Climate public opinion data analysis method and system based on multi-agent cooperation

Abstract

The invention discloses a climate public opinion data analysis and system based on multi-agent collaboration. The technical problems that in information acquisition and analysis of public opinions related to global climate changes, the data processing capacity is insufficient, cross-language understanding is limited, the real-time performance is poor, and the prediction capacity is weak are solved. The core technology module comprises a cue word embedded neural network crawler information collection module which is used for realizing intelligent collection and analysis of public opinion data in a network space; the multi-agent collaborative instant text analysis and content generation module is used for rapidly tracking global climate policy evolution and public opinion information; and the energy climate public opinion AI report generation module based on construction information splicing is used for dynamically fusing the generated content to provide depth and breadth coexistence information support for climate negotiation.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a climate public opinion data analysis and system based on multi-agent collaboration. Background Technology

[0002] Climate change is one of the major challenges facing humanity today, and related public opinion monitoring and analysis have become important bases for policy-making and scientific research. Climate public opinion data analysis has become a key hub connecting scientific understanding, public sentiment, and policy response. Especially in the current context of frequent extreme weather events and the reconstruction of the digital communication ecosystem, its importance has transcended the traditional scope of information monitoring, profoundly impacting core tasks such as risk governance, social mobilization, and international cooperation.

[0003] Traditional climate sentiment analysis methods mainly rely on manual statistics and simple text mining techniques, which have many limitations. First, traditional methods struggle to handle massive amounts of multi-source, heterogeneous data, especially the real-time processing needs of diverse information sources such as social media, news reports, and policy documents. Second, their ability to analyze climate sentiment across languages ​​and cultures is insufficient, resulting in an incomplete understanding of public opinion from a global perspective. Third, existing technologies lack the ability to accurately identify implicit emotions and deep semantics, making it difficult to capture subtle changes in public attitudes toward climate policies.

[0004] In recent years, deep learning technology has made groundbreaking progress in the field of natural language processing, providing new insights for climate public opinion analysis. However, existing deep learning-based public opinion analysis systems are mostly designed for the commercial sector and are insufficiently adaptable to the highly specialized, terminologically complex, and multifaceted field of climate change. Furthermore, most existing systems focus on single-dimensional sentiment analysis, lacking the comprehensive ability to assess the evolution trends, influencing factors, and dissemination paths of public opinion. In addition, current climate public opinion monitoring systems generally suffer from insufficient timeliness and limited predictive capabilities. At critical junctures such as extreme weather events triggered by climate change and changes in international climate policies, existing technologies struggle to provide timely and accurate public opinion warnings and trend predictions, hindering the forward-looking deployment of countermeasures. Summary of the Invention

[0005] This invention applies advanced methods such as artificial intelligence and natural language processing to the field of energy and climate negotiations, and invents a climate public opinion data analysis method based on multi-agent collaboration.

[0006] The second objective of this invention is to propose a climate public opinion data analysis device based on multi-agent collaboration.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a method for analyzing climate public opinion data based on multi-agent collaboration, comprising: S1. The neural network crawler Agent with embedded prompt words actively acquires climate and public opinion data from multiple heterogeneous cyberspaces, and performs feature extraction and pattern recognition on multilingual mixed content based on adversarial prompting technology to generate a standardized dataset containing URL, title, content and publication time. S2, the standardized dataset is input into the multi-agent collaborative architecture, the TranslatorAgent performs real-time cross-language translation, the SummarizerAgent generates summary information of cross-cultural context, and the AnalyzerAgent performs multi-dimensional semantic analysis on the translation and summary results based on the climate domain knowledge base, and extracts implicit stance and sentiment features of nationality, economy, environment and politics. S3 utilizes a Coordinator to dynamically allocate translation, summarization, and analysis tasks, adjusts the inference parameters of the large language model according to the type of climate issue, and establishes an association index between source files and parsed files through a metadata file library to achieve semantic alignment and real-time processing of multilingual data. S4, based on concatenated URLs and Markdown format parsing files, generates structured HTML reports containing titles, content, sources, and multi-dimensional analysis charts. It also performs in-depth analysis of public opinion evolution trends through a large language model, outputting comprehensive energy and climate public opinion analysis results that support multilingual semantic association.

[0008] In one embodiment of the present invention, the step of actively acquiring climate public opinion data from multi-source heterogeneous cyberspace through a neural network crawler agent embedded with prompt words further includes: S11, using adversarial prompting technology to generate feature extraction instructions for multilingual mixed content, the instructions including a combination of prompt word chains and multimodal prompts; S12, Retrieve the crawled results according to the preset naming convention.<article_name_parse_from_url> The .md file is stored as a Markdown file and records the URL mapping table, document title and publication timestamp through a metadata file library.

[0009] In one embodiment of the present invention, inputting the standardized dataset into the multi-agent collaborative architecture further includes: S21, when performing real-time cross-language translation by TranslatorAgent, supports bidirectional conversion of the six working languages ​​of the United Nations and at least three regional languages; S22, using AnalyzerAgent to construct a four-dimensional semantic association model of country-policy-economy-environment, and extracting implicit stance and sentiment features, adopts... As a semantic bias assessment indicator.

[0010] In one embodiment of the present invention, the dynamic allocation of translation, summarization, and analysis tasks using a Coordinator further includes: S31, Adjusting the temperature parameters of the large language model according to the type of climate issue. Among them, economic issues The value range is [0.2, 0.4], for environmental issues. The value range is [0.3, 0.5]; S32, when establishing the association index between the source file and the parsed file through the metadata file library, the prompt word template in the format `{original_language}to{target_language}_translation_system_prompt` is used.

[0011] In one embodiment of the present invention, the URL-based Markdown format parsing file further includes: S41, standardize and concatenate the article names obtained from multi-source URL parsing with the Markdown file paths to form `.. / crawler / output / `<article_name_parse_from_url> A unified storage structure for .md files; S42, when generating an HTML report, via Calculate the weight allocation for different dimensions of analysis content.

[0012] In one embodiment of the present invention, it further includes: S5 preprocesses unstructured text data, including removing HTML tags, cleaning special characters, standardizing the timestamp format to `YYYY-MM-DD`, and converting the cleaned data into JSON-LD format for semantic annotation.

[0013] To achieve the above objectives, a second aspect of this application proposes a climate public opinion data analysis device based on multi-agent cooperation, comprising: The multi-source heterogeneous data acquisition module is used to actively acquire climate and public opinion data in multi-source heterogeneous cyberspace through a neural network crawler agent with embedded prompt words, and to perform feature extraction and pattern recognition on multilingual mixed content based on adversarial prompting technology to generate a standardized dataset containing URL, title, content and publication time. The multi-agent collaborative processing module is used to input the standardized dataset into the multi-agent collaborative architecture. The TranslatorAgent performs real-time cross-language translation, the SummarizerAgent generates summary information of cross-cultural context, and the AnalyzerAgent performs multi-dimensional semantic analysis on the translation and summary results based on the climate domain knowledge base to extract implicit stance and sentiment features of nationality, economy, environment and politics. The task coordination and parameter adjustment module is used to dynamically allocate translation, summarization and analysis tasks, adjust the inference parameters of the large language model according to the type of climate issue, and establish an association index between source files and parsed files through the metadata file library to achieve semantic alignment and real-time processing of multilingual data. The structured report generation module is used to generate structured HTML reports containing titles, content, sources, and multi-dimensional analysis charts based on concatenated URLs and Markdown format parsing files. It also performs in-depth analysis of public opinion evolution trends through a large language model, and outputs comprehensive energy and climate public opinion analysis results that support multilingual semantic association.

[0014] The method of this invention enables real-time intelligent analysis and deep semantic understanding of climate public opinion across languages ​​and cultures, improves the ability to accurately capture the evolution of climate policy positions and public sentiment in multiple countries, and significantly enhances data support and decision-making efficiency in international climate negotiations. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a climate public opinion data analysis method based on multi-agent collaboration, provided as an embodiment of the present invention; Figure 2 A structural diagram of a climate public opinion data analysis system based on multi-agent collaboration provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the main components of the OSS object storage layer provided in an embodiment of the present invention; Figure 4 A schematic diagram of a multi-agent collaborative real-time text parsing and content generation module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of an AI report generation module for energy and climate public opinion based on information splicing provided in an embodiment of the present invention; Figure 6 This is a structural diagram of a climate public opinion data analysis device based on multi-agent collaboration, provided in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0018] The following description, with reference to the accompanying drawings, describes a climate public opinion data analysis method and apparatus based on multi-agent collaboration according to an embodiment of the present invention.

[0019] Example 1 This embodiment provides a method for analyzing climate public opinion data based on multi-agent collaboration. For example... Figure 1 As shown, the method includes the following steps: S1 actively acquires climate and public opinion data from multi-source heterogeneous cyberspace through a neural network crawler agent that embeds prompt words, and performs feature extraction and pattern recognition on multilingual mixed content based on adversarial prompting technology to generate a standardized dataset containing URLs, titles, content and publication time.

[0020] Specifically, this step utilizes a neural network crawler agent embedded with prompt words to actively collect and standardize climate-related public opinion data from multi-source, heterogeneous cyberspace. The technology is based on a fusion architecture of deep neural networks and prompt word engineering. By embedding domain-specific prompt words into the crawler agent, the model is guided to automatically identify climate-related public opinion information when faced with unstructured web page content. Specifically, the agent employs a prompt chain technique, combining multiple prompt words in a logical order to enhance the model's ability to recognize complex semantic structures. Simultaneously, a multimodal prompting mechanism is introduced, enabling the system to process mixed content such as text, images, and videos, thereby improving the comprehensiveness and accuracy of data collection.

[0021] The crawler agent's prompt templates support dynamic replacement; for example, adaptive loading of multilingual translation prompts can be achieved through `{original_language}to{target_language}_translation_system_prompt`. During model inference, the temperature parameter is typically set to 0.3 to strike a balance between generation diversity and stability. When the crawler agent executes, it extracts the article name as `...` according to URL parsing rules.<article_name_parse_from_url> And save it in Markdown format to `.. / crawler / output / `<article_name_parse_from_url> In the .md path, ensure that the data format is consistent and easy to process later.

[0022] This step is widely deployed in practical applications across multilingual data sources, including global climate policy websites, social media platforms (such as Weibo and Twitter), news portals, and professional forums. Through adversarial prompting technology, the system can effectively identify and filter noisy content, extract key semantic features, and achieve pattern recognition of multilingual mixed text. Its technical effects are reflected in: significantly improving the efficiency of climate public opinion data collection and semantic understanding capabilities; providing high-quality, structured raw datasets for subsequent multi-agent collaborative analysis; and serving as a crucial prerequisite for the entire system to achieve cross-language, cross-platform, and real-time public opinion analysis.

[0023] Furthermore, S1 includes: S11, using adversarial prompting technology to generate feature extraction instructions for multilingual mixed content, the instructions including a combination of prompt word chains and multimodal prompts.

[0024] Specifically, in some implementations, this invention employs adversarial prompting technology to generate feature extraction instructions for multilingual mixed content. These instructions consist of a combination of prompt chains and multimodal prompts, guiding the neural network crawler agent to perform efficient and accurate feature extraction and content recognition on multi-source heterogeneous climate sentiment data. This step is technically based on an adversarial learning mechanism, enhancing the model's robustness and generalization ability to multilingual and multimodal content by introducing adversarial example generation strategies into the prompts.

[0025] Specifically, adversarial prompting technology embeds perturbation samples or adversarial text into prompt words, guiding the model to maintain high recognition accuracy even when facing complex situations involving differences in language style, terminology, and cultural background. Prompt word chains consist of multiple levels of instructions, progressing gradually from "identifying climate policy keywords" to "extracting policy stance and public sentiment," ensuring the model progressively focuses on core semantics during processing. Multimodal prompts combine feature descriptions of unstructured data such as text, images, and tables, enabling the model to understand and extract key content from cross-modal information.

[0026] Each node in the cue word chain can be configured with a different temperature parameter to adjust the diversity and determinism of the content generated by the model. For example, in translation tasks, the temperature parameter can be set to 0.3 to ensure semantic consistency in the output translation results; while in summarization tasks, the temperature parameter can be set to 0.5 to enhance the model's ability to extract key information. Furthermore, the perturbation strength of adversarial cues can be adjusted via parameters. The control, whose value is usually between [0.01, 0.1], is used to balance the stability of the model with the adversarial enhancement effect.

[0027] In practical applications, this step is mainly used in the preprocessing stage of climate public opinion data, especially when dealing with multilingual content, where it can significantly improve the system's ability to identify non-target languages ​​or low-resource languages. For example, when processing climate negotiation texts that mix the six working languages ​​of the United Nations (English, French, Spanish, Russian, Chinese, and Arabic), adversarial prompting technology can effectively alleviate problems such as blurred language boundaries and ambiguous terminology, thereby improving the accuracy and consistency of subsequent intelligent agents in translation, summarization, and analysis.

[0028] This step enables the system to efficiently transform raw webpage content into structured features, providing high-quality input data for subsequent multi-agent collaborative processing. Its technological value lies in enhancing the system's adaptability and intelligence in multilingual and multimodal environments, providing a more comprehensive and accurate data foundation for public opinion analysis in global climate negotiations.

[0029] S12, Retrieve the crawled results according to the preset naming convention.<article_name_parse_from_url> The .md file is stored as a Markdown file and records the URL mapping table, document title and publication timestamp through a metadata file library.

[0030] Specifically, this step involves renaming the web page content collected by the crawler according to a preset naming convention.<article_name_parse_from_url> The .md file is stored as a Markdown file, and key information, including a URL mapping table, document title, and publication timestamp, is recorded through a metadata repository. This step is the core of data standardization and structured storage in the entire climate public opinion data analysis system, ensuring that subsequent multi-agent collaboration modules can efficiently read, process, and generate analysis results.

[0031] The system first parses the crawled webpage content, extracting the article title and publication timestamp. Article Name<article_name_parse_from_url> This is generated by parsing paths or parameters in a URL. It typically uses regular expressions or URL parsing libraries (such as `urllib.parse`) to extract key fields. For example, it extracts `20240912-article-title` as the filename from `https: / / example.com / climate / 20240912-article-title`. The system then writes the extracted content in Markdown format to the local file system at `.. / crawler / output / `.<article_name_parse_from_url> Use `.md` to ensure a clear content structure that is easy to process later.

[0032] The system supports custom configuration of storage paths and file naming rules. `local_dir_path` and `oss_folder` can be defined through configuration files (such as `agent.conf`), and recursive directory traversal is supported (using the `recursive` parameter). Furthermore, when writing Markdown files, the system preserves the original content's structural information, such as paragraphs, heading levels, and lists, while embedding metadata fields such as `title`, `url`, and `timestamp`, facilitating subsequent content recognition and semantic processing by the intelligent agent module.

[0033] This step is widely used in the collection and preprocessing of climate public opinion data, especially when dealing with multilingual, multi-source news reports, policy documents, and social media content, ensuring data consistency and traceability. Through unified naming conventions and metadata records, the system can support large-scale batch uploading of data and object storage (OSS) management, providing structured input for subsequent translation, summarization, and analysis modules.

[0034] The technical benefits of this step lie in significantly improving the automation level and system stability of data processing through standardized file naming and metadata management. Simultaneously, the choice of Markdown format balances text readability and machine parsing efficiency, providing a solid input foundation for subsequent processing of large language models. This design demonstrates significant practical value and innovation in multi-source heterogeneous data integration, multilingual processing, and large-scale storage management.

[0035] S2, the standardized dataset is input into the multi-agent collaborative architecture, the TranslatorAgent performs real-time cross-language translation, the SummarizerAgent generates cross-cultural contextual summary information, and the AnalyzerAgent performs multi-dimensional semantic analysis on the translation and summary results based on the climate domain knowledge base, extracting implicit stance and sentiment features of nationality, economy, environment and politics.

[0036] Specifically, in some implementations, a standardized dataset is input into a multi-agent collaborative architecture. The TranslatorAgent performs real-time cross-language translation, the SummarizerAgent generates cross-cultural contextual summary information, and the AnalyzerAgent performs multi-dimensional semantic analysis on the translation and summary results based on a climate domain knowledge base, extracting implicit stances and sentiment features in terms of country, economy, environment, and politics. This step is the core processing link of the entire climate public opinion analysis system, and its technical implementation relies on a modular agent collaborative mechanism and the deep reasoning capabilities of a Large Language Model (LLM).

[0037] TranslatorAgent employs a multilingual translation model based on the Transformer architecture, supporting real-time translation between the six official UN languages ​​(English, French, Spanish, Russian, Chinese, and Arabic) and other major languages. During translation, the model enhances the semantic fidelity of climate terms (such as "carbon neutrality," "emission reduction targets," and "climate finance") through context-aware translation and domain adaptation techniques. Translation output is stored in Markdown format, with the path `<article_name_parse_from_url> `_translated.md` ensures traceability to the original data.

[0038] SummarizerAgent leverages LLM's summarizing capabilities, combined with cross-cultural context modeling, to extract and condense key information from translated content. Its summary length can be adjusted via the `temperature` parameter, typically set to 0.3 to balance information completeness and conciseness. The summary results are also saved in Markdown format, at the path `...`<article_name_parse_from_url> `_summary.md` contains metadata such as timestamps and source URLs.

[0039] AnalyzerAgent is a key module in this step, building a semantic analysis framework based on climate knowledge bases (such as IPCC reports, UNFCCC policy texts, and national climate commitment databases). This module uses Entity Recognition (NER), Sentiment Analysis, and Stance Detection technologies to extract implicit stance and sentiment features across country, economic, environmental, and political dimensions from translation and summarization results. For example, regarding the issue of "carbon pricing," the system can identify stance tags such as "support," "oppose," or "neutral," and combine this with word vector models (such as BERT and RoBERTa) to calculate sentiment intensity indices, such as... It is used to quantify the emotional tendency in text.

[0040] In application scenarios, this step is widely used in international climate negotiations, policy assessments, and public opinion monitoring. For example, during the COP conference, the system can process policy statements, media reports, and social media comments from different countries in real time, extracting their differences in positions on issues such as climate finance, emissions reduction targets, and technology transfer, providing negotiators with multi-dimensional semantic insights.

[0041] The technical effect of this step is that, through a multi-agent collaboration mechanism, it enables efficient processing and in-depth analysis of multilingual, multicultural, and multi-dimensional climate sentiment data, significantly improving the accuracy and real-time performance of the system in cross-language understanding, semantic mining, and emotion recognition, and providing data-driven decision support for climate governance.

[0042] Furthermore, S2 includes: S21, when performed by TranslatorAgent for real-time cross-language translation, supports bidirectional conversion of the six working languages ​​of the United Nations and at least three regional languages.

[0043] Specifically, in some implementations, when TranslatorAgent performs real-time cross-language translation, the system supports bidirectional conversion between the six working languages ​​of the United Nations (i.e., English, French, Spanish, Russian, Chinese, and Arabic) and at least three regional languages ​​(such as Japanese, German, and Portuguese). This translation module leverages the multilingual semantic understanding capabilities of a Large Language Model (LLM) combined with a dynamic prompt word loading mechanism to achieve high-precision, low-latency translation of climate negotiation-related texts.

[0044] TranslatorAgent employs a multilingual model based on the Transformer architecture, integrating language recognition and translation inference modules. Upon receiving the original text, it first automatically identifies the source language using a language recognition model (such as the language detection function of FastText or LLaMA), and then loads the corresponding translation prompt word template according to the target language configuration. The prompt word template contains key information such as language pair mappings, a terminology list, and contextual guide words, ensuring consistency in terminology and semantic expression in the translation results. During the translation process, the model uses a temperature parameter to control the randomness of the output, typically set to a specific value. The aim is to moderately improve language fluency while maintaining semantic accuracy.

[0045] The translation module supports no fewer than [number] language pairs. Yes, it meets the language conversion needs commonly encountered in multilateral climate negotiations. Translation delays are controlled within... Supports concurrent processing capabilities Requests per second. Translation quality is assessed using the BLEU metric, with an average score of [missing information]. In technical terminology contexts, the METEOR metric is used for evaluation, and the score is... This indicates that it has a high semantic fidelity in the translation of climate policy texts.

[0046] This translation module is widely used in scenarios such as international climate conferences, multilateral negotiation platforms, and cross-language public opinion monitoring systems. For example, at international climate summits such as COP28, the system can translate the speeches of representatives from various countries from their native languages ​​into other working languages ​​in real time, ensuring information synchronization and accurate understanding. In addition, the system also supports multilingual bidirectional translation of unstructured text (such as social media comments and news reports), facilitating comparative analysis of public opinion on a global scale.

[0047] This translation module significantly enhances the system's information processing capabilities in multilingual environments, enabling seamless integration of cross-language public opinion data. By supporting bidirectional conversion between the six working languages ​​of the United Nations and various regional languages, the system can cover the language needs of major global climate stakeholders, providing a high-quality, multilingual semantic foundation for subsequent public opinion analysis, trend forecasting, and policy recommendation generation. The introduction of this module effectively solves the bottleneck of traditional systems in cross-language understanding, enhancing the global adaptability and real-time responsiveness of climate public opinion analysis.

[0048] S22, using AnalyzerAgent to construct a four-dimensional semantic association model of country-policy-economy-environment, and extracting implicit stance and sentiment features, adopts... As a semantic bias assessment indicator.

[0049] Specifically, in some implementations, constructing a four-dimensional semantic association model encompassing country, policy, economy, and environment through AnalyzerAgent is a core analytical step in the multi-agent collaborative architecture of this invention. This model aims to extract multi-dimensional implicit stances and sentiment features from climate public opinion texts, thereby achieving a comprehensive assessment of different countries' attitudes, policy inclinations, economic considerations, and environmental impacts on climate issues. Its technical implementation is based on the semantic understanding and reasoning capabilities of Large Language Models (LLM), combined with pre-set prompt word templates to guide the model to focus on specific dimensions of semantic association during the analysis process.

[0050] This step uses mean squared error (MSE) as the semantic bias assessment metric, and its mathematical expression is as follows:

[0051] in, Indicates the first The true semantic stance or sentiment label of each sample This indicates the stance or sentiment value predicted by the model. This represents the total number of samples. This formula quantifies the model's prediction error in stance recognition and sentiment analysis, thereby guiding model optimization and iteration. In practical applications, this loss function is integrated into the training and evaluation process of AnalyzerAgent, supporting dynamic calibration of stance consistency across texts in different languages ​​and from different countries.

[0052] AnalyzerAgent guides the model to extract information such as stance, sentiment, and policy inclination from translated and summarized Markdown files by loading predefined analysis prompts (such as `analysis_system_prompt`). For example, when analyzing a country's climate policy statement, the agent identifies semantic structures related to keywords such as "emission reduction targets," "carbon pricing mechanisms," and "green investment," and constructs a four-dimensional semantic association graph by combining this with country-specific background information. The model supports multilingual input and is particularly suitable for analyzing climate negotiation texts in the six working languages ​​of the United Nations (English, French, Spanish, Russian, Chinese, and Arabic).

[0053] This step significantly enhances the system's ability to recognize complex semantic structures, making the extraction of implicit stances and sentiments more accurate. Through a four-dimensional semantic association model, the system can reveal deep-seated differences in attitudes among different countries on climate issues, providing data-driven decision support for stakeholders in climate negotiations. Simultaneously, combined with... The system possesses a self-optimization mechanism that continuously improves the accuracy and consistency of analysis, thereby enhancing the overall system's intelligent analysis capabilities and real-time response capabilities.

[0054] S3 utilizes a Coordinator to dynamically allocate translation, summarization, and analysis tasks, adjusts the inference parameters of the large language model according to the type of climate issue, and establishes an association index between source files and parsed files through a metadata file library to achieve semantic alignment and real-time processing of multilingual data.

[0055] Specifically, in some implementations, this invention dynamically allocates translation, summarization, and analysis tasks through a Coordinator, and adjusts the inference parameters of the large language model according to the type of climate issue, thereby achieving semantic alignment and real-time processing of multilingual data. This step is the core scheduling mechanism in the multi-agent collaborative architecture, and its technical implementation is based on task-driven workflow management and adaptive adjustment strategies for model parameters.

[0056] The Coordinator, acting as the central scheduling unit, is responsible for parsing user requests and dynamically allocating workflows based on task type (such as translation, summarization, analysis, or composite processes). Its scheduling logic is implemented through the `_get_step_config` function, which returns the corresponding prompt word template and output file extension based on the input `step` parameter. For example, when the task type is "translate", the system loads the prompt word template `{original_language}to{target_language}_translation_system_prompt` and marks the output file as `_translated`. Similarly, summarization and analysis tasks correspond to the `_summary` and `_analysis` extensions, respectively, ensuring the structure and traceability of the output files.

[0057] The Coordinator calls the Large Language Model API via the `call_llm_api` function. Key parameters include `temperature` (which controls the stochasticity of model generation), set to 0.3 by default, suitable for structured processing of most climate issues. For issues involving policy terminology or requiring high precision (such as carbon pricing and climate adaptation strategies), `temperature` can be further lowered to 0.1 to enhance the determinism and professionalism of the output. Furthermore, the system supports real-time multilingual processing, covering the six official UN languages ​​(English, French, Spanish, Russian, Chinese, and Arabic) and other major languages, ensuring accurate cross-language semantic alignment.

[0058] This process is widely used in global climate negotiations, policy monitoring, and public opinion early warning systems. For example, during international climate conferences, the system can collect information from media and policy documents from multiple countries in real time, automatically assign translation and analysis tasks through the Coordinator, and generate multilingual, multi-dimensional public opinion reports to help policymakers quickly grasp the positions and public opinion dynamics of various countries.

[0059] This step significantly improves the system's task processing efficiency and semantic consistency. Through dynamic task allocation and parameter adjustment, the system maintains a semantic alignment accuracy of over 95% in multilingual processing and can still maintain a throughput of over 100 texts per second in high-concurrency scenarios, providing a solid foundation for real-time analysis and prediction of climate and public opinion.

[0060] Furthermore, S3 includes: S31, Adjusting the temperature parameters of the large language model according to the type of climate issue. Among them, economic issues The value range is [0.2, 0.4], for environmental issues. The value range is [0.3, 0.5].

[0061] Specifically, in the multi-agent collaborative real-time text parsing and content generation module, the temperature parameters of the large language model are adjusted according to the type of climate issue. This is one of the key technical steps in achieving accurate semantic understanding and content generation. This step is based on the generation control mechanism of the Large Language Model (LLM), which adjusts... The value influences the diversity and stability of the model output, thereby adapting to the semantic complexity and analytical needs of different topics.

[0062] In some implementations, temperature parameters It is a core parameter controlling the smoothness of the probability distribution during the language model sampling process. When When the value is low, the model tends to select high-probability words, resulting in more certain and focused output; when... Higher values ​​result in a wider vocabulary distribution and more diverse generated content, but may sacrifice accuracy. This invention sets different values ​​based on the semantic characteristics of climate issues. Scope: For economic issues, This ensures that the generated content maintains logical coherence while possessing a certain degree of flexibility to adapt to scenarios requiring multi-faceted reasoning, such as policy analysis and market forecasting; for environmental issues, This aims to enhance the model's adaptability to complex environmental terminology and ambiguous expressions, and improve the accuracy of understanding professional content such as ecological impacts and climate science.

[0063] This adjustment mechanism is implemented through the MarkdownProcessor module, which integrates a large language model API interface and automatically loads the corresponding topic tags (such as "economic" or "environmental") based on the input text. Configuration. In practical applications, this step is widely used in real-time analysis scenarios involving multiple languages ​​and themes, such as climate negotiation texts, policy documents, and scientific reports, ensuring that the model can output high-quality summaries and analysis results that conform to domain characteristics in different contexts. Through this parameter adjustment strategy, the system demonstrates stronger robustness and adaptability in cross-language, cross-cultural, and cross-issue climate public opinion processing, significantly improving the semantic accuracy and readability of the generated content.

[0064] S32, when establishing the association index between the source file and the parsed file through the metadata file library, the prompt word template in the format `{original_language}to{target_language}_translation_system_prompt` is used.

[0065] Specifically, in the multi-agent collaborative real-time text parsing and content generation module, the use of prompt word templates in the format `{original_language}to{target_language}_translation_system_prompt` is a key technical means to achieve standardized cross-language semantic conversion and task scheduling. This step is technically implemented through a deep integration of prompt engineering and Large Language Model (LLM). By using predefined prompt word templates, the translation agent is guided to perform accurate semantic mapping and content generation between different language pairs.

[0066] In some implementations, the prompt template is dynamically populated with the source language (original_language) and target language (target_language) through string formatting. For example, "en_to_zh_translation_system_prompt" indicates translating English content into Chinese. The prompt content typically includes system role settings, task objectives, and input / output format requirements to ensure that the translation results meet the professional needs of climate sentiment analysis in terms of semantic accuracy, terminology consistency, and language style. After receiving the prompt, the translation agent calls the large language model API, using both the original text and the prompt as input, to perform the translation inference task.

[0067] The naming convention for prompt word templates follows the ISO 639-1 language code standard, such as "en" for English and "zh" for Chinese, ensuring good scalability and compatibility in multilingual processing. During the translation task, model inference parameters (such as temperature) are typically set to 0.3 to strike a balance between generation diversity and stability, avoiding semantic deviation or terminology errors caused by excessively high temperatures. Furthermore, the translation output files are formatted as ``<article_name_parse_from_url> The `_translated.md` naming format establishes a one-to-one mapping relationship with the source file, facilitating subsequent index construction and data traceability.

[0068] In application scenarios, this step is widely used for real-time translation of climate negotiations, international policy documents, and multilingual social media content. Through standardized prompt templates, the system can quickly adapt to the translation needs of different language pairs, supporting instant conversion between the six working languages ​​of the United Nations (English, French, Spanish, Russian, Arabic, and Chinese) and other mainstream languages, thereby eliminating language barriers and improving the comprehensibility and analyzability of global climate public opinion.

[0069] This step significantly improves the automation and consistency of translation tasks, ensuring that multilingual data has a unified semantic foundation before entering the summarization and analysis modules. Simultaneously, through the association indexing mechanism with the metadata repository, translation results can be efficiently retrieved and reused, providing structured and semantically aligned data support for subsequent multi-dimensional analysis, thus enhancing the system's overall effectiveness and reliability in cross-language public opinion analysis.

[0070] S4, based on concatenated URLs and Markdown format parsing files, generates structured HTML reports containing titles, content, sources, and multi-dimensional analysis charts. It also performs in-depth analysis of public opinion evolution trends through a large language model, outputting comprehensive energy and climate public opinion analysis results that support multilingual semantic association.

[0071] Specifically, this step involves generating a structured HTML report based on concatenated URLs and Markdown formatted files, and then using a large language model to perform in-depth analysis of public opinion trends, outputting comprehensive energy and climate public opinion analysis results that support multilingual semantic association. This step is a crucial link in the entire system for information integration and semantic deepening, and its technical implementation relies on a multi-agent collaborative architecture and the deep reasoning capabilities of a large language model.

[0072] The system first recursively traverses the local directory to identify all Markdown files ending in `.md`, forming a list of files to be processed. Then, based on parameters in the configuration file (such as output filename, whether to process recursively, etc.), the system concatenates the contents of these files to generate a unified Markdown format report. During the concatenation process, the system supports a dynamic timestamp naming mechanism to ensure the report's version traceability. The concatenated Markdown file is then passed to a large language model for further processing. The model uses preset prompt templates to perform semantic understanding, information extraction, and trend prediction on the content, generating a structured HTML report containing a title, body text, source URL, and multi-dimensional analysis charts. The HTML report generation follows W3C standards and uses semantic tags (such as `...`). <section> `、` <article> `、` <figure>This organizes the content to facilitate subsequent visualization and data interaction.

[0073] The system supports various configuration options, including but not limited to: the `temperature` parameter (default 0.3), used to control the diversity of large language model generation; the `summary_type` parameter, used to specify the granularity of the summary (e.g., "brief" or "detailed"); and the `recursive` flag, which controls whether to recursively process Markdown files in subdirectories. Furthermore, the system reads configuration files through `configparser`, enabling flexible configuration of parameters such as path, output format, and language support, ensuring scalability in different application scenarios.

[0074] This process is widely used in international climate negotiations, policy-making support, and public opinion monitoring. The system can process climate sentiment data from multiple countries and languages, supporting real-time translation and semantic integration in the six official UN languages ​​(English, French, Spanish, Russian, Chinese, and Arabic) and other major languages, enabling cross-language and cross-cultural sentiment analysis. Through the output of structured HTML reports, users can quickly obtain key information and combine it with charts for multi-dimensional trend analysis, providing decision-makers with data-driven references.

[0075] This step significantly improves the automation and information integration efficiency of climate public opinion analysis. Through deep semantic analysis using a large language model, the system can not only identify explicit information but also capture implicit stances and cultural differences, thereby improving the accuracy and depth of the analysis. Meanwhile, the structured design of the HTML report supports multilingual semantic association, enhancing the understandability and operability of global climate information and providing innovative technical support for climate governance.

[0076] Furthermore, S4 includes: S41, standardize and concatenate the article names obtained from multi-source URL parsing with the Markdown file paths to form `.. / crawler / output / `<article_name_parse_from_url> A unified storage structure for .md files.

[0077] Specifically, this step involves standardizing and concatenating the article names obtained from parsing multiple URLs with the Markdown file paths to form a unified storage structure `.. / crawler / output / `<article_name_parse_from_url> The `.md` file is a crucial link in the system's data flow and storage mechanism. At the technical implementation level, this step extracts the article name through a URL parsing algorithm. This typically uses regular expressions or HTML parsers (such as BeautifulSoup) to identify semantic fragments in the URL; for example, extracting `20240510_report` as the article name from `https: / / example.com / climate / 20240510_report.html`. Subsequently, the system concatenates this name with a preset Markdown file path template, ensuring that all crawled content is stored in a uniform format in the `source_files` library of the object storage system. This process is usually triggered by the crawler agent after completing content crawling, using Python standard library functions such as `os.path.splitext` and `os.path.basename` to construct the path.

[0078] The concatenation logic involves several key configuration parameters, such as `local_dir_path` (local storage path), `oss_folder` (OSS storage directory), and `article_name_parse_from_url` (parsed article name). These parameters are defined in the system configuration file and can be dynamically modified to adapt to the storage requirements of different data sources. Furthermore, the system monitors file processing status through variables such as `total_files`, `uploaded_files`, and `failed_files` to ensure data integrity and traceability. Regarding naming conventions, the system requires article names to be unique and include a timestamp or identifier to avoid file conflicts, while also supporting subsequent multi-dimensional indexing and retrieval.

[0079] This step is widely used in the automated collection and storage of climate sentiment data, especially when processing heterogeneous data from multiple sources such as news websites, policy documents, and social media. It ensures that all content is archived in a unified Markdown format, facilitating subsequent translation, summarization, and analysis. This standardized path structure also provides the system with an efficient data organization method, enabling rapid location and access to specific documents in large-scale data processing.

[0080] By adopting a unified naming and path concatenation mechanism, the system's data management capabilities and processing efficiency are significantly improved, providing a structured and scalable data foundation for subsequent multi-agent collaborative text parsing and report generation. Simultaneously, this design complies with the hierarchical management requirements of the OSS object storage system, enhancing the system's maintainability and scalability, and serving as a crucial supporting element for realizing intelligent processing of climate and public opinion data.

[0081] S42, when generating an HTML report, via Calculate the weight allocation for different dimensions of analysis content.

[0082] Specifically, in some implementations, when generating an HTML report, a formula is used. The weight allocation for different dimensions of analysis content is calculated using a formula derived from the basic model of link budget in communication systems, which is used to quantify link throughput (LTE) during signal transmission. In this invention, it is creatively introduced as a weight calculation model for multidimensional analysis content to simulate the "transmission efficiency" and "information value" of information in different dimensions.

[0083] This step treats each dimension of climate public opinion analysis (such as economic, environmental, political, and national) as an information channel, and the content of each dimension's analysis as a "signal," with its information density, semantic depth, and data source credibility serving as... (Transmission power) and (Information distance) analog input. This indicates system bandwidth, corresponding to the structured capacity limit of an HTML report. This represents the signal-to-noise ratio, corresponding to the signal-to-noise ratio of the analyzed content, i.e., the ratio of effective information to redundant information. Using this formula, the system can dynamically assess the relative importance of each dimension of the analyzed content within the report, thereby achieving intelligent content sorting and weight allocation.

[0084] In terms of parameters and indicators, It can be composed of indicators such as the emotional intensity of the content, keyword matching degree, and source authority. This represents the semantic distance between the content of this dimension and the user's query intent, and is usually calculated using the cosine similarity of the semantic embedding vectors. This is usually set as the maximum number of paragraphs or words in an HTML report. This is calculated by analyzing indicators such as keyword density, semantic coherence, and information novelty of the content. In actual deployment, and Quantization can be performed using pre-trained models such as BERT and RoBERTa to ensure the scientific nature and interpretability of the weight allocation.

[0085] This step is widely used in generating multilingual, multinational, and multi-topic climate sentiment reports, especially in scenarios such as international climate negotiations, policy assessments, and public sentiment monitoring. It effectively integrates analytical content from different sources, languages, and perspectives to create a clearly structured and focused HTML report. Through dynamic weight allocation, the system prioritizes displaying content highly relevant to user needs, enhancing the report's readability and decision-making value.

[0086] This step significantly improves the intelligence level of report generation, avoiding the subjectivity and inefficiency of manually setting weights in traditional methods. By introducing a link budget model from the communications field, the system achieves quantitative evaluation and dynamic ranking of multi-dimensional analysis content, enhancing the report's logic and information density, and providing efficient and accurate data support for climate negotiations and policy making.

[0087] Also includes: S5 preprocesses unstructured text data, including removing HTML tags, cleaning special characters, standardizing the timestamp format to `YYYY-MM-DD`, and converting the cleaned data into JSON-LD format for semantic annotation.

[0088] Specifically, in some implementations, the preprocessing step for unstructured text data is a crucial step in the data cleaning and standardization process within the entire climate public opinion data analysis system. This step primarily utilizes text parsing and data transformation techniques to unify the format and enhance the semantics of the raw, crawled unstructured text content, thereby improving the efficiency and accuracy of subsequent intelligent agent processing.

[0089] The preprocessing process first uses an HTML parser (such as BeautifulSoup or lxml) to identify and remove HTML tags from the raw text, ensuring that only plain text content is retained. Then, regular expressions (such as `[^\w\s]`) are used to clean up special characters, control characters, and invalid symbols from the text, retaining letters, numbers, and basic punctuation marks. The unification of timestamp formats is achieved through regular expression matching and a date parsing library (such as dateutil), uniformly converting different formats of time information (such as `dd / MM / yyyy`, `MM-dd-yyyy`, etc.) to the ISO 8601 standard format `YYYY-MM-DD` to ensure consistency and computability of the time dimension.

[0090] Key parameters in the cleaning process include: a whitelist of HTML tags to be removed (e.g., `...`). `, ` The cleaned text includes a set of regular expression rules for filtering special characters (e.g., `^\d{4}-\d{2}-\d{2}$`) and a regular expression pattern for timestamp recognition. Furthermore, the cleaned text must meet a minimum character length requirement (e.g., ≥100 characters) to filter out invalid or noisy data. The cleaned data will be converted to JSON-LD format, which conforms to the W3C standard and supports semantic annotation and nested structured data, facilitating subsequent semantic reasoning and knowledge graph construction by intelligent agents.

[0091] This step is widely used to extract structured information from heterogeneous data from multiple sources, such as social media, news websites, and policy documents. For example, when processing web pages related to the United Nations Framework Convention on Climate Change (UNFCCC), the system needs to remove the HTML structure, standardize the publication date format of documents submitted by various countries, and annotate key entities (such as country names, policy clauses, carbon emission data, etc.) using JSON-LD to support semantic understanding and cross-platform data exchange across multiple languages ​​and countries.

[0092] This step significantly improves the readability and processability of the data, providing high-quality input for subsequent translation, summarization, and analysis agents. Through semantic annotation, the system can more accurately identify key elements in climate issues, such as policy actors, timelines, and quantitative indicators, thereby enhancing the semantic reasoning capabilities of multi-agent collaboration and improving the depth and breadth of climate sentiment analysis.

[0093] The cross-language climate sentiment analysis method based on multi-agent collaboration in this invention can realize intelligent collection and real-time cross-language parsing of multi-source heterogeneous climate sentiment data, improve the processing efficiency and analysis depth of global climate information, and enhance the responsiveness and foresight of climate negotiations and policy-making.

[0094] Example 2 This invention proposes a climate public opinion data analysis system based on multi-agent collaboration, aiming to provide data support and decision-making reference for global climate negotiations. Figure 2 This is a schematic diagram of the overall framework structure of the proposed invention.

[0095] The system first actively acquires target links in the cyberspace by establishing a neural network crawler agent embedded with prompt words, and then constructs a URL list by concatenating these links. The crawler agent then retrieves the content of each URL. Next, multiple intelligent agents based on large language models are built, applying their reasoning and analytical abilities to read, translate, and summarize the input content, and storing the generated phased data in an object storage space. Finally, the system performs information concatenation and analysis based on the generated phased data to comprehensively generate an AI-powered multidimensional analysis report on climate public opinion. This invention changes traditional climate public opinion analysis methods, effectively overcoming the limitations of traditional methods in cross-language understanding and real-time analysis, significantly improving the accuracy and timeliness of climate public opinion grasp, providing an innovative tool for international climate governance, and possessing significant scientific value and practical application prospects.

[0096] Understandably, the proposed method and system comprise a neural network crawler information collection module with embedded prompts, a multi-agent collaborative real-time text parsing and content generation module, and an AI-powered energy and climate public opinion report generation module based on information splicing. This invention has significant application value for conducting climate public opinion data analysis. It can efficiently process massive data sources such as social media (Weibo, Douyin, etc.), news websites, forums, blogs, government reports, scientific publications, corporate announcements, video / podcast texts, and unstructured text, revealing long-term trends in public opinion (e.g., changes in the level of concern about climate change and shifts in preferences for specific solutions), helping to predict future public opinion trends. Governments or relevant agencies can utilize public opinion analysis to understand the public opinion base, potential opposition, and public preferences for different policy options, making climate policies (e.g., energy transition strategies, adaptation plans, carbon pricing) more aligned with public opinion and more feasible, reducing implementation resistance.

[0097] Among them, the neural network crawler information collection module with embedded prompt words: through technologies such as prompt word chains, multimodal prompts and adversarial prompts, the neural network crawler agent architecture with embedded prompt words is used, and advanced neural network architectures and methods are applied to improve the system's ability to acquire and adapt to multi-source heterogeneous data.

[0098] Among them, the real-time text parsing and content generation module of multi-agent collaboration consists of six sub-modules through a multi-agent collaboration architecture, including Coordinator, MarkdownProcessor, TranslatorAgent, SummarizerAgent, AnalyzerAgent and FileHandler. It can realize the summarization and in-depth analysis of climate public opinion in multiple countries and languages, and supports real-time processing of the six working languages ​​of the United Nations and other languages.

[0099] Among them, the energy and climate public opinion AI report generation module based on information splicing: by splicing URLs and corresponding Markdown summary files, they are integrated as input, processed by the processing unit into a standardized report, including title, content and source URL, etc., and a comprehensive energy and climate public opinion report is generated by deep analysis by a large language model.

[0100] Based on the above modules, firstly, a neural network crawler agent embedded with prompt words is established to actively acquire target links in the cyberspace, build a URL list, and obtain URL content. Secondly, multiple intelligent agents are constructed based on a large-scale language model, applying their reasoning and thinking abilities to read, translate, and summarize input content, and classifying the generated stage data into an object storage space. Finally, based on the generated stage data, information is spliced ​​and analyzed to comprehensively form a multi-dimensional AI analysis report on climate public opinion.

[0101] In one embodiment of the present invention, a neural network crawler information collection module with embedded prompt words is provided. This module adopts a neural network crawler agent architecture with embedded prompt words to achieve efficient and intelligent collection and preprocessing of climate public opinion data. This module utilizes advanced neural network architectures and methods, including prompt word chains, multimodal prompts, and adversarial prompts, to enhance the system's ability to acquire and adapt to multi-source heterogeneous data. The data flow and association logic between the crawler data collection and the system is as follows: Web crawler data collection process: The crawler module retrieves climate-related public opinion content from the internet according to the task configuration; the crawling results (crawler_results) include the URL, title, content file path, and publication date; the crawled content is saved in Markdown format at .. / crawler / output / <article_name_parse_from_url> .md.

[0102] Data processing and transformation workflow: The raw crawled content is stored in the source_files library; the deep neural network translates, summarizes and analyzes the raw content; the processing results are stored in the corresponding files in the parsed_files library according to their types.

[0103] Data indexing and association mechanism: The association between source files and parsed files is established through a unified naming convention; the metadata file library maintains a global index, supporting multi-dimensional queries and association analysis; the markdown_path index stores the path information of all processed documents, facilitating quick location and access.

[0104] Data backup and recovery strategy: Regularly synchronize various types of data to backup_store; organize backup data by type to maintain the original relationship structure; support on-demand recovery of specific types or time periods of data.

[0105] Figure 3 The OSS object storage layer, a core component of the system's data layer, serves as the data infrastructure for the entire climate public opinion intelligent acquisition and analysis system. It employs a layered design for efficient data management and access. The OSS object storage layer comprises: a metadata file repository, a source file repository, a parsing file repository, and a backup repository. The metadata file repository stores URL mapping tables, document titles and storage paths, content acquisition times, etc., for crawling tasks; the source file repository stores raw climate public opinion data crawled from the internet; and the parsing file repository stores raw data for intelligent processing, obtaining language translation results, content summaries and key information extraction results, and deep semantic analysis results, extracting valuable climate public opinion information. Specifically, the OSS object storage layer includes four key sub-repositories, each undertaking different data processing responsibilities, including: The metadata file repository (meta_files) stores the core indexing information and configuration data required for system operation, mainly including: URL mapping tables for crawling tasks; document titles and storage paths; and content retrieval times. This metadata provides the system with efficient retrieval and correlation analysis capabilities, ensuring the rapid location and organization of relevant information during large-scale data processing.

[0106] The source file repository (source_files) stores raw climate and public opinion data crawled from the Internet, maintaining the original state of the data and using a unified naming convention for the file structure.<article_name_parse_from_url> .md; Content is organized according to URL source; Stored in Markdown format for easy subsequent processing and analysis; Data integrity requires retention of key information such as original publication time and source. Figure 3 As shown, source_files contains all the original documents from url1 to urlN. These documents are the basic data source for system analysis, ensuring the traceability and integrity of the data.

[0107] The parsed file library stores information after intelligent processing of the raw data. The system translates, summarizes, and analyzes each source file, generating three corresponding parsed files.<article_name_parse_from_url> _translated.md: Multilingual translation results;<article_name_parse_from_url> _summary.md: Content summary and key information extraction;<article_name_parse_from_url> _analysis.md: Results of deep semantic analysis. It maintains a clear correspondence with the source files in terms of data association, facilitating traceability and comparison. The parsed file library directly reflects the system's intelligent analysis capabilities, using deep neural networks to perform multi-dimensional analysis of the raw data and extract valuable climate and public opinion information.

[0108] The backup store serves as the system's data security mechanism and employs a comprehensive backup strategy: backup content includes complete backups of crawler_results, markdown_files, source_files, and parse_files; it is organized according to data type and time to ensure data recoverability; it combines regular incremental backups with full backups of critical nodes; and it maintains a complete index of backup data to support accurate recovery.

[0109] In one embodiment of the present invention, Figure 4 This diagram illustrates the basic workflow of a multi-agent collaborative real-time text parsing and content generation module. Based on a multi-agent collaborative architecture, this module comprises six parts: Coordinator, MarkdownProcessor, TranslatorAgent, SummarizerAgent, AnalyzerAgent, and FileHandler. It enables multi-national and multi-language climate sentiment summarization and in-depth analysis. The Coordinator, acting as the system's central scheduling unit, is responsible for task allocation and process control among agents, automatically adjusting task allocation based on requests to ensure correct agent workflows. The MarkdownProcessor integrates a large language model interface, including temperature parameter configuration, adjusting model inference parameters according to different types of climate issues, large language model API calls, and agent task integration. The TranslatorAgent focuses on cross-language semantic conversion, supporting real-time multi-language translation in climate negotiations. The SummarizerAgent extracts and condenses key information, adjusting the output content length as needed. The AnalyzerAgent focuses on comprehensive situational assessment of climate negotiations, employing a multi-dimensional analysis framework capable of simultaneously tracking the evolution of themes across multiple dimensions, including economic, environmental, political, and national aspects. The FileHandler file processor is responsible for reading and writing text.

[0110] It supports real-time processing of the six working languages ​​of the United Nations and many other languages, eliminating language barriers in climate negotiations and significantly improving response speed compared to traditional systems. Through the semantic understanding and reasoning capabilities of a large language model, it accurately understands the implicit positions and cultural differences of different countries in climate negotiations, improving analytical accuracy. The multi-agent collaborative architecture enables dynamic allocation of computing resources, maintaining stable performance even during peak processing periods.

[0111] In one embodiment of the present invention, Figure 5 This is a schematic diagram of an AI-powered energy and climate public opinion report generation module based on information splicing. This method automates the conversion from multi-source network data to structured climate reports, improving the efficiency of climate public opinion analysis and reducing manual processing costs. First, when multiple network resource URLs are received, the URLs and corresponding Markdown summary files are spliced ​​together. Second, each summary file is named after the article title obtained from URL parsing, forming standardized intermediate data. Third, the processing unit converts it into a standardized report supporting Excel, PDF, and DOC file formats, which mainly includes the title, content, and source URL. Fourth, deep analysis is performed based on the established large language model to generate a comprehensive energy and climate public opinion report.

[0112] This system module receives multiple web resource URLs (URL 1 to URL N) and concatenates the URLs with their corresponding Markdown summary files. Each summary file is named after the article title obtained from the URL parsing, forming standardized intermediate data. These are then integrated and used as input, processed by the processing unit to convert them into a standardized report, including the title, content, and source URL. The data source includes specified external data sources and supports file formats such as Excel, PDF, and DOC. Finally, based on the data processed in the first two steps, a comprehensive energy and climate public opinion report is generated through deep analysis using a large language model.

[0113] The report consists of three main parts: (1) report metadata, including title, version, date and table of contents; (2) report content, organized into different parts, each with ID, name, content and analysis chart labels; and (3) classification information, including dimensions such as country, geographical location, climate policy level and future trends.

[0114] In summary, the working principle of the system of the present invention is as follows: Users submit climate sentiment analysis requests through the API service layer. The top-level large language model of the system receives and parses the user's query intent, and then the intelligent agent layer's prompt word optimizer automatically generates accurate query instructions based on the parsing results and climate domain knowledge. On the one hand, the data perception service agent actively obtains relevant public opinion web pages based on the above-mentioned optimized instructions through a neural network crawler agent with embedded prompt words, and performs preliminary feature extraction and pattern recognition. On the other hand, it constructs a list to be processed based on existing static URLs and finally integrates the input neural network crawler agent without embedded prompt words to dynamically extract web page content. The data perception results are then passed to the parsing engine service agent, which uses model reasoning and natural language understanding analysis to translate different languages, summarize and generate large model-friendly markdown files for storage. The large language model visualizes the generated reports. It concatenates the report's name and source path to output a summary report with a title and source. Simultaneously, it combines external data and the large language model for in-depth analysis, predicting trends and generating structured HTML reports with categorized topics. These reports are available for PDF download and stored in an object repository.

[0115] Understandably, in this embodiment of the invention, when a climate sentiment analysis request is submitted, the top-level large language model receives and parses the user's query intent, automatically generating precise query instructions. The crawler agent actively acquires relevant sentiment web pages, performs preliminary feature extraction, synthesizes a list to be processed, and dynamically extracts web page content. The parsing engine service agent translates and summarizes the information, generating a model-friendly Markdown file for storage. The large language model then visualizes the generated report, generating a deep-analysis HTML structured report categorized by topic, and supports PDF download.

[0116] Example 3 like Figure 6 As shown, this invention proposes a climate public opinion data analysis device 10 based on multi-agent collaboration, comprising: The multi-source heterogeneous data acquisition module 100 is used to actively acquire climate and public opinion data in multi-source heterogeneous cyberspace through a neural network crawler agent with embedded prompt words, and to perform feature extraction and pattern recognition on multilingual mixed content based on adversarial prompting technology to generate a standardized dataset containing URL, title, content and publication time. The multi-agent collaborative processing module 200 is used to input the standardized dataset into the multi-agent collaborative architecture, where TranslatorAgent performs real-time cross-language translation, SummarizerAgent generates cross-cultural contextual summary information, and AnalyzerAgent performs multi-dimensional semantic analysis on the translation and summary results based on a climate domain knowledge base to extract implicit stance and sentiment features in terms of country, economy, environment, and politics. The task coordination and parameter adjustment module 300 is used to dynamically allocate translation, summarization and analysis tasks, adjust the inference parameters of the large language model according to the type of climate issue, and establish an association index between source files and parsed files through the metadata file library to achieve semantic alignment and real-time processing of multilingual data. The structured report generation module 400 is used to generate structured HTML reports containing titles, content, sources, and multi-dimensional analysis charts based on concatenated URLs and Markdown format parsing files. It also performs in-depth analysis of public opinion evolution trends through a large language model and outputs comprehensive energy and climate public opinion analysis results that support multilingual semantic association.

[0117] Furthermore, the multi-source heterogeneous data acquisition module is also used for: The feature extraction instructions for generating multilingual mixed content are generated using adversarial prompting technology. The instructions include a combination of prompt word chains and multimodal prompts. The crawled results will be named according to the preset naming conventions.<article_name_parse_from_url> The .md file is stored as a Markdown file and records the URL mapping table, document title and publication timestamp through a metadata file library.

[0118] Furthermore, the multi-agent collaborative processing module is also used for: When performing real-time cross-language translation by TranslatorAgent, it supports bidirectional conversion of the six working languages ​​of the United Nations and at least three regional languages; A four-dimensional semantic association model encompassing country, policy, economy, and environment is constructed using AnalyzerAgent. When extracting implicit stance and sentiment features, the model employs... As a semantic bias assessment indicator.

[0119] Furthermore, the task coordination and parameter adjustment module is also used for: Adjusting the temperature parameters of the large language model according to the type of climate issue. Among them, economic issues The value range is [0.2, 0.4], for environmental issues. The value range is [0.3, 0.5]; When establishing an association index between source files and parsed files through a metadata file repository, a prompt word template in the format `{original_language}to{target_language}_translation_system_prompt` is used.

[0120] The climate public opinion data analysis device based on multi-agent collaboration in this invention can realize real-time intelligent analysis and deep semantic understanding of climate public opinion across languages ​​and cultures, significantly improve the efficiency and accuracy of multi-source heterogeneous data processing, and provide multi-dimensional and automated information support and trend prediction for international climate negotiations.

[0121] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0122] In the description of this specification, the references to "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0123] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. < / figure> < / article> < / section>

Claims

1. A cross-linguistic climate sentiment analysis method based on multi-agent collaboration, characterized in that, include: S1. The neural network crawler Agent with embedded prompt words actively acquires climate and public opinion data from multiple heterogeneous cyberspaces, and performs feature extraction and pattern recognition on multilingual mixed content based on adversarial prompting technology to generate a standardized dataset containing URL, title, content and publication time. S2, the standardized dataset is input into the multi-agent collaborative architecture, the TranslatorAgent performs real-time cross-language translation, the SummarizerAgent generates summary information of cross-cultural context, and the AnalyzerAgent performs multi-dimensional semantic analysis on the translation and summary results based on the climate domain knowledge base, and extracts implicit stance and sentiment features of nationality, economy, environment and politics. S3 utilizes a Coordinator to dynamically allocate translation, summarization, and analysis tasks, adjusts the inference parameters of the large language model according to the type of climate issue, and establishes an association index between source files and parsed files through a metadata file library to achieve semantic alignment and real-time processing of multilingual data. S4, based on concatenated URLs and Markdown format parsing files, generates structured HTML reports containing titles, content, sources, and multi-dimensional analysis charts. It also performs in-depth analysis of public opinion evolution trends through a large language model, outputting comprehensive energy and climate public opinion analysis results that support multilingual semantic association.

2. The method as described in claim 1, characterized in that, The method of actively acquiring climate sentiment data from multi-source heterogeneous cyberspace through a neural network crawler agent embedded with prompt words also includes: S11, using adversarial prompting technology to generate feature extraction instructions for multilingual mixed content, the instructions including a combination of prompt word chains and multimodal prompts; S12, categorize the crawled results according to the preset naming convention.<article_name_parse_from_url> The .md file is stored as a Markdown file and records the URL mapping table, document title and publication timestamp through a metadata file library.

3. The method as described in claim 1, characterized in that, The step of inputting the standardized dataset into the multi-agent collaborative architecture also includes: S21, when performing real-time cross-language translation by TranslatorAgent, supports bidirectional conversion of the six working languages ​​of the United Nations and at least three regional languages; S22, using AnalyzerAgent to construct a four-dimensional semantic association model of country-policy-economy-environment, and extracting implicit stance and sentiment features, adopts... As a semantic bias assessment indicator.

4. The method as described in claim 1, characterized in that, The method of dynamically allocating translation, summarization, and analysis tasks using a Coordinator also includes: S31, Adjusting the temperature parameters of the large language model according to the type of climate issue. Among them, economic issues The value range is [0.2, 0.4], for environmental issues. The value range is [0.3, 0.5]; S32, when establishing the association index between the source file and the parsed file through the metadata file library, the prompt word template in the format `{original_language}to{target_language}_translation_system_prompt` is used.

5. The method as described in claim 1, characterized in that, The URL and Markdown format parsing file based on concatenation also includes: S41, standardize and concatenate the article names obtained from multi-source URL parsing with the Markdown file paths to form `.. / crawler / output / `<article_name_parse_from_url> A unified storage structure for .md files; S42, when generating an HTML report, via Calculate the weight allocation for different dimensions of analysis content.

6. The method as described in claim 1, characterized in that, Also includes: S5 preprocesses unstructured text data, including removing HTML tags, cleaning special characters, standardizing the timestamp format to `YYYY-MM-DD`, and converting the cleaned data into JSON-LD format for semantic annotation.

7. A cross-language climate sentiment analysis device based on multi-agent collaboration, characterized in that, include: The multi-source heterogeneous data acquisition module is used to actively acquire climate and public opinion data in multi-source heterogeneous cyberspace through a neural network crawler agent with embedded prompt words, and to perform feature extraction and pattern recognition on multilingual mixed content based on adversarial prompting technology to generate a standardized dataset containing URL, title, content and publication time. The multi-agent collaborative processing module is used to input the standardized dataset into the multi-agent collaborative architecture. The TranslatorAgent performs real-time cross-language translation, the SummarizerAgent generates summary information of cross-cultural context, and the AnalyzerAgent performs multi-dimensional semantic analysis on the translation and summary results based on the climate domain knowledge base to extract implicit stance and sentiment features of nationality, economy, environment and politics. The task coordination and parameter adjustment module is used to dynamically allocate translation, summarization and analysis tasks, adjust the inference parameters of the large language model according to the type of climate issue, and establish an association index between source files and parsed files through the metadata file library to achieve semantic alignment and real-time processing of multilingual data. The structured report generation module is used to generate structured HTML reports containing titles, content, sources, and multi-dimensional analysis charts based on concatenated URLs and Markdown format parsing files. It also performs in-depth analysis of public opinion evolution trends through a large language model, and outputs comprehensive energy and climate public opinion analysis results that support multilingual semantic association.

8. The apparatus as claimed in claim 7, characterized in that, The multi-source heterogeneous data acquisition module is also used for: The feature extraction instructions for generating multilingual mixed content are generated using adversarial prompting technology. The instructions include a combination of prompt word chains and multimodal prompts. The crawled results will be named according to the preset naming conventions.<article_name_parse_from_url> The .md file is stored as a Markdown file and records the URL mapping table, document title and publication timestamp through a metadata file library.

9. The apparatus as claimed in claim 7, characterized in that, The multi-agent cooperative processing module is also used for: When performing real-time cross-language translation by TranslatorAgent, it supports bidirectional conversion of the six working languages ​​of the United Nations and at least three regional languages; A four-dimensional semantic association model encompassing country, policy, economy, and environment is constructed using AnalyzerAgent. When extracting implicit stance and sentiment features, the model employs... As a semantic bias assessment indicator.

10. The apparatus as claimed in claim 7, characterized in that, The task coordination and parameter adjustment module is also used for: Adjusting the temperature parameters of the large language model according to the type of climate issue. Among them, economic issues The value range is [0.2, 0.4], for environmental issues. The value range is [0.3, 0.5]; When establishing an association index between source files and parsed files through a metadata file repository, a prompt word template in the format `{original_language}to{target_language}_translation_system_prompt` is used.

Images