Internet information monitoring system and method

Abstract

The present application relates to the field of intelligent analysis, and specifically discloses an Internet information monitoring system and method thereof, which processes the user's comment data on hot events to obtain the user's attitude towards the hot event. In this way, it helps managers to quickly determine the information trend of the hot event and formulate corresponding information crisis handling plans in a targeted manner.

Description

Technical Field

[0001] The present application relates to the field of intelligent analysis, and more specifically, to an Internet information monitoring system and method thereof. Background Art

[0002] Internet information monitoring is an information analysis and monitoring solution based on Internet technology. It can collect, organize, analyze and display various information on the Internet to help users understand and grasp the current information trends and promptly discover and solve possible problems.

[0003] Internet information monitoring mainly includes information data collection, information data cleaning, information data storage, sentiment analysis, topic analysis, event tracking, and information report generation, etc. Traditional information monitoring analysis simply determines the emotional tendency of information through keyword matching, which often cannot accurately determine the user's emotional expression.

[0004] Therefore, an optimized Internet information monitoring solution is needed. Summary of the invention

[0005] In order to solve the above technical problems, this application is proposed. The embodiment of this application provides an Internet information monitoring system and method thereof, which processes the user's comment data on hot events to obtain the user's attitude towards the hot event. In this way, it is helpful to help managers quickly judge the information trend of the hot event and formulate corresponding information crisis handling plans in a targeted manner.

[0006] According to one aspect of the present application, there is provided an Internet information monitoring system, which includes:

[0007] A data acquisition module is used to obtain user comment data on hot events;

[0008] A preprocessing module, used for performing data cleaning and preprocessing on the comment data to obtain preprocessed comment data;

[0009] A word segmentation module, used for performing word segmentation processing on the preprocessed comment data to obtain a sequence of comment words;

[0010] A comment word encoding module, used for passing the comment word sequence through a semantic encoder including a word embedding layer to obtain a plurality of comment word context semantic feature vectors;

[0011] A multi-scale convolutional coding module, used for arranging the multiple comment word context semantic feature vectors into a one-dimensional feature vector and then passing it through a multi-scale convolutional structure including a first convolutional layer and a second convolutional layer to obtain a comment data phrase granularity feature vector;

[0012] A cascading module, used for cascading the multiple comment word context semantic feature vectors to obtain a comment data word granularity feature vector;

[0013] A fusion module, used for fusing the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector; and

[0014] The judgment module is used to pass the classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative.

[0015] In the above-mentioned Internet information monitoring system, the comment word encoding module includes: a word vector conversion unit, which is used to use the word embedding layer of the semantic encoder to map each comment word in the sequence of the comment words into multiple comment word vectors to obtain a sequence of comment word vectors; and a semantic encoding unit, which is used to use the semantic encoder including the word embedding layer to perform global context semantic encoding on the sequence of comment word vectors to obtain multiple comment word context semantic feature vectors.

[0016] In the above-mentioned Internet information monitoring system, the multi-scale convolution structure includes: a first convolution layer and a second convolution layer in parallel with each other, and a multi-scale cascade layer connected to the first convolution layer and the second convolution layer, wherein the first convolution layer and the second convolution layer use one-dimensional convolution kernels with different scales.

[0017] In the above-mentioned Internet information monitoring system, the multi-scale convolutional coding module includes: a first scale extraction unit, which is used to use the first convolution layer of the multi-scale convolutional structure to perform one-dimensional convolutional coding on the one-dimensional feature vector using the following first convolutional formula to obtain a first scale phrase granularity feature vector; wherein the first convolutional formula is:

[0018]

[0019] Wherein, a is the width of the first convolution kernel in the x direction, F(a) is the first convolution kernel parameter vector, G(xa) is the local vector matrix operated with the first convolution kernel function, w is the size of the first convolution kernel, X represents the one-dimensional feature vector, Cov1(X) is the first scale phrase granularity feature vector; a second scale extraction unit is used to use the second convolution layer of the multi-scale convolution structure to perform one-dimensional convolution encoding on the one-dimensional feature vector using the following second convolution formula to obtain a second scale phrase granularity feature vector; wherein, the second convolution formula is:

[0020]

[0021] Wherein, b is the width of the second convolution kernel in the x direction, F(b) is the second convolution kernel parameter vector, G(xb) is the local vector matrix operated with the second convolution kernel function, m is the size of the second convolution kernel, X represents the one-dimensional feature vector, Cov2(X) is the second-scale phrase granularity feature vector; and a multi-scale cascade unit, used to use the cascade layer of the multi-scale convolution structure to cascade the first-scale phrase granularity feature vector and the second-scale phrase granularity feature vector to obtain the comment data phrase granularity feature vector.

[0022] In the above-mentioned Internet information monitoring system, the cascade module is used to: cascade the multiple comment word context semantic feature vectors using the following cascade formula to obtain the comment data word granularity feature vector; wherein the cascade formula is:

[0023] V e =Concat[V a , V b ]

[0024] Among them, V a , V b represents the contextual semantic feature vectors of the multiple comment words, Concat[·] represents the cascade function, V e Represents the word granularity feature vector of the comment data.

[0025] In the above-mentioned Internet information monitoring system, the fusion module includes: a sparse coding unit for sparse coding the comment data phrase granularity feature vector and the comment data word granularity feature vector. Sparse Coding to obtain a first sparse feature vector and a second sparse feature vector; a first JS divergence calculation unit, used to calculate the first JS divergence of the first sparse feature vector relative to the second sparse feature vector; a second JS divergence calculation unit, used to calculate the second JS divergence of the second sparse feature vector relative to the first sparse feature vector; a normalization unit, used to normalize the first JS divergence and the second JS divergence to obtain a normalized first JS divergence and a normalized second JS divergence; and a classification feature vector generation unit, used to fuse the first sparse feature vector and the second sparse feature vector with the normalized first JS divergence and the normalized second JS divergence as weights to obtain the classification feature vector.

[0026] In the above-mentioned Internet information monitoring system, the first JS divergence calculation unit is used to: calculate the first JS divergence of the first sparse feature vector relative to the second sparse feature vector using the following first JS divergence formula; wherein the first JS divergence formula is:

[0027]

[0028] Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD1 represents the first JS divergence.

[0029] In the above-mentioned Internet information monitoring system, the second JS divergence calculation unit is used to: calculate the second JS divergence of the second sparse feature vector relative to the first sparse feature vector using the following second JS divergence formula; wherein the second JS divergence formula is:

[0030]

[0031] Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD2 represents the second JS divergence.

[0032] In the above-mentioned Internet information monitoring system, the judgment module includes: a fully connected encoding unit, which is used to use the fully connected layer of the classifier to fully connect the classification feature vector to obtain an encoded classification feature vector; and a classification result generation unit, which is used to pass the encoded classification feature vector through the Softmax classification function of the classifier to obtain the classification result.

[0033] According to another aspect of the present application, there is provided an Internet information monitoring method, which includes:

[0034] Obtain user comment data on hot events;

[0035] Performing data cleaning and preprocessing on the comment data to obtain preprocessed comment data;

[0036] Performing word segmentation processing on the preprocessed comment data to obtain a sequence of comment words;

[0037] Passing the sequence of comment words through a semantic encoder including a word embedding layer to obtain multiple comment word context semantic feature vectors;

[0038] Arranging the multiple comment word context semantic feature vectors into a one-dimensional feature vector and then passing the vector through a multi-scale convolution structure including a first convolution layer and a second convolution layer to obtain a comment data phrase granularity feature vector;

[0039] Cascading the multiple comment word context semantic feature vectors to obtain a comment data word granularity feature vector;

[0040] Merging the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector; and

[0041] The classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative.

[0042] Compared with the prior art, the Internet information monitoring system and method provided by the present application processes the user's comment data on hot events to obtain the user's attitude towards the hot event. This helps managers to quickly determine the information trend of the hot event and formulate corresponding information crisis handling plans in a targeted manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] 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.

[0044] Figure 1 4 is a system block diagram of an Internet information monitoring system according to an embodiment of the present application.

[0045] Figure 2 Schematic diagram of an Internet information monitoring system according to an embodiment of the present application.

[0046] Figure 3 It is a block diagram of a comment word encoding module in an Internet information monitoring system according to an embodiment of the present application.

[0047] Figure 4 This is a block diagram of a multi-scale convolutional coding module in an Internet information monitoring system according to an embodiment of the present application.

[0048] Figure 5 It is a block diagram of the fusion module in the Internet information monitoring system according to an embodiment of the present application.

[0049] Figure 6 It is a block diagram of the judgment module in the Internet information monitoring system according to an embodiment of the present application.

[0050] Figure 7 Flow chart of an Internet information monitoring method according to an embodiment of the present application. DETAILED DESCRIPTION

[0051] 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.

[0052] Application Overview

[0053] As mentioned in the above background technology, traditional Internet information analysis methods generally judge the emotional tendency of information through simple keyword matching, but this method often cannot accurately judge the user's emotional expression. Therefore, an optimized Internet information monitoring solution is needed.

[0054] In recent years, deep learning and neural networks have been widely used in computer vision, natural language processing, text signal processing and other fields. In addition, deep learning and neural networks have also demonstrated a level close to or even beyond that of humans in image classification, object detection, semantic segmentation, text translation and other fields. The development of deep learning and neural networks has provided new solutions and solutions for Internet information monitoring.

[0055] Specifically, we first obtain the user's comment data on hot events. By collecting the user's comment data, we can quickly obtain the public's attention and feedback on the current hot events, so that managers can respond to information changes in a timely manner and adjust corresponding policies and measures according to public opinion. Analyzing these comment data can help managers grasp the mainstream public opinion, understand the public's needs and wishes, and thus better understand the social direction and development trends.

[0056] Next, the comment data is cleaned and preprocessed to obtain preprocessed comment data. It should be understood that there may be some irrelevant or interfering noise data in the comment data, which cannot help understand the user's true thoughts on a topic and may even mislead the information analysis results, so the comment data needs to be cleaned. There are various lexical expressions in the comment data. If they are not standardized, comments with the same point of view may be calculated as different points of view, thereby affecting the accuracy of the information analysis results. By performing sentiment analysis and keyword extraction on the comment data, we can better understand the user's attitude towards the topic. If these data are not processed, these elements are difficult to highlight and discover. By cleaning and preprocessing the comment data, the accuracy of the analysis and the reliability of the results can be improved, thereby more effectively mining useful information.

[0057] Then, the pre-processed comment data is segmented to obtain a sequence of comment words. The word segmentation process can split a sentence into basic text units, which are more conducive to the computer's understanding and analysis of the text content. In many cases, a word in a comment cannot accurately reflect the sentiment of the entire comment. Only by considering the word and the surrounding content can its meaning be truly understood. The use of word segmentation can help retain the contextual information of the comment and reduce excessive reliance on a single word.

[0058] Then, the sequence of comment words is passed through a semantic encoder including a word embedding layer to obtain multiple comment word context semantic feature vectors. Specifically, each comment word can be mapped to a high-dimensional vector space through the word embedding layer, so that similar words will be close in the vector space, which helps to better identify and process complex semantics. Passing the sequence of comment words through a semantic encoder can obtain the context semantic feature vector of each comment word, and further extract the key features of the text data and identify the attributes such as the contained emotions based on the similarity and distance information between the vectors.

[0059] Next, the multiple comment word context semantic feature vectors are arranged into a one-dimensional feature vector and then passed through a multi-scale convolution structure including a first convolution layer and a second convolution layer to obtain a comment data phrase granularity feature vector. Specifically, by merging and rearranging multiple comment word context semantic feature vectors, a complete text data vector can be obtained and input into a multi-scale convolution structure for convolution operation. This structure uses convolution kernels of different sizes to extract substrings of different lengths in the text feature sequence at different levels, further capture the emotional information at the comment phrase level, and thereby construct a comment phrase granularity feature vector. Compared with traditional convolutional networks, the semantic information extracted by multi-scale convolutional structures is richer and more diverse, and can better obtain local information of text data and improve the accuracy of detailed descriptions.

[0060] Then, the multiple comment word context semantic feature vectors are cascaded to obtain a comment data word granularity feature vector. It should be understood that, compared with using multiple comment word context semantic feature vectors separately, cascading can retain the order information of the vocabulary, thereby better reflecting the semantic features of the comment. Cascading multiple comment word context semantic feature vectors can also capture the connection and interaction between different vocabulary, which is conducive to improving the classification accuracy and sentiment analysis effect of the model.

[0061] Next, the comment data phrase-granularity feature vector and the comment data word-granularity feature vector are fused to obtain a classification feature vector. Specifically, the comment data phrase-granularity feature vector captures semantic information more carefully and comprehensively, while the word-granularity feature vector pays more attention to the order and contextual relationship of vocabulary. Fusion of them can make the classification feature vector more expressive. Moreover, compared with using a feature vector alone, the fused classification feature vector exhibits greater robustness and reliability. Finally, the classification feature vector is passed through a classifier to obtain a classification result indicating whether the user's evaluation of the hot event is positive or negative. In this way, we can better grasp public opinion, have a deeper understanding of public opinion trends, and improve managers' predictive capabilities and decision-making levels.

[0062] In particular, in the technical solution of the present application, since the comment data phrase granularity feature vector and the comment data word granularity feature vector are both obtained by encoding the multiple comment word context semantic feature vectors, the two feature vectors will provide partially similar information. If the comment data phrase granularity feature vector and the comment data word granularity feature vector are directly fused by cascading, a certain degree of data redundancy will be caused. In addition, due to errors, noise and uncertainty in data collection and processing, these factors will also cause data noise. It is expected that a fusion coding method can reduce the influence of noise and redundant information to improve the feature fusion effect. In this way, the information loss in the feature fusion process can be effectively reduced, the important information in the original feature distribution can be retained, and the data quality and credibility after feature fusion can be improved.

[0063] Based on this, in the technical solution of the present application, the comment data phrase granularity feature vector and the comment data word granularity feature vector are merged to obtain a classification feature vector, including: performing classification on the comment data phrase granularity feature vector and the comment data word granularity feature vector. Sparse Coding To obtain a first sparse feature vector and a second sparse feature vector; calculate a first JS divergence of the first sparse feature vector relative to the second sparse feature vector; calculate a second JS divergence of the second sparse feature vector relative to the first sparse feature vector; normalize the first JS divergence and the second JS divergence to obtain a normalized first JS divergence and a normalized second JS divergence; use the normalized first JS divergence and the normalized second JS divergence as weights, fuse the first sparse feature vector and the second sparse feature vector to obtain the classification feature vector.

[0064] Specifically, in the technical solution of the present application, the first JS divergence of the first sparse feature vector relative to the second sparse feature vector is calculated using the following first JS divergence formula; wherein the first JS divergence formula is:

[0065]

[0066] Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD1 represents the first JS divergence. .

[0067] Specifically, in the technical solution of the present application, the second JS divergence of the second sparse feature vector relative to the first sparse feature vector is calculated using the following second JS divergence formula; wherein the second JS divergence formula is:

[0068]

[0069] Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD2 represents the second JS divergence.

[0070] The above-mentioned feature distribution fusion algorithm utilizes the idea of ​​sparse coding to effectively capture the structure and pattern information between two feature distributions without being affected by noise and redundant information, thereby improving the feature fusion effect. In this way, it can effectively reduce the information loss in the feature fusion process, retain the important information in the original feature distribution, and improve the data quality and credibility after feature fusion. At the same time, it can also effectively reduce the data dimension after feature fusion, reduce data redundancy and noise, and effectively enhance the data expression ability after feature fusion, extract more implicit information and potential rules, and improve the ability and level of data mining and knowledge discovery, thereby improving the accuracy of the decoding result obtained by the decoder of the decoded feature vector.

[0071] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0072] Exemplary Systems

[0073] Figure 1 : is a system block diagram of an Internet information monitoring system according to an embodiment of the present application. Figure 1As shown, in the Internet information monitoring system 100, it includes: a data acquisition module 110, which is used to obtain user comment data on hot events; a preprocessing module 120, which is used to perform data cleaning and preprocessing on the comment data to obtain preprocessed comment data; a word segmentation module 130, which is used to perform word segmentation processing on the preprocessed comment data to obtain a sequence of comment words; a comment word encoding module 140, which is used to pass the sequence of comment words through a semantic encoder including a word embedding layer to obtain multiple comment word context semantic feature vectors; a multi-scale convolutional encoding module 150, which is used to convert the multiple comment word context semantic features into After being arranged into a one-dimensional feature vector, the comment data are passed through a multi-scale convolution structure including a first convolution layer and a second convolution layer to obtain a comment data phrase granularity feature vector; a cascade module 160 is used to cascade the multiple comment word context semantic feature vectors to obtain a comment data word granularity feature vector; a fusion module 170 is used to fuse the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector; and a judgment module 180 is used to pass the classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative.

[0074] Figure 2 FIG. 1 is an architecture diagram of an Internet information monitoring system according to an embodiment of the present application. Figure 2 As shown, in this architecture, first, the user's comment data on hot events is obtained. Then, the comment data is cleaned and preprocessed to obtain preprocessed comment data. Then, the comment data is cleaned and preprocessed to obtain preprocessed comment data. Then, the sequence of comment words is passed through a semantic encoder including a word embedding layer to obtain multiple comment word context semantic feature vectors. Then, the multiple comment word context semantic feature vectors are arranged into a one-dimensional feature vector and then passed through a multi-scale convolution structure including a first convolution layer and a second convolution layer to obtain a comment data phrase granularity feature vector. Then, the multiple comment word context semantic feature vectors are cascaded to obtain a comment data word granularity feature vector. Next, the comment data phrase granularity feature vector and the comment data word granularity feature vector are fused to obtain a classification feature vector. Finally, the classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative.

[0075] In the Internet information monitoring system 100, the data acquisition module 110 is used to obtain the user's comment data on hot events. It should be understood that by collecting the user's comment data, the public's attention and feedback on the current hot events can be quickly obtained, so that managers can respond to information changes in a timely manner and adjust corresponding policies and measures according to public opinion, and by analyzing these comment data, it can help managers grasp the mainstream public opinion, understand public needs and wishes, and thus better understand the social direction and development trend. In the technical solution of the present application, the user's comment data on hot events can be obtained from channels such as social media, forums, and news websites on the Internet through a crawler program.

[0076] In the Internet information monitoring system 100, the preprocessing module 120 is used to clean and preprocess the comment data to obtain preprocessed comment data. It should be understood that there may be some irrelevant or interfering noise data in the comment data, which cannot help understand the user's true thoughts on a topic and may even mislead the information analysis results, so the comment data needs to be cleaned. There are various lexical expressions in the comment data. If they are not standardized, comments with the same point of view may be calculated as different points of view, thereby affecting the accuracy of the information analysis results. The method of data cleaning and preprocessing of the comment data here mainly includes removing useless information such as emoticons and special characters, removing duplicate data, and tagging the vocabulary with parts of speech.

[0077] In the Internet information monitoring system 100, the word segmentation module 130 is used to perform word segmentation processing on the pre-processed comment data to obtain a sequence of comment words. Word segmentation processing can split a sentence into basic text units, which are more conducive to the computer's understanding and analysis of the text content. In many cases, a word in a comment cannot accurately reflect the emotional tendency of the entire comment. Only by considering this word and the surrounding content can its meaning be truly understood. The use of word segmentation processing can help retain the contextual information of the comment and reduce excessive reliance on a single word.

[0078] In the Internet information monitoring system 100, the comment word encoding module 140 is used to pass the sequence of comment words through a semantic encoder including a word embedding layer to obtain multiple comment word context semantic feature vectors. The semantic encoder including a word embedding layer is a deep learning model that converts a text sequence into a semantic feature vector. This model is usually composed of multiple layers, including an input layer, a word embedding layer, a convolution layer, a pooling layer, and a fully connected layer. Specifically, each comment word can be mapped to a high-dimensional vector space through the word embedding layer, so that similar words will be close in the vector space, which helps to better identify and process complex semantics. Passing the sequence of comment words through a semantic encoder can obtain the context semantic feature vector of each comment word, and based on the similarity and distance information between the vectors, further extract the key features of the text data, identify the attributes such as the contained emotions, etc.

[0079] Figure 3 FIG. 1 is a block diagram of a comment word encoding module in an Internet information monitoring system according to an embodiment of the present application. Figure 3 As shown, the comment word encoding module 140 includes: a word vector conversion unit 141, used to use the word embedding layer of the semantic encoder to map each comment word in the sequence of comment words into multiple comment word vectors to obtain a sequence of comment word vectors; and a semantic encoding unit 142, used to use the semantic encoder including the word embedding layer to perform global context semantic encoding on the sequence of comment word vectors to obtain multiple comment word context semantic feature vectors.

[0080] In the Internet information monitoring system 100, the multi-scale convolutional coding module 150 is used to arrange the multiple comment word context semantic feature vectors into a one-dimensional feature vector and then pass it through a multi-scale convolutional structure including a first convolutional layer and a second convolutional layer to obtain a comment data phrase granularity feature vector. Specifically, by merging and rearranging multiple comment word context semantic feature vectors, a complete text data vector can be obtained and input into the multi-scale convolutional structure for convolution operation. This structure uses convolution kernels of different sizes to extract substrings of different lengths in the text feature sequence at different levels, further capture the emotional information at the comment phrase level, and construct a comment phrase granularity feature vector. Compared with traditional convolutional networks, the semantic information extracted by the multi-scale convolutional structure is richer and more diverse, which can better obtain the local information of the text data and improve the accuracy of the detailed description. Specifically, the multi-scale convolution structure includes: a first convolution layer and a second convolution layer in parallel with each other, and a multi-scale cascade layer connected to the first convolution layer and the second convolution layer, wherein the first convolution layer and the second convolution layer use one-dimensional convolution kernels with different scales.

[0081] Figure 4FIG. 1 is a block diagram of a multi-scale convolutional coding module in an Internet information monitoring system according to an embodiment of the present application. Figure 4 As shown, the multi-scale convolutional coding module 150 includes: a first scale extraction unit 151, which is used to use the first convolution layer of the multi-scale convolutional structure to perform one-dimensional convolution coding on the one-dimensional feature vector using the following first convolution formula to obtain a first-scale phrase granularity feature vector; wherein the first convolution formula is:

[0082]

[0083] Wherein, a is the width of the first convolution kernel in the x direction, F(a) is the first convolution kernel parameter vector, G(xa) is the local vector matrix operated with the first convolution kernel function, w is the size of the first convolution kernel, X represents the one-dimensional feature vector, Cov1(X) is the first scale phrase granularity feature vector; the second scale extraction unit 152 is used to use the second convolution layer of the multi-scale convolution structure to perform one-dimensional convolution encoding on the one-dimensional feature vector using the following second convolution formula to obtain a second scale phrase granularity feature vector; wherein the second convolution formula is:

[0084]

[0085] Wherein, b is the width of the second convolution kernel in the x direction, F(b) is the second convolution kernel parameter vector, G(xb) is the local vector matrix operated with the second convolution kernel function, m is the size of the second convolution kernel, X represents the one-dimensional feature vector, Cov2(X) is the second-scale phrase granularity feature vector, and a multi-scale cascade unit 153 is used to use the cascade layer of the multi-scale convolution structure to cascade the first-scale phrase granularity feature vector and the second-scale phrase granularity feature vector to obtain the comment data phrase granularity feature vector.

[0086] In the Internet information monitoring system 100, the cascade module 160 is used to cascade the multiple comment word context semantic feature vectors to obtain a comment data word granularity feature vector. It should be understood that compared with using multiple comment word context semantic feature vectors separately, cascading can retain the order information of the vocabulary, thereby better reflecting the semantic characteristics of the comment. Cascading multiple comment word context semantic feature vectors can also capture the connection and interaction between different vocabularies, which is conducive to improving the classification accuracy and sentiment analysis effect of the model.

[0087] Specifically, in the Internet information monitoring system 100, the cascade module 160 is used to: cascade the multiple comment word context semantic feature vectors using the following cascade formula to obtain the comment data word granularity feature vector; wherein the cascade formula is:

[0088] V e =Concat[V a , V b ]

[0089] Among them, V a , V b represents the contextual semantic feature vectors of the multiple comment words, Concat[·] represents the cascade function, V e Represents the word granularity feature vector of the comment data.

[0090] In the Internet information monitoring system 100, the fusion module 170 is used to fuse the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector. Specifically, the comment data phrase granularity feature vector captures semantic information more carefully and comprehensively, while the word granularity feature vector pays more attention to the order and contextual relationship of the vocabulary. Fusion of them can make the classification feature vector more expressive. Moreover, compared with using a feature vector alone, the fused classification feature vector shows greater robustness and reliability.

[0091] In particular, in the technical solution of the present application, since the comment data phrase granularity feature vector and the comment data word granularity feature vector are both obtained by encoding the multiple comment word context semantic feature vectors, the two feature vectors will provide partially similar information. If the comment data phrase granularity feature vector and the comment data word granularity feature vector are directly fused by cascading, a certain degree of data redundancy will be caused. In addition, due to errors, noise and uncertainty in data collection and processing, these factors will also cause data noise. It is expected that a fusion coding method can reduce the influence of noise and redundant information to improve the feature fusion effect. In this way, the information loss in the feature fusion process can be effectively reduced, the important information in the original feature distribution can be retained, and the data quality and credibility after feature fusion can be improved.

[0092] Figure 5 FIG. 1 is a block diagram of a fusion module in an Internet information monitoring system according to an embodiment of the present application. Figure 5 As shown, the fusion module 170 includes: a sparse coding unit 171 for performing sparse coding on the comment data phrase granularity feature vector and the comment data word granularity feature vector. Sparse Codingto obtain a first sparse feature vector and a second sparse feature vector; a first JS divergence calculation unit 172, used to calculate the first JS divergence of the first sparse feature vector relative to the second sparse feature vector; a second JS divergence calculation unit 173, used to calculate the second JS divergence of the second sparse feature vector relative to the first sparse feature vector; a normalization unit 174, used to normalize the first JS divergence and the second JS divergence to obtain a normalized first JS divergence and a normalized second JS divergence; and a classification feature vector generation unit 175, used to fuse the first sparse feature vector and the second sparse feature vector with the normalized first JS divergence and the normalized second JS divergence as weights to obtain the classification feature vector.

[0093] Specifically, in the Internet information monitoring system 100, the first JS divergence calculation unit 172 is used to calculate the first JS divergence of the first sparse feature vector relative to the second sparse feature vector using the following first JS divergence formula; wherein the first JS divergence formula is:

[0094]

[0095] Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD1 represents the first JS divergence.

[0096] Specifically, in the Internet information monitoring system 100, the second JS divergence calculation unit 173 is used to calculate the second JS divergence of the second sparse feature vector relative to the first sparse feature vector using the following second JS divergence formula; wherein the second JS divergence formula is:

[0097]

[0098] Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD2 represents the second JS divergence.

[0099] In one embodiment of the present application, the comment data phrase granularity feature vector and the comment data word granularity feature vector are analyzed based on dictionary learning. Sparse CodingTo obtain a first sparse feature vector and a second sparse feature vector, specifically, the comment data phrase granularity feature vector and the comment data word granularity feature vector are combined into a large feature vector for dictionary learning to obtain a dictionary, and then the reconstruction error of the feature vector by the dictionary is minimized, and each feature vector is sparsely represented to obtain the first sparse feature vector and the second sparse feature vector.

[0100] In another embodiment of the present application, the comment data phrase granularity feature vector and the comment data word granularity feature vector are sparsely encoded based on a joint sparse encoder. It should be understood that a joint sparse encoder is a method for jointly encoding multiple related features. It can regard different feature vectors as different input layers and share a sparse encoder on the output layer. Specifically, the input layer of the joint sparse encoder is the comment data phrase granularity feature vector and the comment data word granularity feature vector, and the comment data phrase granularity feature vector and the comment data word granularity feature vector are represented as a matrix of size m×n, where m represents the number of feature vectors and n represents the dimension of the feature vector. The output layer can be a coding vector of size p, where p represents the output dimension of the encoder. Through the joint sparse encoder, each feature vector of multiple groups of related feature vectors can be encoded and parameters can be shared.

[0101] The above-mentioned feature distribution fusion algorithm utilizes the idea of ​​sparse coding to effectively capture the structure and pattern information between two feature distributions without being affected by noise and redundant information, thereby improving the feature fusion effect. In this way, it can effectively reduce the information loss in the feature fusion process, retain the important information in the original feature distribution, and improve the data quality and credibility after feature fusion. At the same time, it can also effectively reduce the data dimension after feature fusion, reduce data redundancy and noise, and effectively enhance the data expression ability after feature fusion, extract more implicit information and potential rules, and improve the ability and level of data mining and knowledge discovery, thereby improving the accuracy of the decoding result obtained by the decoder of the decoded feature vector.

[0102] In the Internet information monitoring system 100, the judgment module 180 is used to pass the classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative. In this way, public opinion can be better grasped, public opinion trends can be more deeply understood, and managers' prediction ability and decision-making level can be improved.

[0103] Figure 6 FIG. 1 is a block diagram of a judgment module in an Internet information monitoring system according to an embodiment of the present application. Figure 6As shown, the judgment module 180 includes: a fully connected encoding unit 181, used to use the fully connected layer of the classifier to fully connect the classification feature vector to obtain an encoded classification feature vector; and a classification result generation unit 182, used to pass the encoded classification feature vector through the Softmax classification function of the classifier to obtain the classification result.

[0104] In summary, the Internet information monitoring system 100 according to the embodiment of the present application is described, which processes the user's comment data on the hot event to obtain the user's attitude towards the hot event. In this way, it helps managers to quickly determine the information trend of the hot event and formulate corresponding information crisis handling plans in a targeted manner.

[0105] As described above, the Internet information monitoring system 100 according to the embodiment of the present application can be implemented in various terminal devices, such as a server for Internet information monitoring. In one example, the Internet information monitoring system 100 according to the embodiment of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the Internet information monitoring system 100 can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the Internet information monitoring system 100 can also be one of the many hardware modules of the terminal device.

[0106] Alternatively, in another example, the Internet information monitoring system 100 and the terminal device may also be separate devices, and the Internet information monitoring system 100 may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0107] Exemplary Methods

[0108] Figure 7 FIG. 1 is a flow chart of an Internet information monitoring method according to an embodiment of the present application. Figure 7As shown, in the Internet information monitoring method, it includes: S110, obtaining user comment data on hot events; S120, obtaining user comment data on hot events; S130, performing word segmentation processing on the pre-processed comment data to obtain a sequence of comment words; S140, passing the sequence of comment words through a semantic encoder including a word embedding layer to obtain multiple comment word context semantic feature vectors; S150, arranging the multiple comment word context semantic feature vectors into a one-dimensional feature vector and then passing them through a multi-scale convolution structure including a first convolution layer and a second convolution layer to obtain a comment data phrase granularity feature vector; S160, cascading the multiple comment word context semantic feature vectors to obtain a comment data word granularity feature vector; S170, fusing the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector; and, S180, passing the classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative.

[0109] In one example, in the above-mentioned Internet information monitoring method, the sequence of comment words is passed through a semantic encoder including a word embedding layer to obtain a plurality of comment word context semantic feature vectors, including: using the word embedding layer of the semantic encoder to map each comment word in the sequence of comment words into a plurality of comment word vectors to obtain a sequence of comment word vectors; and, using the semantic encoder including the word embedding layer to perform global context semantic encoding on the sequence of comment word vectors to obtain a plurality of comment word context semantic feature vectors.

[0110] In one example, in the above-mentioned Internet information monitoring method, the multi-scale convolution structure includes: a first convolution layer and a second convolution layer in parallel with each other, and a multi-scale cascade layer connected to the first convolution layer and the second convolution layer, wherein the first convolution layer and the second convolution layer use one-dimensional convolution kernels with different scales.

[0111] In one example, in the above-mentioned Internet information monitoring method, the step of arranging the multiple comment word context semantic feature vectors into a one-dimensional feature vector and then passing through a multi-scale convolution structure including a first convolution layer and a second convolution layer to obtain a comment data phrase granularity feature vector comprises: using the first convolution layer of the multi-scale convolution structure to perform one-dimensional convolution encoding on the one-dimensional feature vector using the following first convolution formula to obtain a first-scale phrase granularity feature vector; wherein the first convolution formula is:

[0112]

[0113] Wherein, a is the width of the first convolution kernel in the x direction, F(a) is the first convolution kernel parameter vector, G(xa) is the local vector matrix operated with the first convolution kernel function, w is the size of the first convolution kernel, X represents the one-dimensional feature vector, Cov1(X) is the first scale phrase granularity feature vector; the second convolution layer of the multi-scale convolution structure is further used to perform one-dimensional convolution encoding on the one-dimensional feature vector using the following second convolution formula to obtain a second scale phrase granularity feature vector; wherein the second convolution formula is:

[0114]

[0115] Wherein, b is the width of the second convolution kernel in the x direction, F(b) is the second convolution kernel parameter vector, G(xb) is the local vector matrix operated with the second convolution kernel function, m is the size of the second convolution kernel, X represents the one-dimensional feature vector, Cov2(X) is the second-scale phrase granularity feature vector, and the first-scale phrase granularity feature vector and the second-scale phrase granularity feature vector are cascaded using the cascade layer of the multi-scale convolution structure to obtain the comment data phrase granularity feature vector.

[0116] In one example, in the above-mentioned Internet information monitoring method, the step of cascading the multiple comment word context semantic feature vectors to obtain the comment data word granularity feature vector comprises: cascading the multiple comment word context semantic feature vectors to obtain the comment data word granularity feature vector according to the following cascading formula; wherein the cascading formula is:

[0117] V e =Concat[V a , V b ]

[0118] Among them, V a , V b represents the contextual semantic feature vectors of the multiple comment words, Concat[·] represents the cascade function, V e Represents the word granularity feature vector of the comment data.

[0119] In one example, in the above-mentioned Internet information monitoring method, the fusion of the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector includes: Sparse CodingTo obtain a first sparse feature vector and a second sparse feature vector; calculate a first JS divergence of the first sparse feature vector relative to the second sparse feature vector; calculate a second JS divergence of the second sparse feature vector relative to the first sparse feature vector; normalize the first JS divergence and the second JS divergence to obtain a normalized first JS divergence and a normalized second JS divergence; and, using the normalized first JS divergence and the normalized second JS divergence as weights, fuse the first sparse feature vector and the second sparse feature vector to obtain the classification feature vector.

[0120] In one example, in the above-mentioned Internet information monitoring method, the calculating the first JS divergence of the first sparse feature vector relative to the second sparse feature vector includes: calculating the first JS divergence of the first sparse feature vector relative to the second sparse feature vector using the following first JS divergence formula; wherein the first JS divergence formula is:

[0121]

[0122] Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD1 represents the first JS divergence.

[0123] In one example, in the above-mentioned Internet information monitoring method, the calculating the second JS divergence of the second sparse feature vector relative to the first sparse feature vector includes: calculating the second JS divergence of the second sparse feature vector relative to the first sparse feature vector using the following second JS divergence formula; wherein the second JS divergence formula is:

[0124]

[0125] Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD2 represents the second JS divergence.

[0126] In one example, in the above-mentioned Internet information monitoring method, the classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative, including: using the fully connected layer of the classifier to fully connect encode the classification feature vector to obtain an encoded classification feature vector; and, passing the encoded classification feature vector through the Softmax classification function of the classifier to obtain the classification result.

[0127] In summary, the Internet information monitoring method based on the embodiment of the present application is explained, which processes the user's comment data on hot events to obtain the user's attitude towards the hot event. In this way, it helps managers to quickly determine the information trend of the hot event and formulate corresponding information crisis handling plans in a targeted manner.

Claims

1. An Internet information monitoring system, characterized in that: include: A data acquisition module is used to obtain user comment data on hot events; A preprocessing module, used for performing data cleaning and preprocessing on the comment data to obtain preprocessed comment data; A word segmentation module, used for performing word segmentation processing on the preprocessed comment data to obtain a sequence of comment words; A comment word encoding module, used for passing the comment word sequence through a semantic encoder including a word embedding layer to obtain a plurality of comment word context semantic feature vectors; A multi-scale convolutional coding module, used for arranging the multiple comment word context semantic feature vectors into a one-dimensional feature vector and then passing it through a multi-scale convolutional structure including a first convolutional layer and a second convolutional layer to obtain a comment data phrase granularity feature vector; A cascading module, used for cascading the multiple comment word context semantic feature vectors to obtain a comment data word granularity feature vector; A fusion module, used for fusing the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector; as well as A judgment module, used for passing the classification feature vector through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative; The multi-scale convolution structure includes: a first convolution layer and a second convolution layer in parallel with each other, and a multi-scale cascade layer connected to the first convolution layer and the second convolution layer, wherein the first convolution layer and the second convolution layer use one-dimensional convolution kernels with different scales; Wherein, the multi-scale convolutional coding module includes: A first scale extraction unit, configured to use the first convolution layer of the multi-scale convolution structure to perform one-dimensional convolution encoding on the one-dimensional feature vector using the following first convolution formula to obtain a first-scale phrase granularity feature vector; Among them, the first convolution formula is: Wherein, a is the width of the first convolution kernel in the x direction, F(a) is the first convolution kernel parameter vector, G(xa) is the local vector matrix operated with the first convolution kernel function, w is the size of the first convolution kernel, X represents the one-dimensional feature vector, Cov1(X) is the first scale phrase granularity feature vector; A second scale extraction unit, configured to use the second convolution layer of the multi-scale convolution structure to perform one-dimensional convolution encoding on the one-dimensional feature vector using the following second convolution formula to obtain a second-scale phrase granularity feature vector; Among them, the second convolution formula is: Wherein, b is the width of the second convolution kernel in the x direction, F(b) is the second convolution kernel parameter vector, G(xb) is the local vector matrix operated with the second convolution kernel function, m is the size of the second convolution kernel, X represents the one-dimensional feature vector, Cov2(X) is the second scale phrase granularity feature vector; A multi-scale cascade unit, configured to cascade the first-scale phrase granularity feature vector and the second-scale phrase granularity feature vector using the cascade layer of the multi-scale convolution structure to obtain the comment data phrase granularity feature vector; Wherein, the fusion module includes: A sparse coding unit, configured to perform sparse coding on the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a first sparse feature vector and a second sparse feature vector; A first JS divergence calculation unit, configured to calculate a first JS divergence of the first sparse feature vector relative to the second sparse feature vector; A second JS divergence calculation unit, used to calculate a second JS divergence of the second sparse feature vector relative to the first sparse feature vector; a normalization unit, configured to perform normalization processing on the first JS divergence and the second JS divergence to obtain a normalized first JS divergence and a normalized second JS divergence; and A classification feature vector generating unit is used to fuse the first sparse feature vector and the second sparse feature vector by taking the normalized first JS divergence and the normalized second JS divergence as weights to obtain the classification feature vector.

2. The Internet information monitoring system according to claim 1, characterized in that: The comment word encoding module includes: a word vector conversion unit, configured to map each comment word in the sequence of comment words into a plurality of comment word vectors using the word embedding layer of the semantic encoder to obtain a sequence of comment word vectors; and A semantic encoding unit is used to use the semantic encoder including the word embedding layer to perform global context semantic encoding on the sequence of comment word vectors to obtain multiple comment word context semantic feature vectors.

3. The Internet information monitoring system according to claim 2, characterized in that: The cascade module is used to: cascade the multiple comment word context semantic feature vectors using the following cascade formula to obtain the comment data word granularity feature vector; Wherein, the cascade formula is: In e =Concat[V a ,In b ] Among them, V a , V b represents the contextual semantic feature vectors of the multiple comment words, Concat[·] represents the cascade function, V e Represents the word granularity feature vector of the comment data.

4. The Internet information monitoring system according to claim 3, characterized in that: The first JS divergence calculation unit is used to: calculate the first JS divergence of the first sparse feature vector relative to the second sparse feature vector using the following first JS divergence formula; Among them, the first JS divergence formula is: Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD1 represents the first JS divergence.

5. The Internet information monitoring system according to claim 4, characterized in that: The second JS divergence calculation unit is used to: calculate the second JS divergence of the second sparse feature vector relative to the first sparse feature vector using the following second JS divergence formula; Among them, the second JS divergence formula is: Wherein, S1 is the first sparse feature vector, S2 is the second sparse feature vector, S is the average distribution of the first sparse feature vector and the second sparse feature vector, KL represents KL divergence, and JSD2 represents the second JS divergence.

6. The Internet information monitoring system according to claim 5, characterized in that: The judging module comprises: a fully connected encoding unit, configured to perform fully connected encoding on the classification feature vector using the fully connected layer of the classifier to obtain an encoded classification feature vector; and The classification result generating unit is used to pass the encoded classification feature vector through the Softmax classification function of the classifier to obtain the classification result.

7. A method for monitoring Internet information, characterized in that: include: Obtain user comment data on hot events; Performing data cleaning and preprocessing on the comment data to obtain preprocessed comment data; Performing word segmentation processing on the preprocessed comment data to obtain a sequence of comment words; Passing the sequence of comment words through a semantic encoder including a word embedding layer to obtain multiple comment word context semantic feature vectors; Arranging the multiple comment word context semantic feature vectors into a one-dimensional feature vector and then passing the vector through a multi-scale convolution structure including a first convolution layer and a second convolution layer to obtain a comment data phrase granularity feature vector; Cascading the multiple comment word context semantic feature vectors to obtain a comment data word granularity feature vector; Merging the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector; as well as Passing the classification feature vector through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the user's evaluation of the hot event is positive or negative; The multi-scale convolution structure includes: a first convolution layer and a second convolution layer in parallel with each other, and a multi-scale cascade layer connected to the first convolution layer and the second convolution layer, wherein the first convolution layer and the second convolution layer use one-dimensional convolution kernels with different scales; The plurality of comment word context semantic feature vectors are arranged into a one-dimensional feature vector and then passed through a multi-scale convolution structure including a first convolution layer and a second convolution layer to obtain a comment data phrase granularity feature vector, including: Using the first convolution layer of the multi-scale convolution structure to perform one-dimensional convolution encoding on the one-dimensional feature vector using the following first convolution formula to obtain a first-scale phrase granularity feature vector; Among them, the first convolution formula is: Wherein, a is the width of the first convolution kernel in the x direction, F(a) is the first convolution kernel parameter vector, G(xa) is the local vector matrix operated with the first convolution kernel function, w is the size of the first convolution kernel, X represents the one-dimensional feature vector, Cov1(X) is the first scale phrase granularity feature vector; Using the second convolution layer of the multi-scale convolution structure to perform one-dimensional convolution encoding on the one-dimensional feature vector using the following second convolution formula to obtain a second-scale phrase granularity feature vector; Among them, the second convolution formula is: Wherein, b is the width of the second convolution kernel in the x direction, F(b) is the second convolution kernel parameter vector, G(xb) is the local vector matrix operated with the second convolution kernel function, m is the size of the second convolution kernel, X represents the one-dimensional feature vector, Cov2(X) is the second scale phrase granularity feature vector; Using the cascade layer of the multi-scale convolution structure to cascade the first-scale phrase granularity feature vector and the second-scale phrase granularity feature vector to obtain the comment data phrase granularity feature vector; The fusion of the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a classification feature vector includes: Sparsely encoding the comment data phrase granularity feature vector and the comment data word granularity feature vector to obtain a first sparse feature vector and a second sparse feature vector; Calculating a first JS divergence of the first sparse feature vector relative to the second sparse feature vector; Calculating a second JS divergence of the second sparse feature vector relative to the first sparse feature vector; Normalizing the first JS divergence and the second JS divergence to obtain a normalized first JS divergence and a normalized second JS divergence; and The first sparse feature vector and the second sparse feature vector are fused with the normalized first JS divergence and the normalized second JS divergence as weights to obtain the classification feature vector.

Images