Keyword-driven whole network public opinion analysis system and method based on autoGLM

Abstract

The application discloses a keyword-driven full-network public opinion analysis system and method based on AutoGLM, and belongs to the technical field of artificial intelligence. The method comprises the following steps: collecting full-network multi-source data based on keyword driving; cleaning, deduplicating and standardizing the collected data for preprocessing; tracing and clustering the source of public opinion based on the semantic correlation of AutoGLM; guiding fine-grained sentiment tendency determination by a sceneized prompt word; evaluating the risk level by dynamic weight of multi-dimensional indexes; extracting public opinion summary and key findings by AutoGLM optimization; performing semantic deduplication on the extracted results and checking logical consistency; and outputting structured data and visually displaying the propagation path. The problems of incomplete public opinion source tracing, insufficient sentiment analysis accuracy, inaccurate risk level evaluation, low core information extraction efficiency and poor model adaptability in the prior art are solved, and efficient, accurate and comprehensive full-network public opinion analysis is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, natural language processing, and online public opinion analysis, and particularly to a keyword-driven online public opinion analysis system and method based on AutoGLM. Background Technology

[0002] With the rapid development of internet technology, social media, news platforms, forums, and other channels have generated massive amounts of public opinion information. The importance of public opinion analysis in areas such as brand monitoring, public affairs management, and risk warning is becoming increasingly prominent. Users are increasingly eager to search for target public opinion using keywords, requiring them to quickly obtain key information such as the source of the opinion, its sentiment, risk level, and core information.

[0003] Current mainstream public opinion analysis technologies mostly rely on traditional natural language processing models or general search engines, which have significant limitations: 1. Incomplete tracing of public opinion sources: Existing technologies mostly focus on data collection from mainstream platforms, with insufficient coverage of scattered channels such as niche forums and vertical communities, and it is difficult to accurately locate the initial dissemination node of public opinion, resulting in fragmented source analysis.

[0004] 2. Insufficient accuracy in sentiment analysis: Traditional models have limited ability to identify ambiguous semantics, metaphorical expressions, and context-dependent expressions, which can easily lead to misjudgment of sentiment. Furthermore, they cannot dynamically adjust the analysis dimensions in combination with the public opinion dissemination scenario, making it difficult to meet the needs of refined analysis.

[0005] 3. Risk level assessment is highly subjective: Existing risk assessments mostly rely on fixed rule bases and lack comprehensive consideration of dynamic factors such as the speed of public opinion dissemination, the scope of influence, and the characteristics of the participants. As a result, the assessment results deviate significantly from the actual risks.

[0006] 4. Low efficiency in extracting core information: Faced with massive amounts of public opinion data, traditional models struggle to quickly filter key information, resulting in redundant summary content, lack of emphasis on key findings, and issues of information duplication or omission, thus failing to provide users with efficient decision support.

[0007] 5. Insufficient model adaptability: The general-purpose large language model is not optimized for public opinion analysis scenarios. When processing short text public opinion and multimodal public opinion (text + image / video caption), there are problems such as insufficient contextual understanding and low extraction accuracy. Summary of the Invention

[0008] This invention provides a keyword-driven online public opinion analysis system and method based on AutoGLM. It addresses the problems of incomplete source tracing, insufficient accuracy of sentiment analysis, inaccurate risk level assessment, low efficiency in extracting core information, and poor model adaptability in existing technologies, achieving efficient, accurate, and comprehensive online public opinion analysis.

[0009] According to one aspect of this disclosure, a keyword-driven, full-network public opinion analysis method based on AutoGLM is provided, the method comprising: S1, based on keyword-driven multi-source data collection across the entire network; S2, performs cleaning, deduplication, and standardization preprocessing on the collected data; S3, Public Opinion Source Tracing and Clustering Based on AutoGLM Semantic Association; S4, fine-grained judgment of sentiment guided by the contextual prompt word "Prompt"; S5, risk level assessment with dynamic weights of multi-dimensional indicators; S6, AutoGLM-optimized sentiment summarization and key finding extraction; S7, semantic deduplication and logical consistency verification of the extracted results; S8 provides a visual representation of structured data output and propagation paths.

[0010] In one possible implementation, S1 involves keyword-driven multi-source data collection across the entire network, including: Based on user-input keywords, relevant data is obtained through various methods such as API calls, web crawlers, or RSS subscriptions. Dynamic collection frequency is set, and the collection interval is adaptively adjusted according to the popularity of public opinion.

[0011] In one possible implementation, S2 performs cleaning, deduplication, and standardization preprocessing on the collected data, including: The collected raw data is cleaned, invalid information is removed, the data format is standardized, and the text content, publication time, publishing account, dissemination volume and other metadata are extracted to build a standardized data pool. Data deduplication and noise reduction: Text fingerprint comparison technology is used to remove duplicate data, and AutoGLM is used to identify and filter irrelevant information to ensure the purity and validity of the data pool; Invalid information includes advertisements and irrelevant replies; dissemination data includes likes, comments, and shares.

[0012] In one possible implementation, S3, based on AutoGLM semantic association for tracing and clustering public opinion sources, includes: AutoGLM-based public opinion source tracing and clustering leverages AutoGLM's semantic understanding and association analysis capabilities to achieve accurate tracing and clustering of public opinion sources, including: Semantic association analysis: The preprocessed data is input into AutoGLM. By designing a special prompt word, the model is guided to analyze the semantic similarity, chronological order of publication, and account association of each piece of public opinion data, and to construct a semantic network for public opinion dissemination. Initial source location: Based on semantic network and time series analysis, identify the earliest published public opinion data with propagation and diffusion characteristics, and determine the initial propagation node information of public opinion, including the publishing account, platform, and time; Propagation path tracing: By analyzing the correlations of citations, reposts, and comments among various public opinion data through AutoGLM, the diffusion path of public opinion from the initial node to various dissemination channels is reconstructed, and a source tracing map is generated; Source Clustering: Based on the type of dissemination channel, account attributes, and geographical distribution, AutoGLM is used to perform cluster analysis to classify the sources of public opinion into different categories, making it easier for users to quickly grasp the source structure; The sources of public opinion include official media, self-media, ordinary users, and industry organizations.

[0013] In one possible implementation, S4, the contextualized prompt word Prompt, guides fine-grained sentiment assessment, including: A fine-grained sentiment analysis model was built based on AutoGLM, and the prompt word engineering and analysis process was optimized: Prompt Design: Construct contextualized sentiment analysis prompts, explicitly requiring AutoGLM to combine the contextual information of the public opinion release scenario to perform fine-grained analysis of the sentiment tendency of the text, output positive, neutral, and negative judgment results, and attach a sentiment intensity score. The higher the score, the stronger the sentiment. The release scenarios include: product reviews, public event discussions, and industry comments. Complex semantic processing: For complex emotional expressions such as ambiguity, metaphor, and irony, AutoGLM's deep semantic understanding capabilities are used to make inferences and judgments in combination with contextual information, avoiding misjudgments based on surface semantics; Sentiment Consistency Verification: After performing sentiment analysis on multiple relevant data points for the same public opinion event, the consistency of sentiment tendencies is verified using AutoGLM. If significant contradictions exist, the reasons for the contradictions are further analyzed and annotated. The reasons for the contradictions include information bias and different stances.

[0014] In one possible implementation, S5, a risk level assessment with dynamic weights for multi-dimensional indicators, includes: A multi-dimensional dynamic risk level assessment model is constructed to establish a multi-dimensional risk assessment indicator system, and AutoGLM is used for comprehensive reasoning and assessment. The evaluation indicator system is constructed, including dissemination, impact, sentiment, and content dimensions. The dissemination dimension includes dissemination speed, number of platforms covered by dissemination, and cumulative dissemination volume. The impact dimension includes the number of users participating in the discussion, user activity, and media attention. The sentiment dimension includes the proportion of negative sentiment and the intensity of negative sentiment. The content dimension includes the sensitivity of public opinion content and whether it involves illegal or irregular information. Indicator weight allocation: The initial weights of each indicator are determined by the analytic hierarchy process, and then dynamically adjusted using AutoGLM combined with historical public opinion data to ensure that the weight allocation conforms to the actual risk impact pattern. Risk level reasoning: Input the data of various dimensions into AutoGLM, guide the model to perform comprehensive reasoning through the special prompt word Prompt, output three risk levels of low, medium and high according to the preset risk threshold range, and generate a risk assessment report explaining the key influencing factors.

[0015] In one possible implementation, S6, AutoGLM-optimized sentiment summarization and key finding extraction includes: Public opinion summary generation: Design the summary generation prompt word Prompt. AutoGLM is required to extract the core events, main viewpoints and dissemination status based on all relevant public opinion data, and generate a concise and comprehensive summary. The summary length can be adjusted according to user needs to ensure that no important information is omitted and there is no redundant content. Key findings extraction: AutoGLM is guided by the prompt word "Prompt" to extract key findings from public opinion data, including core points of contention, popular viewpoints, potential trends, and key stakeholders, highlighting the core value information that users care about; Deduplication and optimization: The summaries and key findings generated by AutoGLM are deduplicated by comparing semantic similarity to remove duplicate content, ensuring the uniqueness and validity of the information; at the same time, the prompt words are continuously optimized based on user feedback to improve the extraction quality.

[0016] In one possible implementation, S8, the visualization of structured data output and propagation paths, includes: Structured output: The public opinion analysis results are output in a standardized format, including: keywords, data collection time range, public opinion source statistics, sentiment analysis results, risk level and assessment basis, public opinion summary, key findings, and support for export in Excel and PDF formats; among them, the public opinion source statistics include: initial source, main dissemination channels, source clustering results, and sentiment analysis results include: positive / neutral / negative percentage and sentiment intensity distribution. Visualization: Construct a visual interface to display the sentiment ratio, source distribution, and risk level composition through charts, and to show the path of public opinion dissemination through graphs, presenting the analysis results intuitively and making it easy for users to understand and use; charts include: pie charts, bar charts, or line charts.

[0017] A keyword-driven, full-network public opinion analysis system based on AutoGLM, comprising: The data acquisition module is based on keyword-driven multi-source data acquisition from the entire network; The preprocessing module performs cleaning, deduplication, and standardization preprocessing on the collected data; The public opinion source tracing and clustering module is based on AutoGLM semantic association for public opinion source tracing and clustering; The sentiment bias fine-grained judgment module uses the contextualized prompt word "Prompt" to guide the fine-grained judgment of sentiment bias. Risk level assessment module: Risk level assessment with dynamic weights of multi-dimensional indicators; The public opinion summary and key discovery extraction module uses AutoGLM to optimize public opinion summarization and key discovery extraction. The semantic deduplication and logical consistency verification module performs semantic deduplication and logical consistency verification on the extracted results. The visualization module provides a visual representation of the structured data output and propagation path.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. Comprehensive and accurate traceability of public opinion sources: By integrating data from mainstream platforms and niche channels through a multi-source data collection engine, and combining it with AutoGLM's semantic association analysis capabilities, the initial dissemination nodes and diffusion paths of public opinion can be accurately located, achieving a balance between comprehensive source coverage and accurate traceability.

[0019] 2. Significantly improved sentiment analysis accuracy: Leveraging AutoGLM's deep understanding of complex semantics, the design of sentiment analysis prompts has been optimized, and a contextualized semantic adaptation mechanism has been introduced. This enables accurate identification of complex sentiment tendencies such as ambiguous expressions and metaphors, fine-grained differentiation of positive, neutral, and negative sentiments, and a significant reduction in the false positive rate.

[0020] 3. Objective and accurate risk level assessment: Construct a multi-dimensional assessment indicator system (spread speed, scope of impact, participating population, emotional intensity, etc.), and integrate data from various dimensions through AutoGLM for comprehensive reasoning to avoid the limitations of fixed rules and achieve objective and accurate risk level (low, medium, high).

[0021] 4. Highly efficient and accurate extraction of core information: By leveraging AutoGLM's summarization and key information extraction capabilities, the prompt word engineering is optimized to ensure that the public opinion summary is concise and comprehensive, highlighting key findings. At the same time, the deduplication algorithm avoids information duplication, significantly improving the efficiency and quality of information extraction.

[0022] 5. Strong model adaptability: AutoGLM has been fine-tuned and optimized for public opinion analysis scenarios to adapt to various forms of public opinion data such as short texts and multimodal captions, balancing processing speed and analysis accuracy to meet the needs of public opinion analysis in different scenarios.

[0023] 6. Results are presented in an intuitive and easy-to-understand manner: Through structured output and visualization, information such as the source of public opinion, sentiment, risk level, summary, and key findings are clearly presented, making it easy for users to quickly obtain core value information and support decision-making.

[0024] A keyword-driven online public opinion analysis system and method based on AutoGLM is proposed to solve the problems of incomplete source tracing, insufficient accuracy of sentiment analysis, inaccurate risk level assessment, low efficiency of core information extraction, and poor model adaptability in existing technologies, thereby achieving efficient, accurate, and comprehensive online public opinion analysis. Attached Figure Description

[0025] Figure 1 The flowchart of a keyword-driven online public opinion analysis method based on AutoGLM according to an embodiment of this disclosure is shown.

[0026] Figure 2 A flowchart of a keyword-driven online public opinion analysis method based on AutoGLM, according to another embodiment of this disclosure, is shown. Detailed Implementation

[0027] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0028] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0029] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0030] A keyword-driven, full-network public opinion analysis method based on AutoGLM, the method comprising: S1, based on keyword-driven multi-source data collection across the entire network; S2, performs cleaning, deduplication, and standardization preprocessing on the collected data; S3, Public Opinion Source Tracing and Clustering Based on AutoGLM Semantic Association; S4, fine-grained judgment of sentiment guided by the contextual prompt word "Prompt"; S5, risk level assessment with dynamic weights of multi-dimensional indicators; S6, AutoGLM-optimized sentiment summarization and key finding extraction; S7, semantic deduplication and logical consistency verification of the extracted results; S8 provides a visual representation of structured data output and propagation paths.

[0031] In one possible implementation, S1 involves keyword-driven multi-source data collection across the entire network, including: Based on user-input keywords, relevant data is obtained through various methods such as API calls, web crawlers, or RSS subscriptions. Dynamic collection frequency is set, and the collection interval is adaptively adjusted according to the popularity of public opinion.

[0032] In one possible implementation, S2 performs cleaning, deduplication, and standardization preprocessing on the collected data, including: The collected raw data is cleaned, invalid information is removed, the data format is standardized, and the text content, publication time, publishing account, dissemination volume and other metadata are extracted to build a standardized data pool. Data deduplication and noise reduction: Text fingerprint comparison technology is used to remove duplicate data, and AutoGLM is used to identify and filter irrelevant information to ensure the purity and validity of the data pool; Invalid information includes advertisements and irrelevant replies; dissemination data includes likes, comments, and shares.

[0033] In one possible implementation, S3, based on AutoGLM semantic association for tracing and clustering public opinion sources, includes: AutoGLM-based public opinion source tracing and clustering leverages AutoGLM's semantic understanding and association analysis capabilities to achieve accurate tracing and clustering of public opinion sources, including: Semantic association analysis: The preprocessed data is input into AutoGLM. By designing a special prompt word, the model is guided to analyze the semantic similarity, chronological order of publication, and account association of each piece of public opinion data, and to construct a semantic network for public opinion dissemination. Initial source location: Based on semantic network and time series analysis, identify the earliest published public opinion data with propagation and diffusion characteristics, and determine the initial propagation node information of public opinion, including the publishing account, platform, and time; Propagation path tracing: By analyzing the correlations of citations, reposts, and comments among various public opinion data through AutoGLM, the diffusion path of public opinion from the initial node to various dissemination channels is reconstructed, and a source tracing map is generated; Source Clustering: Based on the type of dissemination channel, account attributes, and geographical distribution, AutoGLM is used to perform cluster analysis to classify the sources of public opinion into different categories, making it easier for users to quickly grasp the source structure; The sources of public opinion include official media, self-media, ordinary users, and industry organizations.

[0034] In one possible implementation, S4, the contextualized prompt word Prompt, guides fine-grained sentiment assessment, including: A fine-grained sentiment analysis model was built based on AutoGLM, and the prompt word engineering and analysis process was optimized: Prompt Design: Construct contextualized sentiment analysis prompts, explicitly requiring AutoGLM to combine the contextual information of the public opinion release scenario to perform fine-grained analysis of the sentiment tendency of the text, output positive, neutral, and negative judgment results, and attach a sentiment intensity score. The higher the score, the stronger the sentiment. The release scenarios include: product reviews, public event discussions, and industry comments. Complex semantic processing: For complex emotional expressions such as ambiguity, metaphor, and irony, AutoGLM's deep semantic understanding capabilities are used to make inferences and judgments in combination with contextual information, avoiding misjudgments based on surface semantics; Sentiment Consistency Verification: After performing sentiment analysis on multiple relevant data points for the same public opinion event, the consistency of sentiment tendencies is verified using AutoGLM. If significant contradictions exist, the reasons for the contradictions are further analyzed and annotated. The reasons for the contradictions include information bias and different stances.

[0035] In one possible implementation, S5, a risk level assessment with dynamic weights for multi-dimensional indicators, includes: A multi-dimensional dynamic risk level assessment model is constructed to establish a multi-dimensional risk assessment indicator system, and AutoGLM is used for comprehensive reasoning and assessment. The evaluation indicator system is constructed, including dissemination, impact, sentiment, and content dimensions. The dissemination dimension includes dissemination speed, number of platforms covered by dissemination, and cumulative dissemination volume. The impact dimension includes the number of users participating in the discussion, user activity, and media attention. The sentiment dimension includes the proportion of negative sentiment and the intensity of negative sentiment. The content dimension includes the sensitivity of public opinion content and whether it involves illegal or irregular information. Indicator weight allocation: The initial weights of each indicator are determined by the analytic hierarchy process, and then dynamically adjusted using AutoGLM combined with historical public opinion data to ensure that the weight allocation conforms to the actual risk impact pattern. Risk level reasoning: Input the data of various dimensions into AutoGLM, guide the model to perform comprehensive reasoning through the special prompt word Prompt, output three risk levels of low, medium and high according to the preset risk threshold range, and generate a risk assessment report explaining the key influencing factors.

[0036] In one possible implementation, S6, AutoGLM-optimized sentiment summarization and key finding extraction includes: Public opinion summary generation: Design the summary generation prompt word Prompt. AutoGLM is required to extract the core events, main viewpoints and dissemination status based on all relevant public opinion data, and generate a concise and comprehensive summary. The summary length can be adjusted according to user needs to ensure that no important information is omitted and there is no redundant content. Key findings extraction: AutoGLM is guided by the prompt word "Prompt" to extract key findings from public opinion data, including core points of contention, popular viewpoints, potential trends, and key stakeholders, highlighting the core value information that users care about; Deduplication and optimization: The summaries and key findings generated by AutoGLM are deduplicated by comparing semantic similarity to remove duplicate content, ensuring the uniqueness and validity of the information; at the same time, the prompt words are continuously optimized based on user feedback to improve the extraction quality.

[0037] In one possible implementation, S8, the visualization of structured data output and propagation paths, includes: Structured output: The public opinion analysis results are output in a standardized format, including: keywords, data collection time range, public opinion source statistics, sentiment analysis results, risk level and assessment basis, public opinion summary, key findings, and support for export in Excel and PDF formats; among them, the public opinion source statistics include: initial source, main dissemination channels, source clustering results, and sentiment analysis results include: positive / neutral / negative percentage and sentiment intensity distribution. Visualization: Construct a visual interface to display the sentiment ratio, source distribution, and risk level composition through charts, and to show the path of public opinion dissemination through graphs, presenting the analysis results intuitively and making it easy for users to understand and use; charts include: pie charts, bar charts, or line charts.

[0038] A keyword-driven, full-network public opinion analysis system based on AutoGLM, comprising: The data acquisition module is based on keyword-driven multi-source data acquisition from the entire network; The preprocessing module performs cleaning, deduplication, and standardization preprocessing on the collected data; The public opinion source tracing and clustering module is based on AutoGLM semantic association for public opinion source tracing and clustering; The sentiment bias fine-grained judgment module uses the contextualized prompt word "Prompt" to guide the fine-grained judgment of sentiment bias. Risk level assessment module: Risk level assessment with dynamic weights of multi-dimensional indicators; The public opinion summary and key discovery extraction module uses AutoGLM to optimize public opinion summarization and key discovery extraction. The semantic deduplication and logical consistency verification module performs semantic deduplication and logical consistency verification on the extracted results. The visualization module provides a visual representation of the structured data output and propagation path.

[0039] Application Example: A keyword-driven, full-network public opinion analysis method based on AutoGLM, the method including: S1, based on keyword-driven multi-source data collection across the entire network; S2, perform cleaning, deduplication, and standardization preprocessing on the collected data; after this step, determine whether the data cleaning, deduplication, and standardization preprocessing are complete. If not, re-execute the data cleaning and preprocessing and continue collecting multi-source data. If yes, proceed to the next step. S3, Public Opinion Source Tracing and Clustering Based on AutoGLM Semantic Association; S4, fine-grained sentiment assessment guided by the contextualized prompt word Prompt; determine whether this step is completed. If not, re-execute the sentiment assessment, and then continue to perform public opinion source tracing and clustering. If yes, proceed to the next step. S5, risk level assessment with dynamic weights of multi-dimensional indicators; S6, AutoGLM-optimized public opinion summary and key findings extraction; determine whether this step is complete. If not, re-execute the summary and key findings extraction, then continue with the risk level assessment. If yes, proceed to the next step. S7, semantic deduplication and logical consistency verification of the extracted results; S8 outputs structured data and visualizes its propagation path. It then determines whether the output and visualization are complete. If not, it re-executes the output and visualization, and then continues to perform deduplication and logical consistency checks.

[0040] like Figure 2 As shown, the flowchart starts with keywords, realizes multi-source collection and preprocessing of public opinion across the entire network, completes source tracing, sentiment judgment, summary extraction and multi-dimensional risk classification through AutoGLM, ensures the accuracy of analysis through closed-loop retry, and finally outputs structured results and visualization.

[0041] The core of the public opinion monitoring platform interface in this embodiment is to display the public opinion data and details of each company. The functions of each area are as follows: The left sidebar is a list of monitoring companies, and you can add new monitoring topics by "adding keywords"; Top overview area: Displays core data: total number of public opinion events (10), positive public opinion events (5), negative public opinion events (2), and high-risk public opinion events (1). Intermediate statistics area: Left: A pie chart showing the distribution of sentiment analysis, visually presenting the proportion of positive, negative, and neutral public opinion; Right: A pie chart showing the percentage of high, medium, and low risk levels; The bottom list area is for "Latest Public Opinion News," which can be filtered by date, sentiment, and risk. It displays the company, title, source (such as Weibo or Zhihu), sentiment, risk level, and time for each piece of public opinion news. It also supports exporting to Excel and viewing details.

[0042] The interface enables centralized display, statistical visualization, and detailed query of public opinion data, allowing users to quickly grasp the public opinion dynamics of various enterprises.

[0043] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A keyword-driven, full-network public opinion analysis method based on AutoGLM, characterized in that, The method includes: S1, based on keyword-driven multi-source data collection across the entire network; S2, performs cleaning, deduplication, and standardization preprocessing on the collected data; S3, Public Opinion Source Tracing and Clustering Based on AutoGLM Semantic Association; S4, fine-grained judgment of sentiment tendency guided by the contextual prompt word Prompt; S5, risk level assessment with dynamic weights of multi-dimensional indicators; S6, AutoGLM-optimized public opinion summarization and key finding extraction; S7, semantic deduplication and logical consistency verification of the extracted results; S8 provides a visual representation of structured data output and propagation paths.

2. The keyword-driven online public opinion analysis method based on AutoGLM according to claim 1, characterized in that, S1, based on keyword-driven multi-source data collection across the entire network, including: Data is obtained through various methods, including API calls, web crawlers, and RSS subscriptions, based on keywords input by users. The collection frequency is set dynamically, and the collection interval is adjusted adaptively according to the popularity of public opinion.

3. The keyword-driven online public opinion analysis method based on AutoGLM according to claim 1, characterized in that, S2 performs cleaning, deduplication, and standardization preprocessing on the collected data, including: The collected raw data is cleaned, invalid information is removed, the data format is standardized, and the text content, publication time, publishing account, dissemination volume and other metadata are extracted to build a standardized data pool. Data deduplication and noise reduction: Text fingerprint comparison technology is used to remove duplicate data, and AutoGLM is used to identify and filter irrelevant information to ensure the purity and validity of the data pool; Invalid information includes advertisements and irrelevant replies; dissemination data includes likes, comments, and shares.

4. The keyword-driven online public opinion analysis method based on AutoGLM according to claim 1, characterized in that, S3, based on AutoGLM semantic association for tracing and clustering public opinion sources, includes: AutoGLM-based public opinion source tracing and clustering leverages AutoGLM's semantic understanding and association analysis capabilities to achieve accurate tracing and clustering of public opinion sources, including: Semantic association analysis: The preprocessed data is input into AutoGLM. By designing a special prompt word, the model is guided to analyze the semantic similarity, release time sequence, and account association of each piece of public opinion data, and to construct a semantic network for public opinion dissemination. Initial source location: Based on semantic network and time series analysis, identify the earliest published public opinion data with propagation and diffusion characteristics, and determine the initial propagation node information of public opinion, including the publishing account, platform, and time; Propagation path tracing: By analyzing the correlations of citations, forwards, and comments among various public opinion data through AutoGLM, the diffusion path of public opinion from the initial node to various dissemination channels is reconstructed, and a source tracing map is generated; Source Clustering: Based on the type of dissemination channel, account attributes, and geographical distribution, AutoGLM is used to perform cluster analysis to classify the sources of public opinion into different categories, making it easier for users to quickly grasp the source structure; The sources of public opinion include official media, self-media, ordinary users, and industry organizations.

5. The keyword-driven online public opinion analysis method based on AutoGLM according to claim 1, characterized in that, S4, fine-grained sentiment assessment guided by the contextualized prompt word "Prompt," including: A fine-grained sentiment analysis model was built based on AutoGLM, and the prompt word engineering and analysis process was optimized: Prompt Design: Construct contextualized sentiment analysis prompts, explicitly requiring AutoGLM to combine the contextual information of the public opinion release scenario to perform fine-grained analysis of the sentiment tendency of the text, output positive, neutral, and negative judgment results, and attach a sentiment intensity score. The higher the score, the stronger the sentiment. The release scenarios include: product reviews, public event discussions, and industry comments. Complex semantic processing: For complex emotional expressions such as ambiguity, metaphor, and irony, AutoGLM's deep semantic understanding capabilities are used to make inferences and judgments in combination with contextual information, avoiding misjudgments based on surface semantics; Sentiment Consistency Verification: After performing sentiment analysis on multiple relevant data points for the same public opinion event, the consistency of sentiment tendencies is verified using AutoGLM. If significant contradictions exist, the reasons for the contradictions are further analyzed and annotated. The reasons for the contradictions include information bias and different stances.

6. The keyword-driven online public opinion analysis method based on AutoGLM according to claim 1, characterized in that, S5, risk level assessment with dynamic weights for multi-dimensional indicators, including: A multi-dimensional dynamic risk level assessment model is constructed to establish a multi-dimensional risk assessment indicator system, and AutoGLM is used for comprehensive reasoning and assessment. The evaluation indicator system is constructed, including dissemination, impact, sentiment, and content dimensions. The dissemination dimension includes dissemination speed, number of platforms covered by dissemination, and cumulative dissemination volume. The impact dimension includes the number of users participating in the discussion, user activity, and media attention. The sentiment dimension includes the proportion of negative sentiment and the intensity of negative sentiment. The content dimension includes the sensitivity of public opinion content and whether it involves illegal or irregular information. Indicator weight allocation: The initial weights of each indicator are determined by the Analytic Hierarchy Process (AHP), and then dynamically adjusted using AutoGLM in conjunction with historical public opinion data to ensure that the weight allocation conforms to the actual risk impact pattern. Risk level reasoning: Input the data of various dimensions into AutoGLM, guide the model to perform comprehensive reasoning through the special prompt word Prompt, output three risk levels of low, medium and high according to the preset risk threshold range, and generate a risk assessment report explaining the key influencing factors.

7. The keyword-driven online public opinion analysis method based on AutoGLM according to claim 1, characterized in that, S6, AutoGLM-optimized sentiment summarization and key findings extraction, including: Public opinion summary generation: Design the summary generation prompt word Prompt. AutoGLM is required to extract the core events, main viewpoints and dissemination status based on all relevant public opinion data, and generate a concise and comprehensive summary. The summary length can be adjusted according to user needs to ensure that no important information is omitted and there is no redundant content. Key findings extraction: AutoGLM is guided by the prompt word "Prompt" to extract key findings from public opinion data, including core points of contention, popular viewpoints, potential trends, and key stakeholders, highlighting the core value information that users care about; Deduplication and optimization: The summaries and key findings generated by AutoGLM are deduplicated by comparing semantic similarity to remove duplicate content, ensuring the uniqueness and validity of the information; at the same time, the prompt words are continuously optimized based on user feedback to improve the extraction quality.

8. The keyword-driven online public opinion analysis method based on AutoGLM according to claim 1, characterized in that, S8 provides a visual representation of structured data output and propagation paths, including: Structured output: The public opinion analysis results are output in a standardized format, including: keywords, data collection time range, public opinion source statistics, sentiment analysis results, risk level and assessment basis, public opinion summary, key findings, and support for export in Excel and PDF formats; among them, the public opinion source statistics include: initial source, main dissemination channels, source clustering results, and sentiment analysis results include: positive / neutral / negative percentage and sentiment intensity distribution. Visualization: Construct a visual interface to display the sentiment ratio, source distribution, and risk level composition through charts, and to show the path of public opinion dissemination through graphs, presenting the analysis results intuitively and making it easy for users to understand and use; charts include: pie charts, bar charts, or line charts.

9. A keyword-driven online public opinion analysis system based on AutoGLM, characterized in that, The system includes: The data acquisition module is based on keyword-driven multi-source data acquisition from the entire network; The preprocessing module performs cleaning, deduplication, and standardization preprocessing on the collected data; The public opinion source tracing and clustering module is based on AutoGLM semantic association for public opinion source tracing and clustering; The sentiment bias fine-grained judgment module uses the contextualized prompt word "Prompt" to guide the fine-grained judgment of sentiment bias. Risk level assessment module: Risk level assessment with dynamic weights of multi-dimensional indicators; The public opinion summary and key discovery extraction module uses AutoGLM to optimize public opinion summarization and key discovery extraction. The semantic deduplication and logical consistency verification module performs semantic deduplication and logical consistency verification on the extracted results. The visualization module provides a visual representation of the structured data output and propagation path.

Images