Short text dialogue-oriented multilevel intention recognition agent processing system, method and device, processor and storage medium thereof

Abstract

The invention relates to a multi-level intention recognition agent processing system for short text dialogues, and the system comprises a data processing module which is used for preprocessing diversified multilingual original data, and segmenting the data into a segment form according to semantics; the intention coarse screening module is used for performing intention analysis on the input segment through a large model, so that the large model is matched with the most relevant intention category from a preset intention candidate set, and summarizing the content of the segment in detail to generate a corresponding word cloud picture; the background search module is used for carrying out keyword extraction or background extraction on the input segment and carrying out background search and arrangement through an MCP webpage search tool and a large model; the intention recognition module is used for jointly inputting intention summary content and associated background information into the large model and outputting structured complete intention analysis by utilizing prompt engineering; and the report generation module is used for combining the complete intention analysis content generated by intention recognition with the word cloud picture information to generate a normative intention analysis report.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a workflow building method of an agent system, a prompt word engineering design method of a large model and an agent tool design method, and specifically refers to a multi-level intention recognition agent processing system and method for short text dialogue, an apparatus, a processor and a computer readable storage medium thereof. BACKGROUND

[0002] Today's resource security issues have become the core topic of consultation among governments of various countries, and also the focus of discussion of all sectors of society. Under this background, how to accurately screen whether the dialogue text contains resource security related information, and efficiently identify the intention of the dialogue involving resource security, has become a problem to be solved. The traditional manual screening and intention recognition method has defects such as slow processing speed, easy to miss or misjudge, limited coverage, etc., and has long been unable to meet the analysis needs of massive dialogue texts.

[0003] To address this challenge, existing computer-aided analysis tools have emerged, which have improved data analysis efficiency to some extent through automation technology. However, these tools still have many shortcomings in practical application: first, traditional manual or computer-aided resource security related information screening and intention recognition consumes a lot of manpower, material and financial resources, and is time-consuming, which is difficult to match the efficient analysis needs of modern scenarios; second, such methods cannot combine resource security exclusive semantic background, and are difficult to accurately associate resource related industries, policies, trade and other real scenarios for identification; third, traditional methods cannot visually display resource security related semantic content, which is not convenient for users to intuitively grasp the core information and intention tendency.

[0004] On this basis, the existing voice call intention recognition technology has the following core pain points: 1. Insufficient segmentation accuracy: traditional fixed-length segmentation does not consider semantic coherence, resulting in single segment containing multiple intentions or single intention being split, affecting subsequent analysis; 2. One-sided intention analysis: relying only on dialogue text, lacking news hot spot background information support, unable to accurately judge ambiguous intentions (such as "warranty issues" cannot distinguish between "time limit" vs. "scope"); 3. Imbalance between efficiency and accuracy: full-scale calling large model analysis long text is low in efficiency, and only using rule base matching leads to low accuracy; 4. Unstructured output: results are mostly plain text, lacking visual presentation (such as core semantic extraction) and standardized reporting, making it difficult to reuse and statistics.

[0005] In summary, the existing computer-aided analysis tools and traditional manual methods cannot meet the needs of screening and intent recognition of resource security related information in massive dialogue texts. Therefore, how to use artificial intelligence technology, especially deep learning and natural language processing technology, to build an intelligent agent screening system to realize accurate screening and intelligent intent recognition of resource security related information in dialogue texts and improve the accuracy and efficiency of analysis has become a problem that needs to be solved in the current related field. SUMMARY

[0006] The purpose of the present application is to overcome the above-mentioned shortcomings of the prior art, and to provide a multi-level intent recognition intelligent agent processing system, method, device, processor and computer readable storage medium for short text dialogue.

[0007] In order to achieve the above-mentioned purpose, the multi-level intent recognition intelligent agent processing system, method, device, processor and computer readable storage medium for short text dialogue of the present application are as follows: The multi-level intent recognition intelligent agent processing system for short text dialogue, the main feature is that the system comprises: A data processing module for pre-processing diversified multilingual raw data and segmenting according to semantics in the form of segments; An intent rough screening module connected to the data processing module, for performing intent analysis on the input segment through a large model, allowing the large model to match the most relevant intent category from a pre-set intent candidate set, and making a detailed summary of the segment content to generate a corresponding word cloud diagram; A background search module connected to the data processing module, for extracting keywords or background from the input segment, and performing background search and sorting through an MCP web search tool and a large model; An intent recognition module connected to the intent rough screening module and the background search module, for inputting the intent summary content and associated background information into the large model, and outputting a complete intent analysis with structure through prompt engineering; A report generation module connected to the intent recognition module, for combining the complete intent analysis content generated by the intent recognition and the word cloud diagram information to generate a standardized intent analysis report.

[0008] Preferably, the data processing module specifically performs the following processing: Read multi-source data files, integrate the obtained data into structured text data, and perform data cleaning to remove non-text information, correct errors and grammatical errors, and generate initial semantic text; Based on the natural language processing model, the semantic similarity of the initial semantic text is calculated to determine the segmentation point, and intelligent segmentation is performed based on semantic integrity and logical coherence. An output structured segments list is generated, each segment including a unique segment ID and a coherence score.

[0009] Preferably, the intent coarse screening module specifically performs the following processing: Each segment obtained from the structured segments list is input, a large model is called to read a preset intent candidate set for preliminary intent analysis, and the most relevant intent category is matched; The large model generates a detailed summary for the current segment, including core information, key points, and semantic sentiment orientation; The large model is guided to output core vocabulary using a prompt word engineering, and the core vocabulary is visualized using a word cloud generation tool to generate a corresponding word cloud diagram.

[0010] Preferably, the background search module specifically performs the following processing: The large model extracts core keywords, entities, and background clues in the segment; then, based on the MCP tool, a web search tool is called to search for the extracted key elements to obtain related background knowledge, industry trends, historical data, news hotspots, or similar cases; The large model filters, removes duplicates, and integrates the search results to extract background information highly relevant to the current segment, which is used to provide external knowledge supplement for in-depth understanding of the intent recognition module, thereby outputting structured background information results.

[0011] Preferably, the intent recognition module specifically performs the following processing: The preliminary intent and summary output by the intent coarse screening module and the associated background information provided by the background search module are integrated and input into the large model together; Through the design of the prompt engineering, the large model is guided to perform deep reasoning, and the confidence degree of intent recognition is given. Low-confidence content is manually reviewed, and the reasons and potential motives for intent generation are explained in combination with background information. The analysis results are output according to the preset fixed structure to form a standardized and directly usable complete intent analysis report segment.

[0012] Preferably, the report generation module specifically performs the following processing: The structured analysis content output by the intent recognition module and the word cloud generated by the intent coarse screening module are fused, and content filling is performed according to the preset report template; Natural language polishing is performed on the structured analysis content to ensure smooth and professional expression, and the generated word cloud picture is integrated into the report in the form of a link or embedding, echoing the text content; Finally, format checking is performed to generate a document that meets the MD specification, facilitating review, sharing, and archiving.

[0013] The short text dialogue-oriented multi-level intent recognition agent processing method based on the above-described system, which mainly features that the method includes the following steps: (1) The user uploads the original text dialogue multi-source data through the Agent portal, and the system initializes the process ID; (2) The Agent calls the data processing module to perform data reading and semantic segmentation, and outputs the corresponding segments list; if the processing fails, an exception log is generated and terminated; (3) The Agent simultaneously calls the intent rough screening module and the background search module to input the segments list, the intent rough screening module performs intent matching and summarization, and generates a corresponding word cloud based on keywords, completing the initial screening of intent; the background search module extracts keywords and triggers background information search, outputting the background result; (4) The Agent receives the intent rough screening result and the background screening result, and calls the intent recognition module to output the structured complete intent analysis; (5) The Agent calls the report generation module to input the complete intent analysis content and the word cloud picture, and generates an MD specification document; (6) The Agent returns the MD specification document to the user, and the process ends.

[0014] The short text dialogue-oriented multi-level intent recognition agent processing device, which mainly features that the device includes: a processor configured to execute computer executable instructions; a memory storing one or more computer executable instructions, which, when executed by the processor, implement the steps of the above-described short text dialogue-oriented multi-level intent recognition agent processing method.

[0015] The processor for short text dialogue-oriented multi-level intent recognition agent processing, which mainly features that the processor is configured to execute computer executable instructions, and the computer executable instructions, when executed by the processor, implement the steps of the above-described short text dialogue-oriented multi-level intent recognition agent processing method.

[0016] The computer readable storage medium mainly characterized by storing a computer program thereon, the computer program being executable by a processor to implement the steps of the multi-level intention recognition intelligent agent processing method for short text dialogue.

[0017] The multi-level intention recognition intelligent agent processing system, method, device, processor and computer readable storage medium thereof of the present application have the following advantages: for resource security related short text dialogue intention recognition needs, through the whole process design of "data preprocessing-parallel analysis-accurate recognition-structured output", significant technical effects are achieved. The data preprocessing module combines rule segmentation and semantic segmentation, and the accuracy rate is much higher than that of the traditional segmentation mode, avoiding intention fragmentation; the intention rough screening and background search modules operate in parallel, the former relies on fine Prompt matching to match resource security exclusive intentions and generate word clouds, and the latter calls Bing to supplement external background based on MCP, which greatly improves the intention recognition accuracy, especially the ambiguity intention recognition precision. The system processing efficiency is optimized, and the report can be generated within a short time after the call ends in real-time scenarios, which is significantly faster than the traditional scheme; the report generation module outputs MD format documents, integrates structured analysis and word clouds, greatly reduces the time for users to obtain core information, and at the same time reduces the AI application threshold of small and medium-sized businesses, promotes the conversion of resource security dialogue data from storage cost to decision asset, and takes into account the technical accuracy, economic cost reduction and industry empowerment value. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The multi-level intention recognition intelligent agent processing system for short text dialogue of the present application.

[0019] Figure 2 The multi-level intention recognition intelligent agent processing method for short text dialogue of the present application.

[0020] Figure 3 The processing flowchart of the background search module of the present application.

[0021] Figure 4 The overall timing processing flowchart of the system of the present application.

[0022] Figure 5 The system framework diagram of the present application.

[0023] Figure 6 The word cloud diagram generated in a specific embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to more clearly describe the technical content of the present application, further description will be made in combination with specific embodiments.

[0025] Before the embodiments of the present application are explained in detail, it is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. It is also to be understood that where the description above refers to determining, identifying, or the like, this refers to any process controlled by a processor whether computer implemented or otherwise.

[0026] Before the technical solutions are described in detail, the key terms used in the technical solutions are described as follows: AI Agent: Artificial Intelligence Agent, refers to an automated system that can autonomously receive input, schedule resources, execute tasks, and output results; NLP (Natural Language Processing): Natural Language Processing, a technology that enables computers to understand human language; MD (Markdown): A lightweight markup language that can quickly generate structured documents and is compatible with multiple platforms for viewing; Prompt engineering: A technology that guides large models to output results that meet requirements by designing precise instructions (Prompts).

[0027] MCP (Model Context Protocol) was released by Anthropic in November 2024, aiming to establish a unified, standardized, and safely controllable connection method for large language models (LLM) and external data sources, tools, and services. It is considered as the "universal plug / USB-C" in the AI ecosystem, allowing models to access various capabilities in a consistent manner and enabling two-way communication and composable orchestration, thereby improving interoperability and scalability.

[0028] The present application aims to provide an intelligent agent with the ability of "data preprocessing-intent rough screening-background search-fusion analysis" whole process, which can solve the technical problems of existing intelligent agents in data processing regularity, intent information integrity, background description accuracy and analysis report coherence.

[0029] In order to achieve the above purpose, the present application adopts the following technical solutions, the multi-level intent recognition system for short text dialogue includes: 1. Data processing module: responsible for systematic processing of diversified multilingual raw data (such as text dialogue records). First, data cleaning is performed to remove non-text information (such as special symbols, format markers), correct spelling and grammar errors, and remove meaningless text; then intelligent segmentation is performed based on semantic integrity and logical coherence, which can generate segments using punctuation symbols or fixed length windows (combined with context adjustment); at the same time, metadata tags (speaker identification, etc.) are added to each segment, and reasonable adjustment (split long text, merge short text) is made for segments that are too long or too short to ensure the efficiency and accuracy of subsequent module processing.

[0030] 2. Intent coarse screening module: taking the pre-processed segment as input, calling large model to perform preliminary intent analysis. By constructing fine prompt words (clearly limiting the range of intent list, requiring output intent judgment basis), the model matches the most possible intent category about resource security from the pre-set intent candidate set. At the same time, the model needs to generate a detailed summary of the segment, covering core information, key points and semantic sentiment tendency. In addition, based on the summary content and the keyword extraction ability of the large model, the model outputs core words guided by prompt words engineering, and then visualizes these words through word cloud generation tools to highlight the key information in the segment, providing intuitive reference for subsequent analysis.

[0031] 3. Background search module: parallel processing of the same segment with the intent coarse screening module, focusing on mining associated information behind the content. First, extract the core keywords, entities (such as names, events, professional terms) and background clues in the segment through the large model; then based on MCP (model context protocol) call web search tools to search for these elements, get related background knowledge, industry dynamics, historical data, news hotspots or similar cases; then use the large model to filter, remove duplicates and integrate the search results, extract the background information highly related to the current segment, provide external knowledge supplement for in-depth understanding of intent. When there is no search result, analyze the reason for no result.

[0032] 4. Intention recognition module: As the core analysis link, the preliminary intention output by the intention screening module and the associated background information provided by the summary and background search module are integrated and input into the large model. Through the carefully designed prompt engineering (clearly structured output format, complete logical chain, and emphasis on the correlation analysis of intention and background), the model is guided to conduct deep reasoning: and give the confidence degree of intention recognition, manually review the low confidence content; explain the reasons and potential motives of intention generation combined with background information; output the analysis results according to the fixed structure (intention category, core appeal, supporting basis, associated background, semantic sentiment tendency), forming a standardized and directly usable complete intention analysis report segment. This module is the core innovative module of the patent, and the final conclusion is obtained by integrating the contents of the above steps and based on the customized prompt engineering.

[0033] 5. Report generation module: responsible for fusing the structured analysis content output by the intention recognition module and the word cloud generated by the intention screening module, and finally generating a complete intention analysis report (md format). First, fill in the content according to (including title, table of contents, segment basic information, intention analysis details, word cloud display area, summary and suggestions, etc.); polish the structured analysis content in natural language to ensure smooth and professional expression; integrate the word cloud picture into the report in the form of link or embedding, and echo with the text content; finally, perform format checking (such as title level, list style, picture layout), generate md specification documents for easy reading, sharing and archiving. The structured content and graphics of the previous step are combined to form the final complex document analysis report for user use.

[0034] Further, the intention screening prompt words and intention recognition list are: "Please analyze the following dialogue paragraph, select the most matching intention from the provided intention list, and make a detailed summary of the content, and output the top five words in the text. Cover core information, key points and semantic sentiment tendency: Intention list: {intent_list} Dialogue paragraph: {segment} Please return in JSON format, including "summary", "key points", and "sentiment tendency" fields." ; Among them, intent_list = ["resource security related news", "resource security related policy", "resource security related transaction", "other resource security related conversation", "conversation content unrelated to resource security"].

[0035] Further, the word cloud generation method mainly relies on Chinese and English stop words in the jieba library, and then uses the generate method in the jieba library to generate word clouds and generate unique picture names for each word cloud picture based on the hash method. Save in the same directory and return the word cloud path.

[0036] Further, the MCP of the background search module refers to Model Context Protocol, released by Anthropic in November 2024, aiming to establish a unified, standardized, and safely controllable connection method for large language models (LLM) and external data sources, tools, and services. It is considered as the "universal plug / USB-C" in the AI ecosystem, allowing models to access various capabilities in a consistent manner and enabling bidirectional communication and composable orchestration, thereby improving interoperability and scalability. The web search tool is an MCP server, and this patent selects the Bing search engine, which has the following configuration format: { "mcpServers": { "bingcnmcp": { "command": "cmd", "args": [" / c","npx","bing-cn-mcp"] } } } The tool contained in the MCP server is bing_search, which has the ability to search Bing and get the result list. Parameters: query: search keywords; num_results: return result quantity (default is 5). fetch_webpage, which has the ability to get the content of the corresponding webpage according to the search result ID. Parameter: result_id: result ID returned from bing_search.

[0037] Further, the prompt words of the intent recognition module are: "Please generate a structured and complete intent analysis based on the following information, requiring a complete logical chain, emphasizing the relevance analysis of intent and background, and explaining the causes and potential motivations of intent generation in combination with background information: Preliminary intent identification: {rough_intent} Content summary: {summary} Background information: {background} The return content is returned in strict json format, and the analysis results are output according to the fixed structure (intent category, core appeal, supporting basis, associated background, semantic sentiment tendency, and confidence).

[0038] Referring to Figure 1 As shown in the figure, the system takes the AI Agent entry module as the overall control, coordinates the data flow and logical interaction of the five core function modules, and the module relationship is as follows: Hierarchical relationship: the Agent entry module is the top-level overall control, the data preprocessing module is the "data preparation layer", the intent coarse screening module + background search module is the "parallel analysis layer", the intent recognition module is the "accurate decision layer", and the report generation module is the "result output layer"; Inclusion relationship: no module nesting, each module is an independent functional unit, and only data interaction is realized through the Agent entry; Information interaction relationship: 1. Agent entry→data processing module: transmit raw text conversation data; 2. Data processing module→Agent entry: feedback structured semantic segmentation segments; 3. Agent entry→intent coarse screening module / background search module: parallel distribution of segments; 4. Intent coarse screening module→Agent entry: feedback "initial intent + conversation summary + keywords + word cloud diagram"; 5. Background search module→Agent entry: feedback "segment keywords + background abstract + information source"; 6. Agent entry→intent recognition module: transmit "coarse screening results + background results"; 7. Intent recognition module→Agent entry: feedback "structured complete intent analysis"; 8. Agent entry→report generation module: transmit "complete intent analysis + word cloud diagram"; 9. Report generation module→Agent entry: feedback "MD format analysis report".

[0039] The specific functions of each functional module will be further described in detail below: (1) AI Agent entry module Function: as a system control unit, responsible for receiving input data.

[0040] Working principle: 1. Receive user uploaded raw text conversation data (support formats: pdf, txt, word, support real-time stream or offline file); 2. Initialize module call queue, trigger modules in the order of "preprocessing→parallel coarse screening / background search→intent recognition→report generation"; 3. If a module reports an error, trigger the retry mechanism (up to 3 times), and if the retry fails, generate an exception log and terminate the process; 4. After the process is completed, return the MD report to the user.

[0041] (2) Data preprocessing module Core function: Convert unstructured short text conversation data into structured and semantically independent text segments, providing high-quality input for subsequent analysis.

[0042] Key technical means and working process: 1. Multi-source file reading: Based on data reading methods such as pandas, read multi-source data files and integrate the data into structured text data. By reading different file types, we get the analysis data.

[0043] 2. Semantic segmentation processing: Based on natural language processing (NLP, Natural Language Processing) models (such as BERTbase), perform "semantic coherence scoring" on the initial text: calculate the semantic similarity of consecutive sentences, and when the similarity is lower than the threshold (such as 0.6), determine the segmentation point; Combine rule-based segmentation: if a speaker switch, punctuation (period / exclamation mark) or conversation termination word (such as "goodbye", "just like this") is detected, force trigger segmentation; Output structured segments list: each segment contains "segment ID (unique identifier), coherence score"; Effect: Avoids intent confusion caused by long text, ensures that each segment contains only one core semantic, and improves the accuracy of subsequent intent analysis.

[0044] (3) Intent rough screening module Core function: Quickly narrow down the intent range, generate preliminary intent matching results, sentiment tendency scores, conversation summaries and keywords (to lay the foundation for word cloud generation and accurate analysis).

[0045] Key technical means and working process: 1. Intent list loading: Read the preset intent library, this technical solution sets the application scenario to be related to resource security, so the intent list is set to ["resource security related news", "resource security related policy", "resource security related transaction", "other resource security related conversations", "conversation content unrelated to resource security"]; Effect: Provides clear analysis boundaries for rough screening, avoids large model targetless generation, and improves processing efficiency.

[0046] 2. Large model intent matching and summary: Call large models (such as deepseek, qwen), use Prompt engineering to clarify tasks: Effect: Through the semantic understanding ability of large models, quickly match the intent, and generate concise summary and keywords, reduce the input of subsequent modules.

[0047] 3. Word cloud generation: Call word cloud generation tools (such as Python's wordcloud library), according to "intent classification and aggregation of keywords": combine the keywords of all segments under the same intent, and count the word frequency (keywords with word frequency ≥2 are retained); Generate visual word cloud: use word frequency as font size weight, output PNG format picture; Effect: Intuitively show the core semantics of each intent, which is convenient for users to quickly grasp the key points of the conversation.

[0048] (4) Background search module Core function: dynamically supplement the external background information of segment, solve the problem of "only relying on dialogue text cannot accurately understand the intent".

[0049] Key technical means and working process: 1. Key word extraction: Based on large model technology, select several key words with high value from semantic text segments.

[0050] Prompt word content: """Please extract 3-5 most key words or phrases from the following text, separated by commas: {text} """.

[0051] 2. Multi-source background search: Call search tool: based on MCP tool, call Bing search engine for search; Search parameter setting: return the first 5 related results; Effect: Ensure the relevance of background information, and prioritize internal data to ensure security.

[0052] 3. Background information sorting: Call large model again, Prompt instruction: "Filter the following search results: {Please sort the following background information about "{keywords}" into a coherent text: {background_fragments}}: {segment content} related information, remove redundant / irrelevant content, summarize in simple language (≤100 words), and mark the information source}"; Output structured background results: including "segment ID, keywords, background summary, information source, and search time"; Function: Remove redundant information and provide contextual support for accurate intent analysis (e.g., if a user says "battery life is not enough", the background summary can be supplemented with "this model is rated to have a battery life of 120 minutes".)

[0053] (5) Intent recognition module Core function: Integrate preliminary screening results with background information to output accurate and structured complete intent analysis, resolving the ambiguity of preliminary screening.

[0054] Please see Figure 4 As shown, the overall system timing process is as follows: 1. Startup and Input: Users upload raw text conversation data from multiple sources through the Agent entry point, and the system initializes the process ID; 2. Data Preprocessing: The Agent calls the data preprocessing module to complete data reading → semantic segmentation, and output a list of segments; if it fails, it generates an exception log and terminates. 3. Parallel Analysis: The agent simultaneously calls the intent screening module and the background search module, with both inputs being a list of segments; The intent screening module completes intent matching, summarizing, keyword analysis, and word cloud generation, and outputs the screening results. Background search module: Completes keyword extraction → triggers search → background processing, and outputs background results; 4. Precise Intent Recognition: The Agent receives the coarse screening results and background results, calls the intent recognition module, and outputs a structured and complete intent analysis; 5. Report Generation: The Agent calls the report generation module, inputs the complete intent analysis and word cloud, and generates an MD report; 6. Output: The Agent returns the MD report to the user, and the process ends.

[0055] Key sub-processes: 1. Receive the list of segments distributed by the Agent; 2. Perform large-scale keyword extraction on a single segment; 3. Determine if the keyword contains a proper noun: Yes → Proceed to step 4; No → Output "No background needed" and skip the rest. 4. Call the MCP server to perform the search; 5. Use a large model to process search results and generate background summaries; 6. Output the background results (associated with the segment ID) in the specified format and send them back to the Agent.

[0056] In a specific embodiment of the present invention, the intent recognition process is as follows: Simulated dialogue data: The following is simulated dialogue data between two people regarding resource security (rare earth elements), used as test data: 1. Xiao Li: Xiao Wang, the company is currently promoting a new energy vehicle parts project. Do you know about the uses of rare earth elements in this area? 2. Xiao Wang: Of course I know! Neodymium and dysprosium in rare earth elements are the core materials for making permanent magnet motors, and new energy vehicle drive motors cannot do without them.

[0057] 3. Xiao Li: No wonder rare earth elements are called "industrial vitamins." Besides automobiles, what other fields can't do without them? 4. Xiao Wang: There are too many! Chips, mobile phone screens, military equipment (such as radar and missiles), and even wind turbines all require rare earth elements.

[0058] 5. Xiao Li: Does China have the world's largest rare earth reserves? 6. Xiao Wang: Actually, that's not the case now. Australia and Brazil also have considerable reserves, but China's separation and purification technology is the world's leading.

[0059] 7. Xiao Li: Will we become dependent on imported rare earth elements? 8. Xiao Wang: Some medium and heavy rare earth elements may need to be imported, as domestic reserves of these elements are small and mining costs are high.

[0060] 9. Xiao Li: No wonder the country has been emphasizing the upgrading of the rare earth industry; it truly is a strategic resource! 10. Xiao Wang: That's right! Rare earth recycling technology is receiving more and more attention now. Recycling from waste motors and mobile phones can reduce dependence on imports.

[0061] 11. Xiao Li: Is recycling difficult? What are the costs like? 12. Xiao Wang: In the past, it was difficult and expensive, but now new technologies have improved recycling efficiency and reduced costs.

[0062] 13. Xiao Li: Are there any relevant policies in place? For example, are there any regulations governing mining or recycling? 14. Xiao Wang: Yes! The government controls the total amount of mining and has also introduced recycling subsidy policies to encourage recycling.

[0063] 15. Xiao Li: It seems that for the rare earth industry to develop, both technology and policies need to keep up.

[0064] 16. Xiao Wang: That's right. We need to ensure supply and adopt green mining practices to achieve long-term sustainability.

[0065] 17. Xiao Li: I've learned a lot! I'll need your help again next time I encounter rare earth-related issues in a project. 18. Xiao Wang: Please, let's exchange and learn together! As a preferred embodiment of the present application, the above dialogue information is processed as follows: S1: First, input the data into the agent portal for data preprocessing, and segment according to the semantics into segments Specifically, the data preprocessing module first reads the multi-source multi-language data, and then divides the short text dialogue into different semantic paragraphs by recognizing punctuation marks or semantic similarity for subsequent intent recognition.

[0066] S2: The processed data segments are input into the intent rough screening module and the background search module at the same time, the intent rough screening module forms an intent and a summary, which contains emotional tendency and other contents, and generates a word cloud diagram (as shown below) Figure 6 which shows the important content and keywords in the text.

[0067] S3: The background search module searches the background according to the main content and keywords in the dialogue, and the large model generates a background description.

[0068] The keywords selected by the large model in this embodiment are as follows: rare earth, new energy vehicle, permanent magnet motor, strategic resource, and recycling technology.

[0069] Specifically, according to the aforementioned large model, important background information and key professional terms in the text dialogue data are extracted, and then searched and extracted using the MCP (Model Context Protocol) web search tool. The first n information is given to the large model to sort out the key content.

[0070] S4: The content of the intent rough screening and the background search content are given to the large model to generate complete intent analysis content, and then generate an analysis report.

[0071] Specifically, the results of the intent rough screening and the information obtained by the background search are integrated, and then given to the large model and apply the carefully designed prompt words to let the large model generate a structured intent analysis report.

[0072] Any process or method described in the flowchart or otherwise described herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for implementing specific logic functions (or steps) in the process, and the scope of the preferred embodiments of the present application includes additional implementation in which the functions are performed in different orders, in substantially simultaneous fashion, or in reverse order, according to the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application belong.

[0073] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be realized by software or firmware stored in a memory and executed by a suitable instruction execution device.

[0074] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-mentioned embodiment method can be completed by a program instructing the relevant hardware, and the program can be stored in a computer readable storage medium, and when the program is executed, it includes one or a combination of steps of the embodiment method.

[0075] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0076] In the description of the present specification, the description of the terms "an embodiment", "some embodiments", "an example", "a specific example" or "embodiments" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0077] Although the embodiments of the present application have been shown and described above, it should be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

[0078] The short text oriented multi-level intent recognition intelligent agent processing system, method, device, processor and computer readable storage medium thereof of the present application, for resource security related short text dialogue intent recognition needs, through the whole process design of "data preprocessing-parallel analysis-accurate recognition-structured output", realize the significant technical effect. Among them, the data preprocessing module combines rule segmentation and semantic segmentation accuracy far exceeds the traditional segmentation mode, avoids the fragmentation of intent; the intent rough screening and background search module operates in parallel, the former relies on fine Prompt matching resource security exclusive intent and generates word cloud, the latter based on MCP calls Bing to supplement external background, so that the intent recognition accuracy is greatly improved, especially the ambiguity intent recognition precision is obviously improved. The system processing efficiency is optimized, and the report can be generated within a short time after the call is ended in real-time scenario, which is significantly faster than the traditional scheme; the report generation module outputs MD format document, integrates structured analysis and word cloud, greatly reduces the time of users to obtain core information, at the same time, reduces the AI application threshold of small and medium-sized businesses, promotes the conversion of resource security dialogue data from storage cost to decision asset, and takes into account the technical accuracy, economic cost reduction and industry empowerment value.

[0079] In this specification, the application has been described with reference to specific examples of embodiments thereof. It is, however, apparent that various modifications and changes can be made to the application without departing from the spirit and scope thereof. Accordingly, the specification and drawings should be regarded as illustrative rather than restrictive.

Claims

1. A multi-layered intent recognition intelligent agent processing system for short text dialogues, characterized in that, The system includes: The data processing module is used to preprocess diverse multilingual raw data and segment it into segments according to semantics. The intent coarse screening module, connected to the data processing module, is used to perform intent analysis on the input segment through a large model, so that the large model matches the most relevant intent category from the preset intent candidate set, and makes a detailed summary of the segment content to generate a corresponding word cloud. The background search module is connected to the data processing module and is used to extract keywords or background from the input segment, and to perform background search and organization using the MCP web search tool and the large model. The intent recognition module, connected to the intent coarse screening module and the background search module, is used to input the intent summary content and associated background information into the large model, and use the prompt project to output a structured and complete intent analysis. The report generation module, connected to the intent recognition module, is used to combine the complete intent analysis content generated by intent recognition with word cloud information to generate a standardized intent analysis report.

2. The multi-layered intent recognition intelligent agent processing system for short text dialogues according to claim 1, characterized in that, The data processing module specifically performs the following processing: Read multi-source data files, integrate the acquired data into structured text data, and perform data cleaning to remove non-text information, correct typos and grammatical errors, and generate initial semantic text; Based on a natural language processing model, semantic similarity is calculated on the initial semantic text to determine segmentation points, and intelligent segmentation is performed based on semantic integrity and logical coherence. Output a structured list of segments, each containing a unique identifier (segment ID) and a consistency score.

3. The multi-layered intent recognition intelligent agent processing system for short text dialogues according to claim 2, characterized in that, The intent coarse screening module specifically performs the following processing: Input each segment obtained from the structured segments list, call the large model to read the preset intent candidate set for preliminary intent analysis, and match the most relevant intent category; The large model generates a detailed summary for the current segment, including core information, key viewpoints, and semantic sentiment. The large model is guided to output core vocabulary using the prompt word engineering method, and the core vocabulary is visualized using a word cloud generation tool to generate a corresponding word cloud.

4. The multi-layered intent recognition intelligent agent processing system for short text dialogues according to claim 3, characterized in that, The background search module specifically performs the following processing: The core keywords, entities, and background clues in the segment are extracted using the large model described above; then, based on the MCP tool, a web search tool is called to retrieve the extracted key elements to obtain relevant background knowledge, industry trends, historical data, news hotspots, or similar cases. The large model is used to filter, deduplicate, and integrate the search results, extracting background information that is highly relevant to the current segment. This information is then used to provide external knowledge supplementation for the intent recognition module to gain a deeper understanding, thereby outputting structured background information results.

5. The multi-layered intent recognition intelligent agent processing system for short text dialogues according to claim 4, characterized in that, The intent recognition module specifically performs the following processing: The preliminary intent and summary output by the intent screening module and the related background information provided by the background search module are integrated and input into the large model. The designed prompt project guides the large model to perform deep reasoning and provides the confidence level of intent recognition. Content with low confidence is manually reviewed, and the reasons and potential motivations for the intent are explained in conjunction with background information. The analysis results are output according to a preset fixed structure to form a standardized and directly usable complete intent analysis report segment.

6. The multi-layered intent recognition intelligent agent processing system for short text dialogue according to claim 5, characterized in that, The report generation module specifically performs the following processing: The structured analysis content output by the intent recognition module is fused with the word cloud generated by the intent coarse screening module, and the content is filled in according to the preset report template. Natural language polishing is applied to the structured analysis content to ensure fluent and professional expression; The generated word cloud image will be integrated into the report via links or embeddings, complementing the text content. Finally, format verification is performed to generate documents that conform to the MD standard, making them easy to view, share, and archive.

7. A multi-layered intent recognition intelligent agent processing method for short text dialogue based on the system described above, characterized in that, The method includes the following steps: (1) Users upload original text conversation data from multiple sources through the Agent entry point, and the system initializes the process ID; (2) The Agent calls the data processing module to read data and perform semantic segmentation, and outputs the corresponding list of segments; if the processing fails, an exception log is generated and the process is terminated. (3) The Agent simultaneously calls the intent screening module and the background search module to input the segments list. The intent screening module performs intent matching and summarization, and generates a corresponding word cloud based on the keywords to complete the initial intent screening. The background search module extracts keywords and triggers background information search to output background results. (4) The Agent receives the preliminary intent screening results and background screening results, and calls the intent recognition module to output a structured and complete intent analysis. (5) The Agent calls the report generation module, inputs the complete intent analysis content and word cloud, and generates the MD specification document; (6) The Agent returns the MD specification document to the user, and the process ends.

8. A multi-layered intent recognition intelligent agent processing device for short text dialogues, characterized in that, The device includes: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the multi-level intent recognition agent processing method for short text dialogue as described in claim 7.

9. A processor for multi-layered intent recognition intelligent agent processing for short text dialogues, characterized in that, The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the multi-layered intent recognition agent processing method for short text dialogue as described in claim 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the multi-level intent recognition agent processing method for short text dialogue as described in claim 7.

Images