Multilingual semantic analysis and decision support system for international climate negotiation scene

Abstract

The invention provides a multilingual semantic analysis and decision support system and method for an international climate negotiation scene. The system comprises a data processing module which is used for utilizing a natural language processing model to automatically generate structured data which is composed of multiple languages and corresponds to corpus data related to a target theme; the database construction module is used for carrying out vector database index construction of hierarchical indexing on the structured data according to an issue, a language and time, and generating a corpus vector database; the content generation module is used for responding to the received content generation instruction, converting the content generation instruction into a query vector, performing retrieval from the corpus vector database according to the query vector to obtain target corpus data corresponding to the content generation instruction, and generating the target corpus data according to a pre-trained language generation model; and performing language generation processing on the content generation indication and the target corpus data to obtain a content generation result associated with the content generation indication.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and natural language processing technology, and in particular to a multilingual semantic analysis and decision support system, method, computer equipment and storage medium for international climate negotiations. Background Technology

[0002] Against the backdrop of global climate change, international climate negotiations, such as the United Nations Climate Change Conference (COP) series of meetings, have become the core platform for coordinating emission reduction pathways, funding mechanisms, and technological routes. This process involves a large number of policy documents, news reports, and scientific data, characterized by multilingualism, a large volume of information, and frequent updates.

[0003] Currently, the text is typically processed manually, which is not only inefficient but also prone to missing key information, limiting decision-making capabilities. Therefore, a technology capable of rapidly processing corpora and providing decision-making support based on the processing results is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, this application provides a multilingual semantic analysis and decision support system, method, computer equipment, and storage medium for international climate negotiations.

[0005] Specifically, this application is implemented through the following technical solution:

[0006] In a first aspect, embodiments of this disclosure provide a multilingual semantic analysis and decision support system for international climate negotiations, including:

[0007] The data processing module is used to automatically generate structured data corresponding to corpus data in multiple languages ​​and related to the target topic using a natural language processing model; wherein, the structured data includes: topic tags, summaries, representative keywords, and corpus data indexes for multiple topics corresponding to the target topic;

[0008] The database construction module is used to construct a vector database index by hierarchically indexing the structured data according to topic, language, and time, and generate a corpus vector database.

[0009] The content generation module is configured to, in response to receiving a content generation instruction, convert the content generation instruction into a query vector, and retrieve target corpus data corresponding to the content generation instruction from the corpus vector database based on the query vector, and use a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a content generation result associated with the content generation instruction.

[0010] Optionally, the data processing module is specifically used for:

[0011] Based on the semantic vectors of each corpus data, the corpus data is divided into multiple corpus groups; different corpus groups correspond to different topics, and each corpus group includes at least one piece of corpus data associated with the corresponding topic;

[0012] For at least a portion of the corpus data in each corpus group, the natural language processing model is used to perform semantic summarization processing to generate a summary of the topic corresponding to each corpus group; and, based on the at least a portion of the corpus data, key phrases are extracted and compressed to obtain topic tags and keywords for the topic.

[0013] Based on the semantic information of at least one of the topics, topic tags, and keywords corresponding to each corpus group, and the semantic information of each corpus data in each corpus group, a corpus data index for each corpus group is generated.

[0014] Based on the topic tags, abstracts, representative keywords, and corpus data indexes corresponding to multiple topics, structured data for each topic is constructed.

[0015] Optionally, when the data processing module divides the corpus data into multiple corpus groups based on the semantic vectors of each corpus data, it is specifically used for:

[0016] Using a pre-trained embedding model, semantic understanding processing is performed on each corpus data to obtain semantic vectors that represent the semantics of each corpus data; wherein, the embedding model is trained using a multilingual corpus related to the target topic;

[0017] Based on the similarity between different semantic vectors, clustering is performed on each corpus data to obtain corpus groups corresponding to the multiple topics.

[0018] Optionally, the system further includes: a data acquisition module, used to acquire raw corpus data corresponding to multiple languages ​​from multiple data sources; the raw corpus data corresponding to each language is based on the natural language text corresponding to that language; the multiple sources include at least one of the following: news, policy documents and research reports released by media, international organizations or research institutions;

[0019] For each language, the original corpus data corresponding to each language is preprocessed by at least one of the following: character encoding, hypertext markup language tag cleaning, paragraph segmentation, format standardization, and translation or transcription, to obtain the corpus data corresponding to each language.

[0020] Optionally, the original corpus data includes: mixed corpus data composed of multiple languages;

[0021] The data acquisition module, when translating the acquired raw corpus data corresponding to multiple languages, is used for:

[0022] Based on a pre-established cross-language aligned corpus, the text content in the original corpus data that is not composed of the target language is translated into text content composed of the target language, and the text content composed of the non-target language is retained, thus obtaining the translated corpus data.

[0023] The cross-language aligned corpus includes multiple sets of proprietary terms related to the target topic; each set of proprietary terms includes terms corresponding to multiple languages, and multiple terms in each set of proprietary terms have the same meaning.

[0024] Optionally, the data processing module, when performing semantic summarization processing using the natural language processing model on at least a portion of the corpus data in each corpus group to generate a summary of the topic corresponding to each corpus group; and when extracting and compressing key phrases based on the at least a portion of the corpus data to obtain topic tags and keywords for the topic, is used for:

[0025] Based on the similarity between each piece of corpus data in each corpus group and the cluster center vector corresponding to the corpus group, and based on the similarity between each piece of corpus data in each corpus group, sample corpus data is selected from each corpus group;

[0026] Using the natural language processing model, semantic summarization processing is performed on the sample corpus data to generate topic summaries corresponding to each corpus group. Key phrases are extracted and compressed from the sample corpus data to obtain topic tags and keywords corresponding to the topics.

[0027] Optionally, the content generation module is further configured to generate source information corresponding to the content generation result based on the target corpus data corresponding to the content generation instruction; the source information includes: file source, timestamp, and relevant contextual corpus fragments.

[0028] Optionally, it also includes: a knowledge graph construction module, used to construct a knowledge graph based on the structured data; wherein the nodes in the knowledge graph include at least one of the following: policy documents, national entities, issue tags, institution names, and negotiating positions; wherein each node maps to structured data in multiple languages;

[0029] The edge relationships in the knowledge graph include at least one of the following semantic relationships: text reference, policy evolution, position attribution, and causal influence semantic relationships;

[0030] The content generation module is further configured to: retrieve from the knowledge graph according to the query vector to obtain graph information corresponding to the content generation instruction;

[0031] The content generation module, when using a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a generation result associated with the content generation instruction, is specifically used to: use the pre-trained language generation model to perform language generation processing on the content generation instruction, the target corpus data, and the spectrogram information to obtain a generation result associated with the content generation instruction.

[0032] Optionally, the system further includes: a front-end interaction module;

[0033] The front-end interaction module is used to provide a visual user interface, which is used to receive content generation instructions corresponding to various content generation operations and send the content generation instructions to the content generation module.

[0034] The various content generation operations include at least one of the following: semantic retrieval, intelligent dialogue, data retrieval, and structured reasoning.

[0035] Secondly, this disclosure also provides a method for a multilingual semantic analysis and decision support system for international climate negotiations, including:

[0036] Using a natural language processing model, structured data corresponding to corpus data related to the target topic in multiple languages ​​is automatically generated; wherein, the structured data includes: topic tags, summaries, representative keywords, and corpus data indexes for multiple topics corresponding to the target topic;

[0037] A vector database index is constructed by hierarchically indexing the structured data according to topic, language, and time, thereby generating a corpus vector database;

[0038] In response to receiving a content generation instruction, the content generation instruction is converted into a query vector, and a retrieval is performed from the corpus vector database based on the query vector to obtain target corpus data corresponding to the content generation instruction. Then, a pre-trained language generation model is used to perform language generation processing on the content generation instruction and the target corpus data to obtain a content generation result associated with the content generation instruction.

[0039] Optionally, the step of automatically generating structured data corresponding to corpus data in multiple languages ​​related to the target topic using a natural language processing model includes:

[0040] Based on the semantic vectors of each corpus data, the corpus data is divided into multiple corpus groups; different corpus groups correspond to different topics, and each corpus group includes at least one piece of corpus data associated with the corresponding topic;

[0041] For at least a portion of the corpus data in each corpus group, the natural language processing model is used to perform semantic summarization processing to generate a summary of the topic corresponding to each corpus group; and, based on the at least a portion of the corpus data, key phrases are extracted and compressed to obtain topic tags and keywords for the topic.

[0042] Based on the semantic information of at least one of the topics, topic tags, and keywords corresponding to each corpus group, and the semantic information of each corpus data in each corpus group, a corpus data index for each corpus group is generated.

[0043] Based on the topic tags, abstracts, representative keywords, and corpus data indexes corresponding to multiple topics, structured data for each topic is constructed.

[0044] Optionally, based on the semantic vectors of each corpus data, the corpus data is divided into multiple corpus groups, including:

[0045] Using a pre-trained embedding model, semantic understanding processing is performed on each corpus data to obtain semantic vectors that represent the semantics of each corpus data; wherein, the embedding model is trained using a multilingual corpus related to the target topic;

[0046] Based on the similarity between different semantic vectors, clustering is performed on each corpus data to obtain corpus groups corresponding to the multiple topics.

[0047] Optionally, the method further includes: collecting raw corpus data corresponding to multiple languages ​​from multiple data sources; the raw corpus data corresponding to each language is based on the natural language text of that language; the multiple sources include at least one of the following: news, policy documents and research reports released by media, international organizations or research institutions;

[0048] For each language, the original corpus data corresponding to each language is preprocessed by at least one of the following: character encoding, hypertext markup language tag cleaning, paragraph segmentation, format standardization, and translation or transcription, to obtain the corpus data corresponding to each language.

[0049] Optionally, the original corpus data includes: mixed corpus data composed of multiple languages;

[0050] The translation processing of the collected original corpus data corresponding to multiple languages ​​includes:

[0051] Based on a pre-established cross-language aligned corpus, the text content in the original corpus data that is not composed of the target language is translated into text content composed of the target language, and the text content composed of the non-target language is retained, thus obtaining the translated corpus data.

[0052] The cross-language aligned corpus includes multiple sets of proprietary terms related to the target topic; each set of proprietary terms includes terms corresponding to multiple languages, and multiple terms in each set of proprietary terms have the same meaning.

[0053] Optionally, for at least a portion of the corpus data in each corpus group, semantic summarization processing is performed using the natural language processing model to generate a summary of the topic corresponding to each corpus group; and, based on the at least a portion of the corpus data, key phrase extraction and compression processing is performed to obtain topic tags and keywords for the topic, including:

[0054] Based on the similarity between each piece of corpus data in each corpus group and the cluster center vector corresponding to the corpus group, and based on the similarity between each piece of corpus data in each corpus group, sample corpus data is selected from each corpus group;

[0055] Using the natural language processing model, semantic summarization processing is performed on the sample corpus data to generate topic summaries corresponding to each corpus group. Key phrases are extracted and compressed from the sample corpus data to obtain topic tags and keywords corresponding to the topics.

[0056] Optionally, it also includes: generating source information corresponding to the content generation result based on the target corpus data corresponding to the content generation instruction; the source information includes: file source, timestamp, and relevant contextual corpus fragments.

[0057] Optionally, the method further includes: constructing a knowledge graph based on the structured data; wherein the nodes in the knowledge graph include at least one of the following: policy documents, national entities, issue tags, institution names, and negotiating positions; the edge relationships in the knowledge graph include at least one of the following semantic relationships: text citation, policy evolution, position attribution, and causal influence semantic relationships; wherein each node maps to structured data in multiple languages;

[0058] Based on the query vector, a retrieval is performed from the knowledge graph to obtain graph information corresponding to the content generation instruction;

[0059] The step of using a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a generation result associated with the content generation instruction includes:

[0060] Using the pre-trained language generation model, language generation processing is performed on the content generation instruction, the target corpus data, and the spectrogram information to obtain a generation result associated with the content generation instruction.

[0061] Optionally, a visual user interface is provided, which is used to receive content generation instructions corresponding to various content generation operations and send the content generation instructions to the content generation module;

[0062] The various content generation operations include at least one of the following: semantic retrieval, intelligent dialogue, data retrieval, and structured reasoning.

[0063] Thirdly, an optional implementation of this disclosure also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the second aspect above, or any possible implementation of the second aspect.

[0064] Fourthly, an optional implementation of this disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the second aspect above, or any possible implementation of the second aspect.

[0065] Fifthly, an optional implementation of this disclosure also provides a computer program product carrying program code, the program code including instructions that can be used to perform steps as described in the second aspect or any possible implementation of the second aspect.

[0066] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this disclosure.

[0067] The multilingual semantic analysis and decision support system for international climate negotiations provided in this embodiment includes a data processing module, a database construction module, and a content generation module. The data processing module uses a natural language processing model to automatically generate structured data corresponding to corpus data related to a target topic in multiple languages. The structured data includes topic tags, summaries, representative keywords, and corpus data indexes for multiple issues corresponding to the target topic. The database construction module constructs a vector database index by hierarchically indexing the structured data according to topic, language, and time, generating a corpus vector database. The content generation module, in response to receiving a content generation instruction, converts the instruction into a query vector, retrieves target corpus data corresponding to the instruction from the corpus vector database based on the query vector, and uses a pre-trained language generation model to perform language generation processing on the instruction and the target corpus data to obtain a content generation result associated with the instruction. The system described above uses a deep learning module to automate data processing and corpus structuring, enabling rapid construction of a corpus vector database and quick response to content generation instructions. Based on the corpus vector database and the content generation module, it generates content generation results, improving corpus processing speed and providing more accurate basis for decision-making.

[0068] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0069] Figure 1 This illustration shows a schematic diagram of a multilingual semantic analysis and decision support system for international climate negotiations provided by some embodiments of this disclosure;

[0070] Figure 2 The present disclosure shows a schematic diagram of the structure of a computer device provided in some embodiments;

[0071] Figure 3 A flowchart is shown below illustrating a multilingual semantic analysis and decision support method for international climate negotiations provided by some embodiments of this disclosure;

[0072] Figure 4 This disclosure illustrates an architectural example of a multilingual semantic analysis and decision support system for international climate negotiations, provided by some embodiments thereof. Detailed Implementation

[0073] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0074] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0075] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0076] This disclosure provides a multilingual semantic analysis and decision support system for international climate negotiations, comprising a data processing module, a database construction module, and a content generation module. The data processing module uses a natural language processing model to automatically generate structured data corresponding to corpus data related to a target topic in multiple languages. The structured data includes topic tags, summaries, representative keywords, and corpus data indexes for multiple issues corresponding to the target topic. The database construction module constructs a vector database index by hierarchically indexing the structured data according to topic, language, and time, generating a corpus vector database. The content generation module, in response to receiving a content generation instruction, converts the instruction into a query vector, retrieves target corpus data corresponding to the instruction from the corpus vector database based on the query vector, and uses a pre-trained language generation model to perform language generation processing on the instruction and the target corpus data to obtain a content generation result associated with the instruction. The system described above uses a deep learning module to automate data processing and corpus structuring, enabling rapid construction of a corpus vector database and quick response to content generation instructions. Based on the corpus vector database and the content generation module, it generates content generation results, improving corpus processing speed and providing more accurate basis for decision-making.

[0077] The shortcomings of the above solutions are the result of the inventor's practical experience and careful research. Therefore, the discovery process of the above problems and the solutions proposed in this disclosure below should be considered as the inventor's contribution to this disclosure.

[0078] To facilitate understanding of the technical solutions disclosed herein, the technical terms used in the embodiments of this disclosure will first be explained:

[0079] In this embodiment of the disclosure, the target topic refers to the core object, central idea, or overall scope of the discussion or research, which is the overall framework within a scenario such as discussion, meeting, research, or writing. For example, a meeting topic could be: Global Climate Change and Sustainable Development.

[0080] A topic is a specific issue or direction that supports a theme. It consists of specific issues, discussion points, or sub-directions that revolve around the theme. It is a concretization and refinement of the theme. It consists of several branches that support the theme and is targeted and debatable. For example, topics around the theme of "global climate change and sustainable development" may include: "How to reduce industrial carbon emissions?" and "Green energy transition paths for developing countries".

[0081] To facilitate understanding of this embodiment, a multilingual semantic analysis and decision support system for international climate negotiations disclosed in this disclosure will first be described in detail. In addition to being used in international climate negotiations, the multilingual semantic analysis and decision support system for international climate negotiations provided in this disclosure can also be used in other scenarios that require the integration of multi-source, multilingual corpus data and the ability to obtain a decision corresponding to the problem based on the results of the corpus data integration. This disclosure does not limit the scope of the system.

[0082] The following describes the multilingual semantic analysis and decision support system for international climate negotiations provided by embodiments of this disclosure.

[0083] See Figure 1 As shown in the diagram, this disclosure provides a structural schematic of a multilingual semantic analysis and decision support system for international climate negotiations.

[0084] The embodiments of the multilingual semantic analysis and decision support system for international climate negotiations described in this application can be applied to computer devices. The system embodiments can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical system, it is formed by the processor of the computer device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 2 The diagram shown is a hardware structure diagram of a computer device used in the multilingual semantic analysis and decision support system for international climate negotiations, as described in this application. (Except for...) Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, the computer device in which the device is located in the embodiment may also include other hardware depending on the actual function of the multilingual semantic analysis and decision support system used in the international climate negotiation scenario, which will not be described in detail here.

[0085] See Figure 1 As shown in the embodiments of this disclosure, the multilingual semantic analysis and decision support system for international climate negotiations includes:

[0086] The system includes a data processing module 10, a database construction module 20, and a content generation module 30.

[0087] The data processing module 10 is used to automatically generate structured data corresponding to corpus data related to the target topic in multiple languages ​​using a natural language processing model. The structured data includes topic tags, summaries, representative keywords, and corpus data indexes for multiple topics corresponding to the target topic.

[0088] Database construction module 20: used to construct a vector database index for hierarchical indexing of the structured data according to topic, language and time, and generate a corpus vector database;

[0089] Content generation module 30: In response to receiving a content generation instruction, it converts the content generation instruction into a query vector, retrieves target corpus data corresponding to the content generation instruction from the corpus vector database based on the query vector, and performs language generation processing on the content generation instruction and the target corpus data using a pre-trained language generation model to obtain a content generation result associated with the content generation instruction.

[0090] In addition, other embodiments of this disclosure may include at least one of the following modules:

[0091] The module includes a data acquisition module 40, a knowledge graph construction module 50, and a front-end interaction module 60.

[0092] It should be noted that the data acquisition module 40, knowledge graph construction module 50, and front-end interaction module 60 may not be included in the multilingual semantic analysis and decision support system for international climate negotiations (hereinafter referred to as the semantic analysis and decision support system) provided in this embodiment of the disclosure. The functions implemented by these modules can be implemented by other systems independent of the semantic analysis and decision support system provided in this embodiment of the disclosure. The semantic analysis and decision support system provided in this embodiment of the disclosure can call the functions of the aforementioned other systems or receive data transmitted by other systems to complete the functions to be implemented by the semantic analysis and decision support system.

[0093] The following provides a detailed description of the data processing module 10, the database construction module 20, the content generation module 30, the data acquisition module 40, the knowledge graph construction module 50, and the front-end interaction module 60.

[0094] Regarding the aforementioned data acquisition module 40:

[0095] It is used to collect raw corpus data corresponding to multiple languages ​​from multiple data sources; the raw corpus data corresponding to each language is based on the natural language text of that language; multiple sources include at least one of the following: news, policy documents and research reports released by media, international organizations or research institutions; specifically, the raw corpus data is, for example, data related to the international climate negotiation scenario.

[0096] For each language, perform at least one of the following preprocessing operations on the original corpus data corresponding to each language: character encoding processing, hypertext markup language tag cleaning processing, paragraph segmentation processing, format normalization processing, and translation or transcription processing, to obtain the corpus data corresponding to each language.

[0097] In a specific implementation, the original corpus data corresponding to multiple languages is, for example, a corpus composed of natural languages; for multiple different languages, such as Chinese, English, French, German, etc., the original corpus data composed of natural language texts corresponding to each language can be collected from multiple data sources.

[0098] The original corpus data can generally be in text format, or in other formats, such as picture format, PDF format, etc., which are not limited in the embodiments of the present disclosure.

[0099] For the case where the original corpus data is not in text format, it is possible to first perform character recognition on the original corpus data to parse the specific text content from the original corpus data in other formats.

[0100] The data acquisition module 40 can automatically capture the original corpus data from news, policy documents, and research reports released by various media, international organizations, and research institutions. At the same time, it is also possible to allow users to customize simple information aggregation (Really Simple Syndication, RSS) sources or news web page links, etc., to expand the data capture function of the source orientation.

[0101] After obtaining the original corpus data, at least one of the following preprocessing operations can be first performed on the original corpus data:

[0102] A1: Character encoding processing.

[0103] Specifically, according to a unified encoding method, each character in the original corpus data can be encoded to convert the character into a character vector. For example, the Unicode Transformation Format-8-bit (UTF-8) encoding method is used. Among them, UTF-8 is a variable-length character encoding method for the Unicode character set, which is widely used in the Internet and computer systems and can represent the characters of most languages in the world, such as including Chinese, English, Japanese, Arabic, etc. Character encoding conversion only changes the storage and transmission format of the text, and does not convert synonyms in different languages into the same encoding based on semantics. For example, the words with the same semantics "气候" (Chinese) and "climate" (English) will correspond to completely different byte sequences under UTF-8, so different languages still maintain their independent representations at the encoding level.

[0104] Here, before performing character encoding on the original corpus data, for example, language identification processing can be performed on the original corpus data to obtain the language corresponding to each original corpus data, and then the original corpus data can be encoded according to the encoding method under that language.

[0105] A2: HyperText Markup Language (HTML) tag cleanup.

[0106] Specifically, when scraping raw corpus data, the obtained corpus data may contain HTML tags. HTML tags do not have semantic information related to the target topic, therefore, HTML tags need to be cleaned up.

[0107] A3: Paragraph segmentation.

[0108] Specifically, some of the original corpus data is quite long, and the content of different paragraphs may differ. Furthermore, only some paragraphs in the original corpus data may be relevant to the target topic. Therefore, in this embodiment of the disclosure, after obtaining the original corpus data, it can be segmented to divide a long piece of original corpus data into multiple smaller pieces. This facilitates subsequent processing.

[0109] When performing paragraph segmentation, for example, a maximum and / or minimum length limit can be set; based on this maximum and / or minimum limit, larger sections of the original corpus data can be segmented. Alternatively, the original corpus data can be segmented into paragraphs using its own paragraph division method.

[0110] A4: Format standardization processing.

[0111] Specifically, for example, a specific format can be pre-set, and the original corpus data can be formatted according to this format to facilitate data storage and subsequent use. The specific format used in this embodiment is not limited.

[0112] A5: Translation or transcription processing.

[0113] Specifically, some raw corpus data may contain text corresponding to multiple languages. For example, in raw Chinese corpus data, English words may appear. In such cases, the English words in the raw corpus data can be translated into Chinese text to facilitate subsequent processing.

[0114] The purpose of transcription processing is not to translate semantics, but to convert the pronunciation or writing form of one writing system into another (such as transcribing "Zhang San" from "张三") so as to achieve entity association across writing systems in the subsequent knowledge graph decision-making process.

[0115] In addition, when the data acquisition module performs translation processing on the original corpus data corresponding to multiple languages collected, it is used for:

[0116] According to the pre-established cross-lingual aligned corpus, translate the text content composed of non-target languages in the original corpus data into the text content composed of the target language, and retain the text content composed of the non-target languages, so as to obtain the translated corpus data;

[0117] Among them, the cross-lingual aligned corpus includes multiple groups of specialized vocabulary related to the target topic; each group of specialized vocabulary includes the corresponding vocabulary in multiple languages, and the meanings of multiple vocabulary in each group of specialized vocabulary are the same.

[0118] In another embodiment of the present disclosure, if the above-mentioned multiple preprocessings are performed on a certain original corpus data, generally, character encoding conversion can be completed first to ensure that all original corpus data are in a unified and non-garbled encoding format in memory, and then language detection and mixed language recognition are performed. Subsequently, translation or transcription operations are performed on the detected non-target language segments. The reason for this order is that if the encoding is not unified first, language detection and subsequent processing may fail due to character set conflicts.

[0119] In addition, for low-quality original corpus data, such as low-quality web page content, the system can comprehensively

[0120] Automatically identify and filter indicators such as HTML noise ratio, character abnormality rate, repetition degree, word frequency distribution, and language detection confidence, and optionally configure a relevance scoring mechanism to generate a relevance score corresponding to the target topic for each segment of original corpus data through methods such as keyword matching, semantic similarity calculation, or topic modeling, so as to achieve on-demand screening and weight reduction processing in the subsequent semantic analysis stage, ensuring that the quality of the original corpus data entering the semantic modeling and knowledge extraction link is controllable, the structure is unified, and the content relevance is high.

[0121] This disclosure introduces a climate-specific semantic enhancement mechanism in the aforementioned preprocessing process. The system incorporates multilingual Named Entity Recognition (NER) and a Climate Ontology, automatically identifying and aligning cross-lingual proper nouns such as "Carbon Border Adjustment Mechanism (CBAM)," "Nationally Determined Contribution (NDC)," and "Lost and Damage (L&D)," ensuring semantic consistency in subsequent analyses. Simultaneously, this module establishes a multilingual climate corpus covering past COP documents, IPCC reports, and policy texts from various countries, and continuously fine-tunes the embedding model based on this corpus to achieve accurate cross-lingual semantic mapping. Furthermore, when processing low-quality web pages and mixed-language texts, the system employs a terminology preservation combined with cross-lingual alignment translation strategies to ensure consistency in textual logic and uniformity of professional terminology. This domain-oriented preprocessing innovation effectively improves the accuracy and reliability of subsequent clustering, retrieval, and question-answering modules.

[0122] The aforementioned data acquisition module 40 can provide the data processing module 10 with high-quality, standardized raw corpus input.

[0123] Regarding the aforementioned data processing module 10:

[0124] In specific implementations, a natural language processing model is, for example, a model used to process corpus data composed of multiple languages. In this embodiment of the disclosure, there may be one natural language processing model, which includes multiple processing modules with different natural language processing capabilities.

[0125] There can be multiple natural language processing models. Different natural language processing models have at least one natural language processing capability, and the natural language processing models corresponding to different natural language processing models are different.

[0126] The data processing module 10 provided in this embodiment can call a natural language processing model to automatically generate structured data corresponding to target topic-related corpus data composed of multiple languages.

[0127] Here, the target theme is, for example, a theme related to international climate negotiations, such as climate finance, loss and damage funds, and global carbon market mechanisms. There can be one or more target themes.

[0128] Specifically, when the data processing module 10 automatically generates structured data corresponding to corpus data in multiple languages ​​related to the target topic using a natural language processing model, it can adopt the following methods:

[0129] Based on the semantic vectors of each corpus data, the corpus data is divided into multiple corpus groups; different corpus groups correspond to different topics, and each corpus group includes at least one corpus data associated with the corresponding topic;

[0130] For at least a portion of the corpus data in each corpus group, the natural language processing model is used to perform semantic summarization processing to generate a summary of the topic corresponding to each corpus group; and, based on the at least a portion of the corpus data, key phrases are extracted and compressed to obtain topic tags and keywords for the topic.

[0131] Based on the semantic information of at least one of the topics, topic tags, and keywords corresponding to each corpus group, and the semantic information of each corpus data in each corpus group, a corpus data index for each corpus group is generated.

[0132] Based on the topic tags, abstracts, representative keywords, and corpus data indexes corresponding to multiple topics, structured data for each topic is constructed.

[0133] Here, the semantic vectors of each corpus data are obtained, for example, by feature embedding processing of the original corpus data based on natural language.

[0134] Specifically, in a practical implementation, the data processing module 10 may, for example, divide the corpus data into multiple corpus groups based on the semantic vectors of each corpus data:

[0135] Using a pre-trained embedding model, semantic understanding processing is performed on each corpus data to obtain semantic vectors that represent the semantics of each corpus data.

[0136] Based on the similarity between different semantic vectors, clustering is performed on each corpus data to obtain corpus groups corresponding to the multiple topics.

[0137] In specific implementations, a lightweight embedding model that supports mixed semantic understanding in at least two languages ​​can be pre-trained. This embedding model can be, for example, a deep learning model with multilingual semantic understanding capabilities. For instance, it could be the BGE-m3 model, which supports mixed semantic understanding of Chinese citations and has good cross-lingual semantic alignment capabilities, ensuring comparability and consistency between Chinese and English within the same semantic space. Furthermore, the embedding model described in this embodiment is trained using a multilingual corpus related to the target topic. Thus, by employing a domain-adaptive multilingual model fine-tuning method in the semantic embedding stage, cross-lingual alignment is maintained not only on general corpora but also secondary training is performed on climate issue corpora, enabling the model to possess higher professional sensitivity when processing documents in fields such as COP reports and policy statements.

[0138] After obtaining the semantic vectors corresponding to each corpus data, the corpus data can be clustered based on the similarity between different semantic vectors.

[0139] In clustering processing, unsupervised clustering methods can be employed, such as Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), which features the advantages of not requiring a pre-defined number of clusters and high robustness to outliers. Using this clustering algorithm, the system can automatically discover potential topic groups in news data. For example, corpus data related to "carbon border adjustment mechanisms," "energy transition," and "adaptation funding arrangements" will be grouped into the same cluster. The corpus data assigned to each cluster constitute a corpus group corresponding to a topic related to the target theme.

[0140] In addition to the HDBSCAN clustering method, other unsupervised clustering methods can also be used, such as hierarchical clustering; density-based clustering, such as Ordering Points To Identify the Clustering Structure (OPTICS) algorithm, Mixture Density Networks (MDN), etc. The specific method can be selected according to actual needs, and the embodiments disclosed herein are not limited.

[0141] In terms of clustering algorithms, combining HDBSCAN with incremental clustering mechanisms allows for dynamic updates of clustering results during negotiations without retraining or large-scale computation, adapting to real-time news and policy updates. The system also supports hierarchical clustering. For example, when issues are divided into multiple levels, the first-level issues are clustered first, and then the second-level issues are clustered based on the results. This allows for the differentiation of primary themes such as "mitigation, adaptation, and funding," followed by further subdivision into secondary themes like "energy transition" and "L&D compensation mechanisms," enhancing the analytical capabilities for complex issue structures. The clustering results not only generate automatic summaries and labels but also ensure consistent recognition of the same topic in both Chinese and English through cross-language consistency checks. These innovations overcome the limitations of traditional static semantic clustering in handling dynamic information and multilingual alignment in international multilateral negotiation environments.

[0142] Then, for multiple corpora, at least a portion of the data within each corpus can be processed using natural language processing models to generate structured data corresponding to multiple topics.

[0143] When generating structured data, for example, at least a portion of the data in each corpus can be selected as sample data. The natural language processing model is then used to perform semantic summarization on the sample data to generate a topic summary corresponding to each corpus. Furthermore, key phrases are extracted and compressed from the sample data to obtain topic tags and keywords corresponding to the topics.

[0144] Specifically, the sample corpus data is, for example, representative corpus data that can represent the semantic features of each corpus data within the corresponding corpus group.

[0145] When selecting sample data from a corpus, one method is to calculate the similarity between each data point and the cluster center vector. This involves first calculating the similarity between each data point and the cluster center vector, and then selecting the data points with the highest similarity as sample data.

[0146] Here, for example, the Euclidean distance between the semantic vector and the clustering center vector corresponding to each piece of corpus data can be calculated as the similarity of each piece of corpus data. The larger the Euclidean distance, the lower the corresponding similarity; the smaller the Euclidean distance, the higher the corresponding similarity.

[0147] For example, three data points can be selected as sample data to cover more expressions; if the number of data points in the corpus is small, it can be adaptively reduced to one or two.

[0148] When processing sample corpus data using a natural language processing model, for example, semantic summarization processing is performed on the sample corpus data to generate topic summaries corresponding to each corpus group, and key phrase extraction and compression processing is performed on the sample corpus data to obtain topic tags and keywords corresponding to the topic.

[0149] Here, after extracting multiple sample corpus data, a separate summary is not generated for each sample corpus data. Instead, multiple sample corpus data are sorted by similarity and merged into one input, which is then fed into the language generation model to generate semantic summaries, topic tags, and keywords in a unified manner. This ensures that the summaries, topic tags, and keywords can cover the core information within each corpus group and reduce redundancy.

[0150] After obtaining the semantic summary, topic tags, and keywords of each topic corresponding to each corpus, a corpus data index for each corpus can be generated based on the semantic information of at least one of the summary, topic tags, and keywords of each topic corresponding to each corpus, as well as the semantic information of each corpus data in each corpus.

[0151] Here, we can specifically determine the correlation between the semantic vectors of each piece of corpus data in the corpus, as well as the semantic vectors corresponding to at least one of the semantic summaries, topic tags, and keywords.

[0152] Based on relevance, establish the association between semantic summary, topic tag, and at least one of keywords and each piece of corpus data in the corpus group, and organize the corpus data into an index form based on the association.

[0153] It is important to note here that the semantic summaries, topic tags, and keywords in the structured data are all represented in vector form for ease of use later.

[0154] The index of the corpus data may include, for example, identification information corresponding to the corpus data, such as storage location and name, and may also include the semantic vector of the corpus data. Specific embodiments disclosed herein are not limited.

[0155] Finally, based on the above process, the data processing module 10 converts the clustering results of the corpus data into structured data, which includes the corpus group number, topic label, summary, representative keywords and corpus data index.

[0156] Regarding the aforementioned database construction module 20:

[0157] It is used to construct a vector database index that hierarchically indexes the structured data according to topic, language, and time, thereby generating a corpus vector database.

[0158] In practical implementation, for example, a sharded index structure can be introduced, such as a hierarchical navigable small world (HNSW) structure or an inverted file index + product quantization (IVF+PQ) structure. FAISS (Facebook AISimilarity Search) can be used to build a high-performance vector database index to generate a corpus vector database.

[0159] FAISS, developed by Facebook AI Research (FAIR), is an open-source library for efficiently searching for similar vectors in large-scale datasets. It performs similarity searches by comparing vectors in a high-dimensional space. Its main task is to identify vectors that are "close" to a given query vector based on a specific distance metric. It can quickly perform nearest neighbor searches by indexing vectors and then using efficient algorithms to search for similar vectors.

[0160] HNSW constructs a multi-layered "navigation graph" for vector datasets: the bottom layer is a complete graph containing all vectors; the upper layers are sparse subsets of the bottom layer (randomly sampled by probability), and the higher the layer, the sparser the nodes.

[0161] Each vector node may be connected at different levels, with higher-level nodes acting as "fast navigation signposts" and lower-level nodes responsible for precise matching.

[0162] IVF+PQ combines two technologies. IVF clusters vectors, locating the target cluster first and then searching within it, reducing the number of vectors compared. PQ decomposes high-dimensional vectors into low-dimensional sub-vectors and quantizes each sub-vector individually (replacing the original floating-point vector with integer encoding), thus compressing storage and accelerating distance calculation. Combining the two, IVF narrows the search scope, while PQ compresses storage and accelerates computation. This combination significantly improves search speed, making it suitable for handling vector datasets with hundreds of millions of entries. It also boasts low storage costs: PQ quantization reduces vector storage costs by 10-100 times. Furthermore, adjustable parameters allow for a balance between search accuracy and speed.

[0163] The specific method for constructing the corpus vector database is not limited in this embodiment.

[0164] For the above content generation module 30:

[0165] In response to receiving a content generation instruction, it converts the content generation instruction into a query vector, retrieves target corpus data corresponding to the content generation instruction from the corpus vector database based on the query vector, and uses a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a generation result associated with the content generation instruction.

[0166] In specific implementations, the content generation module 30 may, for example, integrate a Retrieval-Augmented Generation (RAG) question-answering engine and a corresponding content generation model to achieve content generation.

[0167] The type of user input content generation instruction can be set according to actual needs, and this embodiment of the disclosure is not limited. This embodiment of the disclosure takes the content generation instruction as a query question as an example, which can be any language supported by the semantic analysis and decision support system provided in this embodiment of the disclosure. The system first performs language detection and embedding processing on the question to generate a query vector, and then searches for similar content in the generated corpus vector database to return several candidate corpus data with high matching degree.

[0168] Here, when generating the embedding vector of the question, for example, the pre-trained semantic embedding model mentioned above can be used to perform semantic understanding processing on the content generation instruction to obtain a query vector that represents the semantics of the content generation instruction.

[0169] Next, based on the query vector, a retrieval is performed from the corpus vector database to obtain candidate corpus data with a high degree of matching. At this point, the returned candidate corpus data may be in the form of semantic vectors and / or text, or it may be the identification information corresponding to the candidate corpus data. Based on this identification information (such as name, link, etc.), the detailed content of the original corpus data can be obtained from the database.

[0170] The recalled candidate prediction data may or may not be shown to the user; this disclosure does not limit this.

[0171] Furthermore, the semantic analysis and decision support system provided in this embodiment of the present disclosure also utilizes a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a generation result associated with the content generation instruction. This generation result can be used to better implement decisions for subsequent work, thereby achieving decision support.

[0172] In this embodiment of the disclosure, the language generation model may include, for example, BART (Bidirectional and Auto-Regressive Transformers), ChatGLM, or MiniLLM.

[0173] BART, proposed by Facebook AI Research (FAIR), is a Transformer-based pre-trained language model that combines the features of bidirectional encoding (such as BERT) and autoregressive generation (such as GPT), and performs well in various natural language processing (NLP) tasks.

[0174] ChatGLM is a series of open-source conversational language models jointly developed by the Knowledge Engineering Lab (KEG) at Tsinghua University and Zhipu AI. It aims to provide efficient, customizable, and context-appropriate large-scale language model capabilities. Its core features emphasize lightweight design and ease of deployment while ensuring performance, making it suitable for rapid application in various scenarios.

[0175] Mini Large Language Model (MiniLLM) can refer to different projects or models, such as systems that run large language models on consumer-grade graphics processors, and small language models used for knowledge distillation.

[0176] In this embodiment of the disclosure, the above-mentioned basic model can be retrained using corpus samples corresponding to the target topic to obtain a semantic generation model applicable to the semantic analysis and decision support system described in this embodiment of the disclosure.

[0177] Furthermore, the semantic generation model provided in this embodiment can also adopt other model structures, which can realize semantic generation based on content generation instructions and recalled candidate corpus data, and obtain the generation result. This embodiment does not limit the model.

[0178] This query process not only supports single-round question responses but also has multi-round dialogue capabilities. It can automatically maintain the coherence of the topic based on the context of the user's question and supplement the explanation by referencing historical search results. It is particularly suitable for scenarios such as cross-border document comparison, review of historical agreements, and explanation of professional terms.

[0179] In another embodiment of the semantic analysis and decision support system provided by this disclosure, a knowledge graph construction module 50 may also be included.

[0180] The knowledge graph construction module 50 is used to construct a knowledge graph based on the structured data; wherein the nodes in the knowledge graph include at least one of the following: policy documents, national entities, issue tags, institution names, and negotiating positions; wherein each node maps to structured data in multiple languages.

[0181] The edge relationships in the knowledge graph include at least one of the following semantic relationships: text reference, policy evolution, position attribution, and causal influence semantic relationships;

[0182] In this case, the content generation module 30 is further configured to retrieve graph information corresponding to the content generation instruction from the knowledge graph based on the query vector.

[0183] When the content generation module 30 uses a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a generation result associated with the content generation instruction, it is specifically used to: use the pre-trained language generation model to perform language generation processing on the content generation instruction, the target corpus data, and the spectrogram information to obtain a generation result associated with the content generation instruction.

[0184] In this way, the knowledge graph construction module 50 constructs a knowledge graph to describe the relationships between more complex entities. Through the construction of the knowledge graph, the system can support complex queries proposed by users, such as "which countries have opposed the carbon border mechanism in the past three sessions" and "does a certain policy cite the IPCC Sixth Assessment Report?", and provide transparent and interpretable answer logic through graph path reasoning.

[0185] This disclosure employs a temporal ontology modeling approach in the knowledge graph construction process. It not only stores countries, policies, issues, and institutions as static entities but also characterizes policy evolution and position changes through a time dimension, supporting trend reasoning across years and conferences. Simultaneously, the system introduces causal relationship chain modeling into the graph, enabling the identification of complex logic such as "whether a policy was proposed because of a certain report" or "whether a country adjusted its position due to changes in the positions of other countries." For multilingual documents, this invention proposes a cross-language entity alignment mechanism, uniformly mapping Chinese and English text to the same node, achieving unified modeling of cross-language policy semantics. Combined with the path reasoning capabilities of the graph database, the system can answer complex questions such as "which countries have consistently opposed CBAM in the last three COPs" or "whether a certain policy has cited the IPCC Sixth Assessment Report." These innovations make the knowledge graph not only a static knowledge base but also a decision support engine with interpretability and causal reasoning capabilities.

[0186] In another embodiment of the semantic analysis and decision support system provided by this disclosure, for example, it may also include: a front-end interaction module 60, which is used to provide a visual user interface, the user interface being used to receive content generation instructions corresponding to various content generation operations, and to send the content generation instructions to the content generation module;

[0187] The various content generation operations include at least one of the following: semantic retrieval, intelligent dialogue, data retrieval, and structured reasoning.

[0188] Specifically, the front-end interaction module 60, which provides visual interaction, serves as the user interface for the entire semantic analysis and decision support system. For example, it supports users in performing multiple functions such as semantic retrieval, intelligent dialogue, data browsing, and structured reasoning during actual climate negotiations or policy analysis. Upon receiving any of the aforementioned content generation operations, the front-end interaction module 60 converts the content generation operation into a corresponding content generation instruction and passes this instruction to the content generation module 30. The content generation module then executes the content generation process to obtain a content generation result that matches the content generation operation.

[0189] The front-end interaction module 60 can be based on a modular design concept, adopt a responsive World Wide Web (Web) architecture, adapt to various terminals (PC, mobile, screen projection system, etc.), and ensure user identity security and access control through a unified login verification system.

[0190] A visual user interface (i.e., user interface) may include, for example, a function navigation area, an interactive main window, and an operation response area.

[0191] The navigation area integrates functional modules such as "Intelligent Dialogue," "Information Search," "Database," "Analysis Report," and "Export Document," allowing users to quickly access and switch between various system capabilities. For example, clicking the "Intelligent Dialogue" module allows users to enter a dialogue system built on a large language model. In this interface, the system can proactively introduce its functional scope, such as "Identifying Multi-Party Climate Negotiation Positions," "Comparative Analysis of Climate Policies," and "Analyzing NDC Content and Negotiation Text Structure." Users can input relevant questions using natural language. The system will then combine the backend content generation module 30, the corpus vector database built by the database construction module, and / or the knowledge graph built by the knowledge graph construction module to generate structured content, which will then be displayed in real-time in text form on the user interface.

[0192] In addition, the content generation module is also used to generate source information corresponding to the content generation result based on the target corpus data corresponding to the content generation instruction; the source information includes: file source, timestamp, and relevant contextual corpus fragments.

[0193] This source information can be displayed in the interactive window for easy viewing by users.

[0194] In the main interactive window, the system supports content recognition, semantic parsing, strategy matching, and intelligent response to freely input questions from users. Based on the parsing results from the backend model, the system will display the complete response in a dialog box, including cited information sources, point-by-point analysis content, and links for navigation. Taking "What model are you?" as an example, the system will intelligently generate a model description, including modules for supporting document parsing, legal text clause matching, and policy scenario analysis. If the user submits a structured request, such as "Compare the differences in carbon neutrality pathways between 2023 and 2022 in various countries," the system will automatically invoke the backend corpus vector database and graph relationship reasoning capabilities to integrate the multi-round question-and-answer results into a list or card format and return them.

[0195] This publication innovatively designs a temporal and hierarchical vector index structure, supporting not only semantic-based near-nearest neighbor search but also combined retrieval by topic, source, and time dimension, thus supporting scenarios such as "cross-session comparison" and "policy evolution tracking." In the question-and-answer section, the system employs Retrieval Enhanced Generation (RAG) combined with an evidence chain tracing mechanism. While generating natural language answers, it automatically includes source information, including document origin, timestamp, and relevant contextual fragments, enhancing interpretability and credibility. Furthermore, the system supports multilingual mixed queries and cross-round dialogue continuity maintenance, ensuring that users receive coherent, structured, and causally logical answers when asking complex questions (such as "changes in the EU's position on CBAM from 2021-2023"). This innovative approach, combining efficient retrieval, generation, and evidence chain analysis, significantly enhances the professionalism and transparency of the question-and-answer system in international negotiation scenarios.

[0196] Furthermore, the semantic analysis and decision support system provided in this disclosure can also integrate several advanced functional modules, such as:

[0197] Issue analysis function: For example, it can, for a pre-set target issue, call the content generation module based on real-time acquired corpus data associated with the target issue to generate an automatic summary and real-time monitoring results for the target issue; for example, it can display key negotiation issues such as "carbon border adjustment," "financial mechanisms," and "mitigation and adaptation" through a front-end interactive module, and call the content generation module to generate automatic summaries and real-time monitoring results for the above issues; here, for example, when the issue analysis module calls the content generation module, it sends corresponding content instruction information to the content generation module to indicate the specific content to be generated.

[0198] Tool evaluation and recommendation engine: Automatically identifies relevant technology maturity levels, such as Technology Readiness Level (TRL) assessments, through corpora such as technical documents and adaptation reports, and matches action tools accordingly;

[0199] Loss and Damage Reasoning Function: For any issue, the content generation module can be invoked to quickly identify liability, compare policies, and retrieve supporting legal documents;

[0200] Negotiation text comparison and data fusion function: Supports comparative analysis of negotiation drafts in different languages ​​and multiple rounds, and enables entity highlighting and difference extraction.

[0201] To further enhance the user experience, the front-end system supports importing various document formats (such as PDF, Word, and web links) and automatically converts them into structured knowledge units for the content generation module to use. It also supports query record tracking and visual log display, facilitating user review of historical input and reuse of query strategies. All functions are managed through a unified access control platform, supporting multi-level permission settings such as visitor mode, expert mode, and institutional administrator mode. It can also be integrated into on-site screen sharing systems or multilingual collaborative platforms via API.

[0202] like Figure 4 As shown in the embodiments of this disclosure, a specific architecture example of a semantic analysis and decision support system is also provided, including:

[0203] The data acquisition module 40 collects relevant data, such as news data and policy data, to form a news database and a policy text database.

[0204] The data processing module 10 and the database construction module 20 use the corpus data collected by the data acquisition module to construct structured data, and then use the structured data to construct a vector database. The vector database and related vectors constitute a vector index library.

[0205] It is a graph engine (knowledge graph construction module 50), which constructs indicator graphs, and the generated knowledge graphs constitute a graph database.

[0206] The front-end interaction module 60 provides a user interaction interface and integrates related functions, such as information search function, intelligent dialogue function, and data report display function, which are implemented by the information search module, intelligent dialogue interface, and data report display module integrated in the front-end interaction module 60, respectively.

[0207] The semantic analysis and decision support system provided in this disclosure relies on artificial intelligence (AI) technologies such as semantic embedding, vector databases, cluster analysis, retrieval-augmented generation (RAG) question answering, and knowledge graphs to build an intelligent system that supports multilingual semantic consistency, cross-border position comparison, deep text analysis, and policy reasoning capabilities. This system can effectively improve the efficiency, transparency, and decision-making quality of international climate negotiations. The system is particularly suitable for information gathering, trend identification, and analysis tasks conducted in the context of global climate governance, such as at international multilateral conferences, national policy research institutions, think tanks, and university research teams, and has broad prospects for promotion and application value.

[0208] See Figure 3As shown in the embodiments of this disclosure, a semantic analysis and decision support method is also provided. The executing entity of the semantic analysis and decision support method provided in this disclosure is generally a computer device with a certain computing power. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), a mobile device, a user terminal, a terminal, an in-vehicle device, a wearable device, etc. In some possible implementations, the semantic analysis and decision support method can be implemented by a processor calling computer-readable instructions stored in memory.

[0209] The semantic analysis and decision support method includes steps S301 to S303, wherein:

[0210] S301: Using a natural language processing model, automatically generate structured data corresponding to corpus data in multiple languages ​​related to the target topic; wherein, the structured data includes: topic tags, summaries, representative keywords, and corpus data indexes for multiple topics corresponding to the target topic;

[0211] S302: Construct a vector database index by hierarchically indexing the structured data according to topic, language, and time, and generate a corpus vector database;

[0212] S303: In response to receiving a content generation instruction, the content generation instruction is converted into a query vector, and a retrieval is performed from the corpus vector database according to the query vector to obtain target corpus data corresponding to the content generation instruction. Then, a pre-trained language generation model is used to perform language generation processing on the content generation instruction and the target corpus data to obtain a content generation result associated with the content generation instruction.

[0213] Optionally, the step of automatically generating structured data corresponding to corpus data in multiple languages ​​related to the target topic using a natural language processing model includes:

[0214] Based on the semantic vectors of each corpus data, the corpus data is divided into multiple corpus groups; different corpus groups correspond to different topics, and each corpus group includes at least one piece of corpus data associated with the corresponding topic;

[0215] For at least a portion of the corpus data in each corpus group, the natural language processing model is used to perform semantic summarization processing to generate a summary of the topic corresponding to each corpus group; and, based on the at least a portion of the corpus data, key phrases are extracted and compressed to obtain topic tags and keywords for the topic.

[0216] Based on the semantic information of at least one of the topics, topic tags, and keywords corresponding to each corpus group, and the semantic information of each corpus data in each corpus group, a corpus data index for each corpus group is generated.

[0217] Based on the topic tags, abstracts, representative keywords, and corpus data indexes corresponding to multiple topics, structured data for each topic is constructed.

[0218] Optionally, based on the semantic vectors of each corpus data, the corpus data is divided into multiple corpus groups, including:

[0219] Using a pre-trained embedding model, semantic understanding processing is performed on each corpus data to obtain semantic vectors that represent the semantics of each corpus data; wherein, the embedding model is trained using a multilingual corpus related to the target topic;

[0220] Based on the similarity between different semantic vectors, clustering is performed on each corpus data to obtain corpus groups corresponding to the multiple topics.

[0221] Optionally, the method further includes: collecting raw corpus data corresponding to multiple languages ​​from multiple data sources; the raw corpus data corresponding to each language is based on the natural language text of that language; the multiple sources include at least one of the following: news, policy documents and research reports released by media, international organizations or research institutions;

[0222] For each language, the original corpus data corresponding to each language is preprocessed by at least one of the following: character encoding, hypertext markup language tag cleaning, paragraph segmentation, format standardization, and translation or transcription, to obtain the corpus data corresponding to each language.

[0223] Optionally, the original corpus data includes: mixed corpus data composed of multiple languages;

[0224] The translation processing of the collected original corpus data corresponding to multiple languages ​​includes:

[0225] Based on a pre-established cross-language aligned corpus, the text content in the original corpus data that is not composed of the target language is translated into text content composed of the target language, and the text content composed of the non-target language is retained, thus obtaining the translated corpus data.

[0226] The cross-language aligned corpus includes multiple sets of proprietary terms related to the target topic; each set of proprietary terms includes terms corresponding to multiple languages, and multiple terms in each set of proprietary terms have the same meaning.

[0227] Optionally, for at least a portion of the corpus data in each corpus group, semantic summarization processing is performed using the natural language processing model to generate a summary of the topic corresponding to each corpus group; and, based on the at least a portion of the corpus data, key phrase extraction and compression processing is performed to obtain topic tags and keywords for the topic, including:

[0228] Based on the similarity between each piece of corpus data in each corpus group and the cluster center vector corresponding to the corpus group, and based on the similarity between each piece of corpus data in each corpus group, sample corpus data is selected from each corpus group;

[0229] Using the natural language processing model, semantic summarization processing is performed on the sample corpus data to generate topic summaries corresponding to each corpus group. Key phrases are extracted and compressed from the sample corpus data to obtain topic tags and keywords corresponding to the topics.

[0230] Optionally, it also includes: generating source information corresponding to the content generation result based on the target corpus data corresponding to the content generation instruction; the source information includes: file source, timestamp, and relevant contextual corpus fragments.

[0231] Optionally, the method further includes: constructing a knowledge graph based on the structured data; wherein the nodes in the knowledge graph include at least one of the following: policy documents, national entities, issue tags, institution names, and negotiating positions; the edge relationships in the knowledge graph include at least one of the following semantic relationships: text citation, policy evolution, position attribution, and causal influence semantic relationships; wherein each node maps to structured data in multiple languages;

[0232] Based on the query vector, a retrieval is performed from the knowledge graph to obtain graph information corresponding to the content generation instruction;

[0233] The step of using a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a generation result associated with the content generation instruction includes:

[0234] Using the pre-trained language generation model, language generation processing is performed on the content generation instruction, the target corpus data, and the spectrogram information to obtain a generation result associated with the content generation instruction.

[0235] Optionally, a visual user interface is provided, which is used to receive content generation instructions corresponding to various content generation operations and send the content generation instructions to the content generation module;

[0236] The various content generation operations include at least one of the following: semantic retrieval, intelligent dialogue, data retrieval, and structured reasoning.

[0237] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the semantic analysis and decision support method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0238] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the semantic analysis and decision support method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0239] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0240] The computer program or instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; it can also be an optical medium, such as a digital video optical disc; or it can be a semiconductor medium, such as a solid-state drive. The computer-readable storage medium can be a volatile or non-volatile storage medium, or it can include both volatile and non-volatile types of storage media.

[0241] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0242] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A multilingual semantic analysis and decision support system for international climate negotiations, characterized in that, include: The data processing module is used to automatically generate structured data corresponding to corpus data in multiple languages ​​and related to the target topic using a natural language processing model; wherein, the structured data includes: topic tags, summaries, representative keywords, and corpus data indexes for multiple topics corresponding to the target topic; The database construction module is used to construct a vector database index by hierarchically indexing the structured data according to topic, language, and time, and generate a corpus vector database. The content generation module is configured to, in response to receiving a content generation instruction, convert the content generation instruction into a query vector, and retrieve target corpus data corresponding to the content generation instruction from the corpus vector database based on the query vector, and use a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a content generation result associated with the content generation instruction.

2. The system according to claim 1, characterized in that, The data processing module is specifically used for: Based on the semantic vectors of each corpus data, the corpus data is divided into multiple corpus groups; different corpus groups correspond to different topics, and each corpus group includes at least one piece of corpus data associated with the corresponding topic; For at least a portion of the corpus data in each corpus group, the natural language processing model is used to perform semantic summarization processing to generate a summary of the topic corresponding to each corpus group; and, based on the at least a portion of the corpus data, key phrases are extracted and compressed to obtain topic tags and keywords for the topic. Based on the semantic information of at least one of the topics, topic tags, and keywords corresponding to each corpus group, and the semantic information of each corpus data in each corpus group, a corpus data index for each corpus group is generated. Based on the topic tags, abstracts, representative keywords, and corpus data indexes corresponding to multiple topics, structured data for each topic is constructed.

3. The system according to claim 2, characterized in that, The data processing module, when dividing the corpus data into multiple corpus groups based on the semantic vectors of each corpus data, is specifically used for: Using a pre-trained embedding model, semantic understanding processing is performed on each corpus data to obtain semantic vectors that represent the semantics of each corpus data; wherein, the embedding model is trained using a multilingual corpus related to the target topic; Based on the similarity between different semantic vectors, clustering is performed on each corpus data to obtain corpus groups corresponding to the multiple topics.

4. The system according to claim 1, characterized in that, The system further includes: a data acquisition module, used to collect raw corpus data corresponding to multiple languages ​​from multiple data sources; the raw corpus data corresponding to each language is based on the natural language text of that language; the multiple sources include at least one of the following: news, policy documents and research reports released by media, international organizations or research institutions; For each language, the original corpus data corresponding to each language is preprocessed by at least one of the following: character encoding, hypertext markup language tag cleaning, paragraph segmentation, format standardization, and translation or transcription, to obtain the corpus data corresponding to each language.

5. The system according to claim 4, characterized in that, The original corpus data includes: mixed corpus data composed of multiple languages; The data acquisition module, when translating the acquired raw corpus data corresponding to multiple languages, is used for: Based on a pre-established cross-language aligned corpus, the text content in the original corpus data that is not composed of the target language is translated into text content composed of the target language, and the text content composed of the non-target language is retained, thus obtaining the translated corpus data. The cross-language aligned corpus includes multiple sets of proprietary terms related to the target topic; each set of proprietary terms includes terms corresponding to multiple languages, and multiple terms in each set of proprietary terms have the same meaning.

6. The system according to claim 2, characterized in that, The data processing module performs semantic summarization processing on at least a portion of the corpus data in each corpus using the natural language processing model to generate a summary of the topic corresponding to each corpus. And, based on the at least part of the corpus data, when extracting and compressing key phrases to obtain the topic tags and keywords of the topic, it is used for: Based on the similarity between each piece of corpus data in each corpus group and the cluster center vector corresponding to the corpus group, and based on the similarity between each piece of corpus data in each corpus group, sample corpus data is selected from each corpus group; Using the natural language processing model, semantic summarization processing is performed on the sample corpus data to generate topic summaries corresponding to each corpus group. Key phrases are extracted and compressed from the sample corpus data to obtain topic tags and keywords corresponding to the topics.

7. The system according to claim 1, characterized in that, The content generation module is further configured to generate source information corresponding to the content generation result based on the target corpus data corresponding to the content generation instruction; the source information includes: file source, timestamp, and relevant contextual corpus fragments.

8. The system according to any one of claims 1-7, characterized in that, Also includes: A knowledge graph construction module is used to construct a knowledge graph based on the structured data; wherein, the nodes in the knowledge graph include at least one of the following: policy documents, national entities, issue tags, institution names, and negotiating positions; wherein, each node maps to structured data in multiple languages; The edge relationships in the knowledge graph include at least one of the following semantic relationships: text reference, policy evolution, position attribution, and causal influence semantic relationships; The content generation module is further configured to: retrieve from the knowledge graph according to the query vector to obtain graph information corresponding to the content generation instruction; The content generation module, when using a pre-trained language generation model to perform language generation processing on the content generation instruction and the target corpus data to obtain a generation result associated with the content generation instruction, is specifically used to: use the pre-trained language generation model to perform language generation processing on the content generation instruction, the target corpus data, and the spectrogram information to obtain a generation result associated with the content generation instruction.

9. The system according to claim 1, characterized in that, The system also includes: a front-end interaction module; The front-end interaction module is used to provide a visual user interface, which is used to receive content generation instructions corresponding to various content generation operations and send the content generation instructions to the content generation module. The various content generation operations include at least one of the following: semantic retrieval, intelligent dialogue, data retrieval, and structured reasoning.

10. A multilingual semantic analysis and decision support method for international climate negotiations, characterized in that, include: Using a natural language processing model, structured data corresponding to corpus data related to the target topic in multiple languages ​​is automatically generated; wherein, the structured data includes: topic tags, summaries, representative keywords, and corpus data indexes for multiple topics corresponding to the target topic; A vector database index is constructed by hierarchically indexing the structured data according to topic, language, and time, thereby generating a corpus vector database; In response to receiving a content generation instruction, the content generation instruction is converted into a query vector, and a retrieval is performed from the corpus vector database based on the query vector to obtain target corpus data corresponding to the content generation instruction. Then, a pre-trained language generation model is used to perform language generation processing on the content generation instruction and the target corpus data to obtain a content generation result associated with the content generation instruction.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method of claim 10.

12. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it performs the following steps: Using a natural language processing model, structured data corresponding to corpus data related to the target topic in multiple languages ​​is automatically generated; wherein, the structured data includes: topic tags, summaries, representative keywords, and corpus data indexes for multiple topics corresponding to the target topic; A vector database index is constructed by hierarchically indexing the structured data according to topic, language, and time, thereby generating a corpus vector database; In response to receiving a content generation instruction, the content generation instruction is converted into a query vector, and a retrieval is performed from the corpus vector database based on the query vector to obtain target corpus data corresponding to the content generation instruction. Then, a pre-trained language generation model is used to perform language generation processing on the content generation instruction and the target corpus data to obtain a content generation result associated with the content generation instruction.

Images