Social event venation generation method based on double-layer structure entropy graph clustering

Abstract

The invention discloses a social event venation generation method based on double-layer structure entropy graph clustering, which adopts a double-layer clustering method based on structure entropy, performs theme clustering through coarse-grained clustering, performs fine-grained event clustering to form event nodes, and generates abstracts of the event nodes by using a large model; in addition, the composition not only considers the keyword co-occurrence frequency, but also considers the semantics through the SenteceBert calculation similarity, and considers the time span opposite side deletion between the messages, so that a good social message graph can be constructed, the graph clustering effect is further improved, and the social event timeline context analysis is better carried out.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and in particular relates to a method for generating social event context based on double-layer structure entropy graph clustering. Background Art

[0002] In today's era of rapid information development, social media platforms such as Twitter, Facebook and Instagram have become the main channels for people to communicate and share information in their daily lives. These platforms are characterized by a huge number of users and fast information dissemination. In these massive data, there is a large amount of information related to social events and hot topics. Through in-depth mining and analysis of social media data, the dynamic development of social events, especially the propagation trajectory and impact range of emergencies, can be effectively identified. However, it is not easy to accurately extract effective information from massive social media data, identify the evolution process of events, and generate intuitive timeline context. This requires not only the use of artificial intelligence algorithms, but also a deep understanding of language, context and the complexity of events. The massive data generated by these platforms contains rich information about social events. How to identify the dynamic development of events from them has important practical significance in crisis management, public opinion monitoring and decision-making.

[0003] This task is usually divided into a date selection subtask and a date summary subtask. The date selection subtask focuses on identifying dates related to key events from the text. These dates can be explicitly marked dates (such as "September 8, 2023") or implicit time information (such as "yesterday" and "last month"). Through date selection, the system determines which dates are directly related to the events described in the article. The date summary subtask aims to generate a brief event description for each selected date, highlighting the core events around these dates in the article. By generating a timeline, the system helps users better understand the order of events and their relevance. Traditional clustering algorithms have low clustering efficiency, especially when processing large-scale data. In addition, traditional composition methods do not distinguish the influence of different nodes in the semantic edge selection strategy, resulting in some important information being ignored, and the impact of time is not considered. There are still defects to a certain extent. Summary of the invention

[0004] In order to solve the above technical problems, the present invention designs a social event context generation method based on double-layer structural entropy graph clustering, adopts a double-layer, structural entropy-based clustering method, performs topic clustering through coarse-grained clustering, and forms event nodes through fine-grained event clustering, and uses a large model to generate summaries of event nodes; and the composition not only considers the co-occurrence frequency of keywords, but also takes into account the semantics through SentenceBert calculation of similarity, and considers the time span between messages to delete edges, which can construct a good social message graph, thereby improving the graph clustering effect and better performing social event timeline context analysis.

[0005] In order to achieve the above technical effects, the present invention discloses a method for generating social event context based on double-layer structure entropy graph clustering, comprising the following steps:

[0006] S1. Preprocessing of social media data;

[0007] S2. Apply the KeyBERT model to each message to be processed to extract keywords;

[0008] S3, construct a keyword graph and use a clustering algorithm based on structural entropy to perform coarse-grained topic clustering; specifically: when two keywords appear in the same message at the same time, traverse the keywords of each message and calculate the co-occurrence frequency of each keyword pair; and use the deduplicated keywords in S2 as nodes to construct a keyword co-occurrence graph, and build an edge between nodes with a co-occurrence frequency greater than a threshold (generally 2); after the graph is constructed, apply a community detection algorithm based on structural entropy to the graph to perform coarse-grained topic clustering.

[0009] S4, performing community selection on social media messages based on the results of S3;

[0010] S5: compose the messages and perform fine-grained event clustering in each community based on the results of S4; specifically:

[0011] S5.1: Treat each message as a node and construct a message graph; first, construct it by keyword frequency. Based on the keywords extracted from each message in S2, if the number of keywords in two messages is greater than or equal to 2, an edge is constructed between the two nodes;

[0012] S5.2: Then use SentenceBert to calculate the similarity between messages. If the similarity between two messages is less than 0.3 and there is an edge between the two messages, delete the existing edge. If the similarity is greater than 0.7 and there is no edge between the two messages, add an edge. If the similarity is between 0.3 and 0.7, do not make any changes.

[0013] S5.3: After matching the communities based on S4, extract subgraphs based on the messages within each community, and apply the event detection algorithm based on structural entropy within the community;

[0014] S6, extracting the time when the event occurred, and generating multi-document summaries for multiple messages of the event through a large model;

[0015] S7, sorting each event by time to generate an event timeline, that is, sorting the summary and content generated in S6 by time and outputting the result.

[0016] Further, in S4, after applying the community detection algorithm to the keyword co-occurrence graph in S3, each community is obtained as a bag-of-words model, each community is regarded as a document, the similarity between each message and the community is calculated, and the community with the highest similarity is selected for the message, that is, community matching;

[0017] Furthermore, in S6,

[0018] S6.1: Based on S5, the messages in the community are divided into multiple clusters. The messages in each cluster represent an event. Each message corresponds to its release time or timestamp. They are uniformly converted into timestamps. The timestamps of the messages in each cluster are summed and averaged, and then converted into time, accurate to year, month and day, as the time of occurrence of the event.

[0019] S6.2: For the summary of each cluster, use the large model to generate it, concatenate the message contents of the k nodes with the largest degree in the cluster, use "||" to separate each message, and then truncate the concatenated content according to the maximum number of input tokens supported by the large model. GPT3.5-turbo-16k supports a maximum of 16384 tokens. The excess is truncated, and a valid prompt "You are a helpful assistant. Three quotation marks inside is a Chinese multi-document article collection, each article is separated by a'||' separator, please provide a Chinese summarization for the multi-document article set, within 20 words." is constructed;

[0020] Furthermore, in S2, the pre-trained BERT model is used to generate a keyword set that best represents the topic of the message in combination with the content of the message; then the keywords of all messages are put into a set to remove duplication and used as nodes of the co-occurrence graph;

[0021] Furthermore, in said S1,

[0022] S1.1: For the collected social media data, filter out valid content, that is, content with more than 20 words, because short messages are usually regarded as noise or irrelevant information, which is not conducive to event detection and timeline construction;

[0023] S1.2: After the message is filtered, the word segmentation algorithm is used to segment the text.

[0024] The beneficial effects of the present invention are:

[0025] The present invention designs a social event context generation method based on double-layer structure entropy graph clustering. Through double-layer structure entropy clustering, social media messages can be effectively classified at both coarse-grained and fine-grained levels, and potential events and topics can be extracted, thereby significantly improving the accuracy of event detection. This method uses keyword co-occurrence frequency to form a graph, and calculates the similarity weights between messages through the SentenceBert model to more accurately capture the semantic relationship of messages.

[0026] In addition, the impact of time span on message relationships is considered to ensure that the constructed social message graph is more representative and the analysis results are more comprehensive. Using a large model to generate multi-document summaries and sorting events by event can effectively extract the key information of each event, provide a concise and clear event overview, and generate a clear event timeline, so that users can intuitively understand the development process of events and their chronological order. In contrast, traditional clustering algorithms are less efficient in processing large-scale data and are difficult to meet actual needs. The community detection algorithm based on structural entropy improves the efficiency and accuracy of clustering through a more efficient data processing mechanism, which is particularly suitable for complex social media data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the drawings required for describing the embodiment are briefly introduced below.

[0028] Figure 1 It is a flow chart of a method for generating social event context based on double-layer structure entropy graph clustering. DETAILED DESCRIPTION

[0029] Example 1

[0030] The present invention discloses a method for generating social event context based on double-layer structure entropy graph clustering, comprising the following steps:

[0031] Step 1: For the collected social media data, first filter out the valid content, that is, the content with more than 20 words. This is because short messages are usually regarded as noise or irrelevant information, which is not conducive to event detection and timeline construction.

[0032] Step 1.2: After the message is filtered, the word segmentation algorithm is used to segment the text.

[0033] Step 2: Apply the KeyBERT model to each message to be processed to extract keywords. KeyBERT is a keyword extraction algorithm based on the BERT model. It uses BERT's powerful semantic representation ability to extract the most representative keywords from the text. Use the pre-trained BERT model and combine it with the content of the message to generate a keyword set that best represents the topic of the message.

[0034] Step 2.1: We put the keywords of all messages into a set to remove duplication and use them as nodes of the co-occurrence graph.

[0035] Step 3: When two keywords appear in the same message at the same time, it is called a co-occurrence. Traverse the keywords of each message and calculate the co-occurrence frequency of each keyword pair.

[0036] Step 3.1: We use the keywords removed from Step 2.1 as nodes to construct a keyword co-occurrence graph. We construct an edge between nodes whose co-occurrence frequency is greater than a threshold (usually 2). After the graph is constructed, a community detection algorithm based on structural entropy is applied to the graph to perform coarse-grained topic clustering.

[0037] Step 4: After applying the community detection algorithm to the keyword co-occurrence graph in Step 3, each community is actually a bag-of-words model. We regard each community as a document, calculate the similarity between each message and the community, and select the community with the highest similarity for the message, that is, community matching.

[0038] Step 5: We regard each message as a node and construct a message graph. First, we construct it through keyword frequency. In Step 2, we extract keywords from each message. When the number of keywords in two messages is greater than or equal to 2, we construct an edge between the two nodes. Then we use SentenceBert to calculate the similarity between messages. If the similarity between the two messages is less than 0.3 and there is an edge between the two messages, we delete the edge. If the similarity is greater than 0.7 and there is no edge between the two messages, we add an edge. If the similarity is between 0.3 and 0.7, we do not make any changes.

[0039] Step5.1: After the communities in Step4 are matched, subgraphs are extracted based on the messages within each community, and an event detection algorithm based on structural entropy is applied within the community.

[0040] Step 6: In Step 5, we divided the messages within the community into multiple clusters. The messages in each cluster represent an event. Each message has its release time or timestamp. We convert them into timestamps, add up the timestamps of the messages in each cluster and calculate the average, and then convert them into time, accurate to year, month and day, as the time of occurrence of the event.

[0041] Step 6.1: To generate the summary of each cluster, we use the big model to generate it. We concatenate the message contents of the k nodes with the largest degree in the cluster, use "||" to separate each message, and then truncate the concatenated content according to the maximum number of input tokens supported by the big model. GPT3.5-turbo-16k supports a maximum of 16384 tokens. The excess is truncated, and a valid prompt is constructed: "You are a helpful assistant. Three quotation marks inside is a Chinese multi-document article collection, each article is separated by a'||' separator, please provide a Chinese summarization for the multi-document article set, within 20 words." Then pass the concatenated content into the prompt.

[0042] Step7: Output the summary and content generated in Step6 in chronological order.

[0043] In this embodiment, the feasibility of the above method is verified by further practical operation; specific experiments are carried out, and the results obtained are shown in Table 1 and Table 2 below:

[0044] Table 1 The first set of output results of the actual experimental verification

[0045]

[0046] Table 2 The second set of output results actually verified by the experiment

[0047]

[0048] Tables 1 and 2 above are the results of direct output. Now take the results in Table 1 as an example. The specific operation process is as follows:

[0049] The input is as follows. Each message will have a timestamp when it was published. In addition to the messages about Double Eleven and NBA in Examples 1 and 2, there are many other topics, such as telecommunications fraud, Champions League, etc. Each message is attached with Figure 1 The node document d in the table is shown in Table 1.1 below.

[0050] Table 1.1 Original timestamp and message input

[0051]

[0052] After the first level of clustering, related messages will be in the same cluster, that is, messages related to Double 11, NBA, and UEFA Champions League will be in different clusters. Figure 1 After subgraph extraction, the documents are divided into different clusters (two categories in the example in the figure). After this step, the Double Eleven related messages in the above input will be gathered in one cluster. As shown in Table 1.2 below:

[0053] Table 1.2 Clustering of documents after subgraph extraction

[0054]

[0055]

[0056] Next, we perform fine-grained clustering within this cluster, that is, clustering the nodes related to Double Eleven. After clustering, messages describing the same event will gather into a new small cluster, as shown in the following figure. Figure 1 After the community internal event detection, document d appears in different colors, and each color represents a small cluster. In the above table, the first and second messages are a cluster respectively, and the last three messages are a cluster, that is, three events e.

[0057] Finally, we generate time and summary for each event (cluster), add and average the timestamps of each message in the cluster and convert them into time format as the time of event occurrence; for summary generation, taking a cluster with three messages as an example, we sort the three messages in descending order according to the degree of the nodes, and splice the three messages together in this order, intercept them according to the maximum number of input tokens of the large model used, that is, complete the selection of top_k nodes, and after interception, bring the constructed prompt and input it to the large model to get the following summary: "Double Eleven Shopping Festival sales set a record, Jingzhou and Qinghai e-commerce performed outstandingly, and sales of Three Squirrels and Wenzhou footwear companies increased significantly." After executing this process for each event cluster, sort them by time to get the event timeline.

[0058] The content in Table 2 can also be obtained in the same way. After executing this process for each event cluster, the event timeline can be obtained by sorting them by time, and no repetition will be made here.

[0059] Through the above experiments, it can be clearly seen that using a large model to generate multi-document summaries and sorting events by event can effectively extract the key information of each event, provide a concise and clear event overview, and generate a clear event timeline, so that users can intuitively understand the development process of events and their chronological order. The community detection algorithm based on structural entropy improves the efficiency and accuracy of clustering through a more efficient data processing mechanism, which is particularly suitable for complex social media data analysis.

[0060] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described.

Claims

1. A method for generating social event context based on double-layer structure entropy graph clustering, characterized in that: The following steps are involved: S1. Preprocessing of social media data; S2. Apply the KeyBERT model to each message to be processed to extract keywords; S3, construct a keyword graph and use a clustering algorithm based on structural entropy to perform coarse-grained topic clustering; specifically: when two keywords appear in the same message at the same time, traverse the keywords of each message and calculate the co-occurrence frequency of each keyword pair; and use the deduplicated keywords in S2 as nodes to construct a keyword co-occurrence graph, and construct an edge between nodes whose co-occurrence frequency is greater than a threshold (usually 2); After the graph is constructed, a community detection algorithm based on structural entropy is applied to the graph to perform coarse-grained topic clustering; S4, performing community selection on social media messages based on the results of S3; S5: compose the messages and perform fine-grained event clustering in each community based on the results of S4; specifically: S5.1: Treat each message as a node and construct a message graph; first, construct it by keyword frequency. Based on the keywords extracted from each message in S2, if the number of keywords in two messages is greater than or equal to 2, an edge is constructed between the two nodes; S5.2: Then use SentenceBert to calculate the similarity between messages. If the similarity between two messages is less than 0.3 and there is an edge between the two messages, delete the existing edge. If the similarity is greater than 0.7 and there is no edge between the two messages, add an edge. If the similarity is between 0.3 and 0.7, do not make any changes. S5.3: After community matching based on S4, subgraphs are extracted based on messages within each community, and an event detection algorithm based on structural entropy is applied within the community; S6, extracting the time when the event occurred, and generating multi-document summaries for multiple messages of the event through a large model; S7, sorting each event by time to generate an event timeline, that is, sorting the summary and content generated in S6 by time and outputting the result.

2. According to claim 1, a method for generating social event context based on double-layer structure entropy graph clustering is characterized in that: In S4, after applying the community detection algorithm to the keyword co-occurrence graph in S3, each community is obtained as a bag-of-words model. Each community is regarded as a document, the similarity between each message and the community is calculated, and the community with the highest similarity is selected for the message, that is, community matching.

3. According to claim 1, a method for generating social event context based on double-layer structure entropy graph clustering, characterized in that: In S6, S6.1: Based on S5, the messages in the community are divided into multiple clusters. The messages in each cluster represent an event. Each message corresponds to its release time or timestamp. They are uniformly converted into timestamps. The timestamps of the messages in each cluster are summed and averaged, and then converted into time, accurate to year, month and day, as the time of occurrence of the event. S6.2: For the summary of each cluster, use the large model to generate it. The message contents of the k nodes with the largest degree in the cluster are spliced ​​together. Each message is separated by "||". Then the spliced ​​content is truncated according to the maximum number of input tokens supported by the large model. GPT3.5-turbo-16k supports a maximum of 16384 tokens. The excess is truncated, and a valid prompt "You are a helpful assistant. Three quotation marks inside is a Chinese multi-document article collection, each article is separated by a'||'separator, please provide a Chinese summarization for the multi-document article set, within 20 words." is constructed.

4. According to claim 1, a method for generating social event context based on double-layer structure entropy graph clustering, characterized in that: In S2, a pre-trained BERT model is used to generate a keyword set that best represents the topic of the message in combination with the content of the message; then the keywords of all messages are put into a set to remove duplication and used as nodes of the co-occurrence graph.

5. According to claim 1, a method for generating social event context based on double-layer structure entropy graph clustering, characterized in that: In S1, S1.1: For the collected social media data, filter out valid content, that is, content with more than 20 words, because short messages are usually regarded as noise or irrelevant information, which is not conducive to event detection and timeline construction; S1.2: After the message is filtered, the word segmentation algorithm is used to segment the text.

Images