text-image modulation, V′ s represents the intrinsic feature vector after microblog video modulation, represents matrix multiplication, S represents the scale of the vector, M x Represents the fine-grained interaction matrix of microblog content classification, Conv 3×3 (M x ) indicates that M x A convolutional encoding operation based on a convolution kernel of 3×3 is performed, Ω represents the fine-grained semantic information field of microblog content classification, and v kt represents the microblog text-image feature vector after information field mapping, v st represents the feature vector of the microblog video after information field mapping, α and β represent weighted hyperparameters, and vc Represents the microblog content classification feature vector.

[0058] That is, in the technical solution of the present application, the target domain adaptation based on feature space conversion is performed on the microblog text-picture feature vector and the microblog video feature vector. The process first maps the microblog text-picture feature vector and the microblog video feature vector to the intrinsic decomposition space to obtain the microblog text-picture intrinsic feature vector and the microblog video intrinsic feature vector. Those of ordinary skill in the art should know that the process of mapping to the intrinsic decomposition space is a combination of dimensionality reduction and feature extraction. In a specific example, the eigenvalue decomposition technology of principal component analysis (PCA) is used to find the intrinsic decomposition space, and the intrinsic decomposition space mapping is performed. The microblog text-picture intrinsic feature vector and the microblog video intrinsic feature vector after information field mapping not only represent the projection of the original feature vector in the new coordinate system, but also reflect the main direction of change of the data, which helps to avoid the "dimensionality disaster", that is, the problem that the sample density drops sharply as the number of features increases.

[0059] Next, after mapping the feature vector to the intrinsic decomposition space, it is necessary to perform inter-domain feature calibration on the Weibo text-image intrinsic feature vector and the Weibo video intrinsic feature vector to adjust the source domain feature vector to better match the data distribution of the target domain. This operation aims to capture and compensate for the systematic deviation between the two domains to reduce the hidden disturbance of the class expression of the local feature distribution of the Weibo text-image intrinsic feature vector and the Weibo video intrinsic feature vector to the overall class expression. After modulation, the modulated Weibo text-image intrinsic feature vector and the modulated Weibo video intrinsic feature vector can retain the source domain features while being more in line with the actual needs of the target domain, enhancing the model's understanding and application of the target domain data.

[0060] After completing the scale modulation of the feature vector, a microblog content classification fine-grained semantic information field is constructed between the microblog text-picture modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector. The microblog content classification fine-grained semantic information field is used to describe a more detailed relationship between feature vectors. The introduction of the microblog content classification fine-grained semantic information field enables the model to go beyond simple point-to-point comparison and consider more complex patterns and contextual information, thereby enhancing the model's sensitivity to subtle differences and improving the quality of feature representation. Finally, the microblog text-picture modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector are fused based on the microblog content classification fine-grained semantic information field to obtain a microblog content classification feature vector.

[0061] Then, considering that the classifier can quickly process the microblogs to be detected and give the classification results in time. Therefore, in the embodiment of the present application, the microblog content classification feature vector is feature classified by the classifier. This enables the system to take appropriate countermeasures in time to avoid the negative impact of negative emotional topics as much as possible. For example, microblogs involving sensitive topics or negative emotions can be quickly identified and corresponding measures can be taken, such as deleting, restricting dissemination, or providing appropriate guidance and support.

[0062] In a specific embodiment of the present application, the classification unit 133 is used to: use the fully connected layer of the classifier to fully connect the microblog content classification feature vector to obtain a fully connected encoded feature vector; pass the fully connected encoded feature vector through the Softmax classification function of the classifier to obtain a first probability that the microblog content to be detected is a negative emotional topic and a second probability that the microblog content to be detected is not a negative emotional topic; and determine the classification result based on a comparison between the first probability and the second probability.

[0063] It is worth mentioning that in addition to using a classifier to perform feature classification on the microblog content classification feature vector, a rule engine can also be used. A rule engine is a rule-based system in which a series of rules and conditions are defined for judging and classifying input data. In the Internet-oriented negative emotion burst topic detection system, a set of rules can be defined to perform judgment and classification based on the various features and thresholds of the microblog content classification feature vector.

[0064] For example, the following rules can be defined: 1. If a microblog content contains specific negative sentiment words or phrases, it is classified as a negative sentiment topic. 2. If the sentiment score in the microblog content exceeds a certain threshold, it is classified as a negative sentiment topic. 3. If the keywords in the microblog content are related to known negative sentiment topics, it is classified as a negative sentiment topic.

[0065] By defining a set of rules and judging the microblog content classification feature vector according to the rules, the classification result can be obtained. The advantage of the rule engine is that it can be customized according to specific needs and domain knowledge, and classification decisions can be made quickly without a lot of training and tuning processes.

[0066] However, compared with machine learning-based classifiers, rule engines may be more limited by manually defined rules and cannot automatically learn and adapt to new data and situations. Therefore, when designing a system, it is necessary to weigh the use of classifiers or rule engines and choose the most appropriate implementation method based on needs and actual conditions.

[0067] In summary, the method for detecting sudden negative emotional topics on the Internet according to the embodiment of the present application is explained, which takes the text information posted by Weibo users, the picture data posted by Weibo users, and the video data posted by Weibo users as input data, and respectively extracts and analyzes these input data through the convolutional neural network model to obtain classification results. Among them, the convolutional neural network model has excellent performance in feature processing of image data, text data, etc. That is, a feature vector with more characteristic representation can be obtained through the convolutional neural network model. In this way, it is possible to more accurately judge whether the Weibo content to be detected is a negative emotional topic. In this way, appropriate countermeasures can be taken in time to avoid the negative impact as much as possible.

[0068] As described above, the Internet-oriented negative emotion sudden topic detection system 100 according to the embodiment of the present application can be implemented in various wireless terminals, such as a server for Internet-oriented negative emotion sudden topic detection, etc. In one example, the Internet-oriented negative emotion sudden topic detection system 100 according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the Internet-oriented negative emotion sudden topic detection system 100 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the Internet-oriented negative emotion sudden topic detection system 100 can also be one of the many hardware modules of the wireless terminal.

[0069] Alternatively, in another example, the Internet-oriented negative emotion sudden topic detection system 100 and the wireless terminal may also be separate devices, and the Internet-oriented negative emotion sudden topic detection system 100 may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0070] Figure 7 The figure shows a flow chart of a method for detecting sudden negative emotional topics on the Internet according to an embodiment of the present application. Figure 7 As shown, the method for detecting sudden negative emotional topics on the Internet according to the embodiment of the present application includes: S1, collecting text information posted by Weibo users, picture data posted by Weibo users, and video data posted by Weibo users; S2, acquiring feature information of the text information posted by the Weibo users, feature information of the picture data posted by the Weibo users, and feature information of the video data posted by the Weibo users to obtain Weibo text feature vectors, Weibo picture feature vectors, and Weibo video feature vectors; S3, obtaining classification results based on the Weibo text feature vectors, the Weibo picture feature vectors, and the Weibo video feature vectors.

[0071] Here, those skilled in the art can understand that the specific functions and operations of each step in the above-mentioned Internet-oriented negative emotion burst topic detection system have been described in the above reference. Figure 1 The description of the negative emotion burst topic detection method for the Internet has been introduced in detail, and therefore, its repeated description will be omitted.

[0072] Those skilled in the art will appreciate that various aspects of the present application may be illustrated and described by a number of patentable categories or situations, including any new and useful process, machine, product or combination of substances, or any new and useful improvements thereto. Accordingly, various aspects of the present application may be performed entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software may all be referred to as "data blocks", "modules", "engines", "units", "components" or "systems". In addition, various aspects of the present application may be represented as a computer product located in one or more computer-readable media, which includes computer-readable program code.

[0073] The present application uses specific words to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that the mention of "one embodiment" or "an embodiment" or "an alternative embodiment" twice or more in different positions in this specification does not necessarily refer to the same embodiment. In addition, some features, structures or characteristics in one or more embodiments of the present application can be appropriately combined.

[0074] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs. It should also be understood that terms such as those defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology and should not be interpreted in an idealized or highly formal sense, unless explicitly defined as such herein.

[0075] The above is an explanation of the present disclosure and should not be considered as a limitation thereof. Although several exemplary embodiments of the present disclosure are described, it will be readily understood by those skilled in the art that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present disclosure. Therefore, all such modifications are intended to be included within the scope of the present disclosure as defined in the claims. It should be understood that the above is an explanation of the present disclosure and should not be considered to be limited to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present disclosure is defined by the claims and their equivalents.

Claims

1. A negative emotion sudden topic detection system for the Internet, characterized in that: include: The microblog data collection module is used to collect text information posted by microblog users, image data posted by microblog users, and video data posted by microblog users; A microblog data feature extraction module, used to obtain feature information of text information posted by the microblog user, feature information of picture data posted by the microblog user, and feature information of video data posted by the microblog user to obtain a microblog text feature vector, a microblog picture feature vector, and a microblog video feature vector; The microblog classification result generating module is used to obtain a classification result based on the microblog text feature vector, the microblog image feature vector and the microblog video feature vector.

2. The negative emotion sudden topic detection system for the Internet according to claim 1 is characterized in that: The microblog data feature extraction module includes: A microblog text feature acquisition unit, used for extracting features from the text information posted by the microblog user through a sequence encoder based on a Clip model to obtain the microblog text feature vector; A microblog image feature acquisition unit, configured to perform convolution encoding on the image data posted by the microblog user to obtain the microblog image feature vector; The microblog video feature acquisition unit is used to perform convolution encoding on the video data posted by the microblog user to obtain the microblog video feature vector.

3. The Internet-oriented negative emotion sudden topic detection system according to claim 2 is characterized in that: The microblog image feature acquisition unit includes: A picture feature extraction subunit, used for obtaining a microblog picture feature graph by passing the picture data posted by the microblog user through a picture feature extraction module based on a three-dimensional convolutional neural network model; The picture feature pooling subunit is used to perform a global mean pooling operation based on a feature matrix on the microblog picture feature map to obtain the microblog picture feature vector.

4. The Internet-oriented negative emotion sudden topic detection system according to claim 3 is characterized in that: The microblog video feature acquisition unit includes: A video feature extraction subunit, used for obtaining a microblog video feature graph by passing the video data posted by the microblog user through a video feature extraction module based on a spatial attention mechanism; The video feature pooling subunit is used to perform a global mean pooling operation based on a feature matrix on the microblog video feature map to obtain the microblog video feature vector.

5. The Internet-oriented negative emotion sudden topic detection system according to claim 4 is characterized in that: The microblog classification result generating module comprises: A multi-source information fusion unit, used for cascading the microblog text feature vector, the microblog image feature vector and the microblog video feature vector to obtain a microblog content classification feature vector; The classification unit is used to pass the microblog content classification feature vector through a classifier to obtain the classification result.

6. The Internet-oriented negative emotion sudden topic detection system according to claim 5 is characterized in that: The multi-source information fusion unit comprises: A microblog text-image feature fusion subunit, used to fuse the microblog text feature vector and the microblog image feature vector to obtain a microblog text-image feature vector; The microblog content classification feature generation subunit is used to perform target domain adaptation based on feature space conversion on the microblog text-image feature vector and the microblog video feature vector to obtain the microblog content classification feature vector.

7. The Internet-oriented negative emotion sudden topic detection system according to claim 6 is characterized in that: The microblog content classification feature generation subunit is used to: Mapping the microblog text-image feature vector and the microblog video feature vector to an eigendecomposition space to obtain a microblog text-image eigenfeature vector and a microblog video eigenfeature vector; Performing inter-domain feature calibration on the microblog text-picture intrinsic feature vector and the microblog video intrinsic feature vector to obtain a microblog text-picture modulated intrinsic feature vector and a microblog video modulated intrinsic feature vector; Constructing a microblog content classification fine-grained semantic information field between the microblog text-image modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector; The microblog content classification feature vector is obtained by fusing the microblog text-image modulated intrinsic feature vector and the microblog video modulated intrinsic feature vector based on the microblog content classification fine-grained semantic information field.

8. The Internet-oriented negative emotion sudden topic detection system according to claim 7 is characterized in that: The classification result is used to indicate whether the microblog content to be detected is a negative sentiment topic.

9. A method for detecting sudden negative emotional topics on the Internet, characterized in that ,include: Collect text information posted by Weibo users, picture data posted by Weibo users, and video data posted by Weibo users; Acquire feature information of text information posted by the microblog user, feature information of picture data posted by the microblog user, and feature information of video data posted by the microblog user to obtain a microblog text feature vector, a microblog picture feature vector, and a microblog video feature vector; A classification result is obtained based on the microblog text feature vector, the microblog image feature vector and the microblog video feature vector.

10. The method for detecting sudden negative emotional topics on the Internet according to claim 9, characterized in that: Acquiring feature information of text information posted by the microblog user, feature information of picture data posted by the microblog user, and feature information of video data posted by the microblog user to obtain a microblog text feature vector, a microblog picture feature vector, and a microblog video feature vector, including: Extracting features of the text information posted by the microblog user through a sequence encoder based on the Clip model to obtain the microblog text feature vector; Performing convolution encoding on the picture data posted by the microblog user to obtain the microblog picture feature vector; Convolution encoding is performed on the video data posted by the microblog user to obtain the microblog video feature vector.

Images