Multi-source information analysis method and system based on deep learning

Abstract

The invention discloses a multi-source information analysis method based on deep learning, and belongs to the field of network information analysis. According to the method, firstly, multi-document abstract extraction is conducted, sentiment classification is conducted on input documents, the quality and information richness of document abstracts are improved, then information updating of multi-granularity nodes is achieved through a graph attention mechanism, and therefore the problem that a cross-document relation is difficult to model in the prior art is solved. Furthermore, the information popularity in a specific time period is deeply analyzed by combining the popularity analysis of the network main information questions with two aspects of emotional polarity and information popularity, so that the emotional evolution process of the network information can be known, and a basis is provided for scientific decision and active management.

Description

Technical Field

[0001] The present invention belongs to the field of network information analysis, and specifically relates to a multi-source information analysis method and system based on deep learning. Background Art

[0002] Sentiment analysis, also known as opinion mining, is a major branch of natural language processing. It processes and analyzes text with emotional connotations to determine its emotional polarity. Unlike text-level sentiment analysis, aspect-level sentiment analysis shifts the target of expressing emotions from the entire document or sentence to an entity or an aspect of an entity. For example, when a customer buys clothes, they need to understand multiple aspects such as the price, quality, and after-sales service of the clothes. Therefore, they need to conduct sentiment analysis on these aspects to meet the customer's needs.

[0003] In today's digital age, the amount of online information is surging, and transmission speeds are rapidly increasing. In this context, network information analysis is crucial. By collecting and analyzing information from multiple channels, such as social media, news reports, and forum comments, it helps businesses, governments, and organizations understand public sentiment and social dynamics in real time, and promptly identify potential risks and opportunities. Due to the vast amount of online information, current network analysis methods suffer from high redundancy in the analysis process and difficulty capturing information hotspots, resulting in poor results. On the other hand, traditional sentiment analysis primarily uses text data for sentiment classification and recognition. Multimodal sentiment analysis, on the other hand, combines multiple modal data (such as text, images, and audio) to perform sentiment analysis. However, many techniques generally only consider fine-grained text, namely word-level relationships, focusing primarily on local features rather than fully considering the global nature of the text, resulting in poor accuracy. Summary of the Invention

[0004] Purpose of the invention: The present invention aims to provide a multi-source information analysis method and system based on deep learning, which can effectively extract summaries, analyze hot information trends, provide a basis for scientific decision-making and proactive management, avoid blind response, and seize development opportunities.

[0005] Technical solution: A multi-source information analysis method based on deep learning, including the following steps:

[0006] (1) Multi-document summary extraction

[0007] First, the sentiment classification of the input text information data is performed based on media data and AI large models. Then, the sentiment classification of the input document is further performed based on the constructed multi-document summary model to improve the quality and information richness of the document summary. At the same time, the model uses the graph attention mechanism to realize the information update of multi-granular nodes and complete label prediction to reduce the redundancy of the summary.

[0008] (2) Analysis of the popularity of online information topics

[0009] Combining sentiment polarity and information heat, we analyze the information heat within a specific time period to obtain the sentiment evolution process of network information; sentiment polarity analysis includes sentiment classification of text content under each topic based on the AI ​​big model to determine its sentiment tendency, including counting the ratio of positive and negative sentiment tendencies in a certain time period to reflect the sentiment polarity of information data under each topic within that time period; information heat analysis is a comprehensive evaluation based on indicators including the number of posts, comments, and reposts of the information.

[0010] Furthermore, the multi-document summarization model includes an input layer, a feature extraction layer, a graph attention layer, and a sentence selection layer;

[0011] Input layer: Extracts summary sentences from the input text information and represents it as an abstract task of predicting sentence label sequence {l1,l2,…l n}, where l n =m means that the nth sentence in the summary is extracted from the mth sentence in the document, and the input text information is segmented into sentences, and then the review topics are extracted through the LDA topic model;

[0012] Feature extraction layer: The local features of each sentence and topic are captured through a convolutional neural network, and then the global features are obtained using a bidirectional long short-term memory network. Finally, the concatenation of local and global features is used as node features as the input of the graph attention layer;

[0013] Graph Attention Layer: This layer constructs a heterogeneous graph containing topic, sentence, and document nodes. For sentence nodes, it uses a convolutional neural network to extract the n-gram feature vectors of each sentence in the text information and stores the extracted features. It then encodes the sentence node position as a vector feature. The two vectors are concatenated and mapped to a specified size to form the local features of the sentence. The local features and sentence length values ​​are then input into a Bi-LSTM model for further feature extraction. Finally, the two features are concatenated as the global features of the sentence node to improve the expressiveness of sentence features.

[0014] Sentence selection layer: Use the maximum margin correlation algorithm to select the best summary sentence and finally complete the prediction of the candidate summary label.

[0015] Furthermore, in the graph attention layer, the input multi-document representation is D = {D1, D2, ... D nd}, where D i For the i-th input document, the topic in the document is extracted using the LDA model, which is recorded as the topic node T = {T1, T2, ... T nt};

[0016] The concatenation of the global features of the sentence nodes is represented as:

[0017]

[0018] in, represents the global features of the jth sentence, l j Represents the local n-gram features of the jth sentence, g j Represents the features obtained by the Bi-LSTM layer;

[0019] The document node is regarded as a special sentence node, and the average value of all sentence features contained in it is used as the feature representation of the document node. Then, the heterogeneous graph is constructed using the sentences and the extracted topic nodes, which is expressed as follows:

[0020] G=(N,E)

[0021] Where N is a node set including sentence and topic nodes, E represents the weight matrix of undirected edges between nodes, and the frequencies of all topic words appearing in sentences and documents are converted into TF-IDF values ​​as the initial edge weight values ​​between corresponding nodes. Document nodes use TF-IDF values ​​as the edge weights connecting to topic nodes.

[0022] Furthermore, the information of sentence nodes is updated in the heterogeneous graph. For a sentence node, based on its neighboring topic nodes and their weight values, the weight value is normalized to obtain the attention coefficient between the sentence node and the adjacent nodes. According to the attention coefficient, the sentence nodes are weightedly aggregated to obtain the new features of the sentence under the current attention head. The features under all attention heads are spliced ​​to obtain the features of the final sentence node and the heterogeneous graph is updated.

[0023] Furthermore, in the sentence selection layer, considering that there is relevant information in multiple input documents and there is a certain amount of character overlap between them, the maximum margin relevance algorithm selects the summary sentence based on the combination of the relevance of the sentence to the entire document and the novelty of the information contained in the sentence itself. The expression is as follows:

[0024]

[0025] Where R is the set of candidate sentences, S is the set of selected sentences, D is the document set, S i Indicates that the i-th sentence is not selected, S j Indicates that the jth sentence in the sentence set is selected, R\S is the set of unselected sentences, Sim1 represents the similarity between the i-th sentence and the document, and Sim2 represents the similarity between the candidate sentence and the selected sentence.

[0026] Furthermore, this method is used for a certain time period Ti , the information heat is calculated as follows:

[0027]

[0028] Among them, S heat (T i ) represents the time period T i The popularity of nz is T in the time period i The number of topics, b i Indicates the time period T i The number of posts, Represents the time period T i The average number of reposts, comments, and likes for related articles within the website;

[0029] Combining sentiment polarity and information heat, the calculation formula for the comprehensive score of network information sentiment intensity is as follows:

[0030]

[0031] The changes in the emotional intensity of online information with time series characteristics can show the changes in the emotional intensity of online information from its generation to its demise. Based on the changes in emotional intensity, we can analyze and understand the public's emotional attitude towards the information and realize the tracking and analysis of the information.

[0032] The present invention also provides a multi-source information sentiment analysis system based on deep learning, which uses the above method to perform sentiment analysis on text data, including:

[0033] The multi-document summary extraction module improves the quality and information richness of document summaries by performing sentiment classification on input documents. It also uses the graph attention mechanism to update information at multi-granularity nodes and complete label prediction to reduce summary redundancy.

[0034] The network information topic heat analysis module combines sentiment polarity and information heat to analyze the information heat within a certain time period. It is used to understand the emotional evolution process of network information, including social public opinion orientation analysis, and provide a basis for scientific decision-making and proactive management.

[0035] The beneficial effect is that the multi-source information analysis method based on deep learning described in the present invention first performs multi-document summary extraction, and improves the quality and information richness of the document summary by performing sentiment classification on the input document. At the same time, the graph attention mechanism is used to realize the information update of multi-granularity nodes, which solves the problem that the existing model is difficult to model cross-document relationships; the network main information topic heat analysis combines the two aspects of sentiment polarity and information heat, and conducts an in-depth analysis of the information heat within a specific time period, which can understand the emotional evolution process of network information and provide a basis for scientific decision-making and proactive management. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is an implementation flow chart of the method of the present invention. DETAILED DESCRIPTION

[0037] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0038] like Figure 1 As shown, the multi-source information analysis method based on deep learning of the present invention includes the following steps:

[0039] Multi-document summary extraction: A multi-document summary model is built to extract summaries. The model is divided into four layers: input layer, feature extraction layer, graph attention layer, and sentence selection layer. The model constructs a heterogeneous graph by selecting topic nodes, sentence nodes, and document nodes, and uses the graph attention mechanism to update nodes of different granularities, ultimately generating a summary with increased coverage and reduced redundancy.

[0040] Before input, a large model is used to perform sentiment classification on the input text information data. The information sources include multiple platforms on the Internet, and data of the same sentiment category is used as multiple documents as input. The value of "text" in the dataset can be set to a single document or multiple documents. All English strings are lowercase and tokenized, and word segmentation is performed using the Jieba tool. In the input layer, for different groups of input text information, the abstract sentence is extracted and abstracted into the task of predicting the sentence label sequence {l1,l2,…l n}. n =m means that the nth sentence in the summary is extracted from the mth sentence in the document. The input text is segmented into sentences, and then the review topics are extracted using the LDA topic model. In the feature extraction layer, a convolutional neural network is used to capture the local features of each sentence and topic. A bidirectional long short-term memory network is then used to obtain global features. Finally, the concatenation of local and global features is used as node features as input to the graph attention layer.

[0041] In the graph attention layer, a heterogeneous graph containing topics, sentences, and document nodes is first established. Specifically, in the input of multiple documents D = {D1, D2, ... D nd}, where D i For the i-th input document, the topic in the document is extracted using the LDA model, which is recorded as the topic node T = {T1, T2, ... T ntFor sentence nodes, a convolutional neural network is used to extract the n-gram feature vector of each sentence in the text information, and the extracted features are stored. The sentence node position is then encoded as a vector feature. The two vectors are concatenated and mapped to a specified size to form the local feature of the sentence. The local feature and sentence length value are then input into the Bi-LSTM model for further feature extraction. Finally, the two features are concatenated as the global feature of the sentence node to improve the sentence feature expression capability. The specific formula is as follows:

[0042]

[0043] in, represents the global features of the jth sentence, l j Represents the local n-gram features of the jth sentence, g j Represents the features obtained by the Bi-LSTM layer. The document node can be regarded as a special sentence node, and the average value of all sentence features contained in it is used as the feature representation of the document node. Finally, the heterogeneous graph is constructed using the sentences and the extracted topic nodes, which is expressed as the following formula

[0044] G=(N,E)

[0045] Where N is a node set including sentence and topic nodes, E represents the weight matrix of undirected edges between nodes, and the frequencies of all topic words appearing in sentences and documents are converted into TF-IDF values ​​as the initial edge weight values ​​between corresponding nodes. Document nodes use TF-IDF values ​​as the edge weights connecting to topic nodes.

[0046] Then, the information of the sentence nodes is updated in the heterogeneous graph. For a sentence node, based on its neighboring topic nodes and their weight values, the weight value is normalized to obtain the attention coefficient between the sentence node and the adjacent nodes. According to the attention coefficient, the sentence nodes are weightedly aggregated to obtain the new features of the sentence under the current attention head. The features under all attention heads are spliced ​​to obtain the features of the final sentence node and the heterogeneous graph is updated.

[0047] Finally, in the sentence selection layer, the maximum margin relevance algorithm is used to select the best summary sentence to complete the prediction of the candidate summary tag. Because there is related information in multiple input documents and there is a certain amount of character overlap between them, selecting sentences based solely on relevance indicators may result in a high degree of redundancy in the generated summary. The maximum margin relevance algorithm selects summary sentences based on a combination of the relevance of the sentence to the entire document and the novelty of the information contained in the sentence itself. This ensures that the core content is retained while filtering out sentences with a large amount of character overlap with the selected sentences. The formula is as follows:

[0048]

[0049] Where R is the set of candidate sentences, S is the set of selected sentences, D is the document set, S i Indicates that the i-th sentence is not selected, S j Indicates that the jth sentence in the sentence set is selected, R\S is the set of unselected sentences, Sim1 represents the similarity between the i-th sentence and the document, and Sim2 represents the similarity between the candidate sentence and the selected sentence.

[0050] Analysis of the popularity of online information topics: Based on the summary obtained from the heterogeneous graph and combined with network information, the LDA model is used to obtain network topic information. The factors affecting the intensity of network topic information hotness are divided into two parts: the popularity of the information itself and the sentiment polarity of the comments related to the information. For the sentiment polarity analysis of information comments, the large model is first used to perform sentiment classification on the text content under each topic to determine its emotional tendency. The emotional polarity of a single data point is divided into two categories: positive is recorded as 1 and negative is recorded as 0. Then, the ratio of positive and negative sentiment tendencies within a certain time period is calculated to reflect the emotional polarity of the information data under each topic during that time period. The formula is as follows:

[0051]

[0052] Text(t i ) is the time period T i The sentiment tendency of the i-th sample in the time period T i The number of samples in Text(T i ) is the time period T i The sentiment tendency statistics within the group are calculated and the sentiment tendency ratio is normalized. The formula is as follows:

[0053]

[0054] max(Text(T i )) is the maximum value of sentiment tendency statistics in each topic, min(Text(T i )) is the minimum value of sentiment tendency statistics in each topic.

[0055] Regarding the analysis of the popularity of online information, since the popularity of online information is positively correlated with its related information, as the number of specific related information increases and the frequency increases within a unit time, the popularity of the information also increases. The popularity of information reflects its development trend. High popularity means that the information is attracting widespread attention, so it needs to be monitored and analyzed in a timely and accurate manner. On the Internet, users express their attention to information through forwarding, liking, and commenting. The interactive behavior between different users continuously expands the scope of influence of information, prompting the relevant information to spread on the Internet and even evolve into hot information events. Online information is composed of the number of online posts, likes, comments, and forwarding. A rapid increase in the number of relevant posts, likes, comments, and forwarding within a unit time indicates that the popularity of the event has increased, and the number of people paying attention to the event is also increasing. Therefore, the evaluation of the popularity of online information is mainly based on quantitative indicators such as the number of posts, comments, and forwarding of related information. For a certain time period T i , the information heat calculation is as follows:

[0056]

[0057] Among them, S heat (T i ) represents the time period T i The popularity of nz is T in the time period i The number of topics, b i Indicates the time period T i The number of posts, Represents the time period T i The average number of reposts, comments, and likes for related information articles.

[0058] Combining sentiment polarity and information heat, the calculation formula for the comprehensive score of network information sentiment intensity is as follows:

[0059]

[0060] The changes in the emotional intensity of online information with time series characteristics can show the changes in the emotional intensity of online information from its generation to its demise. Based on the changes in emotional intensity, we can analyze and understand the public's emotional attitude towards the information and realize the tracking and analysis of the information.

[0061] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A multi-source information analysis method based on deep learning, characterized in that: The following steps are involved: (1) Multi-document summary extraction First, sentiment classification is performed on the input text information data based on media data and a large AI model. Then, the input document is further sentimentally classified based on the constructed multi-document summarization model to improve the quality and information richness of the document summary. At the same time, the model uses a graph attention mechanism to update information at multiple granularity nodes and complete label prediction to reduce summary redundancy. (2) Analysis of the popularity of online information topics Combine sentiment polarity and information heat to analyze information heat within a specific time period to obtain the sentiment evolution process of network information; Sentiment polarity analysis involves classifying the sentiment of text content under each topic based on a large AI model to determine its sentiment tendency, including counting the ratio of positive and negative sentiment tendencies within a certain time period to reflect the sentiment polarity of information data under each topic within that time period; information heat analysis is a comprehensive evaluation based on indicators including the number of posts, comments, and reposts of the information.

2. The multi-source information analysis method based on deep learning according to claim 1 is characterized in that: The multi-document summarization model includes an input layer, a feature extraction layer, a graph attention layer, and a sentence selection layer; Input layer: Extracts summary sentences from the input text information and represents it as an abstract task of predicting sentence label sequence {l1,l2,…l n }, where l n =m means that the nth sentence in the summary is extracted from the mth sentence in the document, and the input text information is segmented into sentences, and then the review topics are extracted through the LDA topic model; Feature extraction layer: The local features of each sentence and topic are captured through a convolutional neural network, and then the global features are obtained using a bidirectional long short-term memory network. Finally, the concatenation of local and global features is used as node features as the input of the graph attention layer; Graph Attention Layer: This layer constructs a heterogeneous graph containing topic, sentence, and document nodes. For sentence nodes, it uses a convolutional neural network to extract the n-gram feature vectors of each sentence in the text information and stores the extracted features. It then encodes the sentence node position as a vector feature. The two vectors are concatenated and mapped to a specified size to form the local features of the sentence. The local features and sentence length values ​​are then input into a Bi-LSTM model for further feature extraction. Finally, the two features are concatenated as the global features of the sentence node to improve the expressiveness of sentence features. Sentence selection layer: Use the maximum margin correlation algorithm to select the best summary sentence and finally complete the prediction of the candidate summary label.

3. The multi-source information analysis method based on deep learning according to claim 2 is characterized in that: In the graph attention layer, the input multi-document representation is D = {D1, D2, ... D nd }, where D i For the i-th input document, the topic in the document is extracted using the LDA model, which is recorded as the topic node T = {T1, T2, ... T nt }; The concatenation of the global features of the sentence nodes is represented as: in, represents the global features of the jth sentence, l j Represents the local n-gram features of the jth sentence, g j Represents the features obtained by the Bi-LSTM layer; The document node is regarded as a special sentence node, and the average value of all sentence features contained in it is used as the feature representation of the document node. Then, the heterogeneous graph is constructed using the sentences and the extracted topic nodes, which is expressed as follows: G=(N,E) Where N is a node set including sentence and topic nodes, E represents the weight matrix of undirected edges between nodes, and the frequencies of all topic words appearing in sentences and documents are converted into TF-IDF values ​​as the initial edge weight values ​​between corresponding nodes. Document nodes use TF-IDF values ​​as the edge weights connecting to topic nodes.

4. The multi-source information analysis method based on deep learning according to claim 3 is characterized in that: The information of sentence nodes is updated in the heterogeneous graph. For a sentence node, based on its neighboring topic nodes and their weight values, the weight value is normalized to obtain the attention coefficient between the sentence node and the adjacent nodes. The sentence nodes are weightedly aggregated according to the attention coefficient to obtain the new features of the sentence under the current attention head. The features under all attention heads are spliced ​​to obtain the features of the final sentence node and the heterogeneous graph is updated.

5. The multi-source information analysis method based on deep learning according to claim 3 is characterized in that: In the sentence selection layer, considering that there is relevant information in multiple input documents and there is a certain amount of character overlap between them, the maximum margin relevance algorithm selects the summary sentence based on the combination of the relevance of the sentence to the entire document and the novelty of the information contained in the sentence itself. The corresponding expression is as follows: Where R is the set of candidate sentences, S is the set of selected sentences, D is the document set, S i Indicates that the i-th sentence is not selected, S j Indicates that the jth sentence in the sentence set is selected, R\S is the set of unselected sentences, Sim1 represents the similarity between the i-th sentence and the document, and Sim2 represents the similarity between the candidate sentence and the selected sentence.

6. The multi-source information analysis method based on deep learning according to claim 1 or 2, characterized in that: This method is used for a certain time period T i , the information heat is calculated as follows: Among them, S heat (T i ) represents the time period T i The popularity of nz is T in the time period i The number of topics, b i Indicates the time period T i The number of posts, Represents the time period T i The average number of reposts, comments, and likes for related articles within the website; Combining sentiment polarity and information heat, the calculation formula for the comprehensive score of network information sentiment intensity is as follows: The changes in the emotional intensity of online information with time series characteristics are used to show the changes in the emotional intensity of online information from its generation to its demise. Based on the changes in emotional intensity, we can analyze and understand the public's emotional attitude towards the information and realize the tracking and analysis of the information.

7. A multi-source information sentiment analysis system based on deep learning, characterized in that: The system uses the method according to any one of claims 1 to 6 to perform sentiment analysis on text data, including: The multi-document summary extraction module improves the quality and information richness of document summaries by performing sentiment classification on input documents. It also uses the graph attention mechanism to update information at multi-granularity nodes and complete label prediction to reduce summary redundancy. The network information topic heat analysis module combines sentiment polarity and information heat to analyze the information heat within a certain time period. It is used to understand the emotional evolution process of network information, including social opinion analysis, and provide a basis for scientific decision-making and proactive management.

Images