Number query method and AI intelligent agent for number query

Abstract

The invention discloses a number query method, an AI intelligent agent for number query and an automatic workflow driven by the AI intelligent agent, and the workflow deeply integrates the judgment capability of a large language model (LLM), the personalized adaptation capability of a user portrait, the fusion capability of multi-source heterogeneous data and a double verification mechanism of number security. Accurate, safe and natural number information service is realized.

Description

Technical Field

[0001] This invention relates to the field of information query technology, specifically to a number query method and an AI agent for number query. Background Technology

[0002] Telephone numbers serve as crucial identifiers connecting individuals and organizations, leading to increasingly frequent needs for retrieval and verification. Whether it's searching for contact information for corporate customer service or public service institutions, or verifying the identity and security of unknown callers, higher demands are placed on the accuracy, timeliness, and convenience of telephone number information services. To meet these needs, various telephone number retrieval solutions have been developed, which can be broadly categorized as follows:

[0003] Communication applications with number identification and tagging capabilities suffer from the following drawbacks: First, the data source is biased and subjective. User tagging is highly subjective and prone to false positives and malicious tagging, making it difficult to guarantee the objectivity and accuracy of query results. For example, a legitimate customer service number might be incorrectly tagged as a "harassment call" due to negative experiences from a few users. Second, query capabilities are limited. The core function of such applications is "passive identification" rather than "active querying." Their query modules typically only support reverse lookups using precise numbers, offering limited support for complex scenarios such as searching numbers by organization name, fuzzy queries, and knowledge-based question-and-answer sessions. Third, there is the problem of data silos. Their data systems are primarily built around user-tagged data, disconnected from authoritative official yellow pages and security authentication data, failing to form a comprehensive, multi-dimensional view of number information. Summary of the Invention

[0004] The main objective of this invention is to provide a number lookup method and an AI agent for number lookup, in order to address the shortcomings of related technologies.

[0005] To achieve the above objectives, according to a first aspect of the present invention, a number query method is provided, comprising: after obtaining the number query request, detecting whether the login account indicated by the query request corresponds to authorization information; if authorization information exists, obtaining user profile data associated with the currently logged-in account; invoking a large language model to perform semantic analysis on the query request to classify it into a preset intent category; when the classified intent category is a preset query number type, determining whether the natural language text of the query request contains a numeric string conforming to a preset number format; if not, performing named entity recognition on the input natural language text to extract the query entity; fusing the user profile data with the key query entity to generate optimized query conditions; retrieving a yellow page number database according to the optimized query conditions to obtain at least one candidate number; submitting the candidate number to an authentication number database for verification; and dynamically generating differentiated natural language response text based on the verification result.

[0006] According to a second aspect of the present invention, an AI agent for number lookup is provided, the AI ​​agent comprising a client and a server, wherein, when the client detects a user's input operation through a human-computer interaction interface, it generates a number lookup request based on the natural language text of the input operation instruction; after the server obtains the number lookup request, it checks whether the currently logged-in account corresponds to authorization information; if authorization information exists, it obtains user profile data associated with the currently logged-in account; it calls a large language model to perform semantic analysis on the lookup request to classify it into a preset intent category; when the classified intent category is a preset lookup number type, it determines whether the natural language text of the lookup request contains a numeric string that conforms to a preset number format; if not, it performs named entity recognition on the input natural language text to extract the query entity; it merges the user profile data with the key query entity to generate optimized query conditions; it searches a yellow page number database according to the optimized query conditions to obtain at least one candidate number; it submits the candidate number to an authentication number database for verification; and it dynamically generates differentiated natural language response text based on the verification result.

[0007] According to a third aspect of the present invention, a computer program product is provided, which, when executed by a processor, implements the method described in the first aspect.

[0008] This embodiment describes a number query method and an AI agent for number querying. The number query method includes: after obtaining the number query request, detecting whether the login account indicated by the query request corresponds to authorization information; if authorization information exists, obtaining user profile data associated with the currently logged-in account; calling a large language model to perform semantic analysis on the query request to classify it into a preset intent category; when the classified intent category is a preset query number type, determining whether the natural language text of the query request contains a numeric string that conforms to a preset number format; if not, performing named entity recognition on the input natural language text to extract the query entity; fusing the user profile data with the key query entity to generate optimized query conditions; searching a yellow page number database according to the optimized query conditions to obtain at least one candidate number; submitting the candidate number to an authentication number database for verification; and dynamically generating differentiated natural language response text based on the verification results. An automated workflow driven by an AI agent deeply integrates the judgment capabilities of a large language model (LLM), the personalized adaptation capabilities of user profiles, the fusion capabilities of multi-source heterogeneous data, and a dual verification mechanism for number security, thereby achieving accurate, secure, and natural number information services. Attached Figure Description

[0009] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 This is a flowchart of the number query method according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0012] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0014] According to an embodiment of the present invention, a number lookup method is provided, such as... Figure 1 As shown, steps 101 to 103 are included below: Because users express their intentions in diverse ways—including colloquialisms, abbreviations, typos, and even implicit intents—a string of numbers might represent a phone number or simply be ordinary numbers, requiring the system to accurately determine its semantics. When a user queries a common or abbreviated name of an institution, such as "Peking Union Medical College Hospital," the system may encounter numerous results from different regions and of varying natures, necessitating the identification of the user's true intent. Traditional systems struggle to understand users' natural language and deeper intentions, and cannot handle complex queries with context (such as "I want to find the phone number of the nearest China Merchants Bank branch") or knowledge-based questions (such as "What do numbers starting with 123 do?"). Accurately determining user intent and avoiding misjudgments or missing potential needs when user input falls between explicit "number-related queries" and "vague questions" presents a significant challenge.

[0015] This embodiment utilizes LLM for intent recognition, which offers the advantage of deep semantic understanding and analysis, surpassing traditional keyword matching and thus more accurately capturing user intent. By fine-tuning the LLM with domain-specific instructions, it can precisely categorize user input into four preset intent categories. This allows the model to better identify specific patterns and expressions related to number queries.

[0016] The predefined intent categories (IntentionAgent, IntentionPhone, IntentionUnclear, IntentionNegative) and their detailed intent descriptions and notes provide a clear basis for LLM judgment.

[0017] Step 101: After obtaining the number query request, check whether the login account indicated by the query request corresponds to authorization information. If authorization information exists, obtain the user profile data associated with the current login account; call the large language model to perform semantic analysis on the query request to classify it into a preset intent category.

[0018] In this step, when a user enters the agent, the agent requests a user ID from an external application (such as a virtual ID mapped from a mobile phone account to protect user data privacy). After obtaining the user ID from the external application, it is stored in the agent's user account system.

[0019] The system obtains the client's number query request, parses the number query request to obtain the natural language text, and determines whether the currently logged-in account of the client has been authorized. If the user has authorized the user, the system synchronously loads the user profile data associated with the currently logged-in account.

[0020] User profile data may include, but is not limited to: user's geographical location, demographic information, historical communication behavior characteristics (such as industry categories with high-frequency interactions), and frequently used applications.

[0021] Furthermore, an intent recognition module based on a Large Language Model (LLM) is invoked. This module, after targeted instruction-tuning, can perform deep semantic understanding and analysis of the user's input natural language text, accurately classifying the intent into one of four preset categories: querying number-related questions, querying questions related to the agent itself, ambiguous questions, and explicit negative questions. Based on the intent recognition classification results, different processing actions are executed.

[0022] Regarding the relationship between intent recognition and fine-tuning of the large model, the main reason for needing fine-tuning is that user input often contains a large number of colloquial expressions (i.e., the commonly used phrases mentioned earlier), making it difficult to determine whether there is an intent to search for a phone number. Therefore, the large model is fine-tuned to convert colloquial expressions (common phrases) into complete entity names. Model fine-tuning only involves positively weighting common descriptions of number affiliation (such as brand, company, occupation, identity), and negatively weighting sensitive words, thus improving the accuracy of the large model in intent determination.

[0023] After conversion, intent recognition is primarily a semantic understanding process, that is, understanding the user's input, such as what the user specifically wants to do with this entity, or whether they need to look up a number. Therefore, the relationship between the large model and intent recognition is that they are different processes, but fine-tuning the large model can significantly improve the accuracy of intent recognition.

[0024] Accurate intent recognition is a prerequisite for implementing subsequent "intelligent disambiguation based on user profile query intent" and the "retrieval-verification" dual security confirmation mechanism. It ensures from the source that the system can understand the user's true needs, thereby providing the user with an extremely accurate and personalized query experience.

[0025] This embodiment combines the intent recognition capabilities of LLM (Limited Language Management) to enable the intelligent agent to "understand" the user's complex questions and trigger different processing flows based on different intents, bringing users a revolutionary interactive experience that is far superior to keyword search or traditional IVR (Interactive Viewership). Through detailed classification and differentiated processing of user intents, it can adaptively handle various query scenarios, fundamentally solving the problems of existing technologies such as single interaction modes and lack of contextual understanding.

[0026] Step 102: When the classified intent category is a preset query number type, determine whether the natural language text of the query request contains a numeric string that conforms to the preset number format.

[0027] In this step, when the user's intent is identified as a query about a number-related question, a number format recognition module (which may use regular expressions, pattern matching, or other technologies) scans the natural language text entered by the user to determine whether it contains numeric strings that conform to preset rules (such as mobile phone numbers, landline numbers, 400 / 95 numbers, and other special numbers).

[0028] Step 103: If not included, perform named entity recognition on the input natural language text to extract the query entity; fuse the user profile data with the key query entity to generate optimized query conditions; search the yellow page number database according to the optimized query conditions to obtain at least one candidate number; submit the candidate number to the authentication number database for verification; dynamically generate differentiated natural language response text based on the verification results.

[0029] In this step, for text that does not contain numeric strings, Named Entity Recognition (NER) is performed to extract key query entities, such as institution names ("Union Hospital") or service types ("China Southern Airlines booking hotline"). The loaded user profile data is then fused with the extracted query entities. User profile data is not limited to geographic location but can also include information based on user history, such as communication habits and industry preferences. For example, when the geographic location in the user profile is detected as "Xi'an," the system intelligently resolves the fuzzy query "Union Hospital" to a high-priority query for "Union Hospitals in the Xi'an area."

[0030] The integration process mainly involves three steps: Step 1, the large model identifies entities based on user input, and these entities are used to analyze which content of the user profile is needed; Step 2, the corresponding user profile information is retrieved from the knowledge graph; Step 3, the large model combines the user input and profile content to create more accurate and optimized query conditions.

[0031] For example: 1. A user searches for "What is the phone number for Peking Union Medical College Hospital?" 2. The large model may first determine that it needs to find information such as the user's "current / local location", "occupation and company". 3. The knowledge graph shows that this user is currently in "Beijing - Chaoyang District", but there is no occupation information. The data is then passed to the large model. 4. The large model cannot determine whether the user is an internal employee of "Peking Union Medical College Hospital", so it defaults to the identity of an external person and provides the contact information of the hospital branch closer to the user, namely the contact number of "Peking Union Medical College Hospital - Dongdan Branch". 5. Possible branching logic: If the knowledge graph previously returned that the current user is not a Chinese citizen, or if the user asked the question in English, the large model may prioritize returning the external contact number of "Peking Union Medical College Hospital - International Medical Department".

[0032] Furthermore, if the fusion result indicates that the user's intent is to query an organization / unit number, the process proceeds to number retrieval and verification. During the retrieval, the query entity, modified based on the user profile, is used to search the internal yellow page number database (a database compiled over a long period, including data from partners and publicly available internet information). This database stores data on organizations or units and their corresponding numbers. By using the modified query conditions, one or more candidate numbers and their associated full names of organizations are obtained, thus improving query accuracy.

[0033] When users perform number searches (especially when providing vague organization names instead of precise numbers), user profile data is used as key context and dynamically integrated with the user's input query entity to intelligently predict, correct, or prioritize the query target. Applying user profile data to the number search process resolves query ambiguity. This method utilizes any profile dimension to improve query accuracy.

[0034] The retrieved candidate numbers are submitted to an independent, highly authoritative "verified number database" for secondary cross-validation. This step aims to confirm the official status and security of the number. Verification failures include: the number not existing in the database, or a significant difference between the name of the verification organization corresponding to the number and the query entity in the optimized query conditions, such as the organization name queried by the user. If the number exists in the database, the verification is successful. Based on the verification results, the process proceeds to two different output stages.

[0035] For example, when a user queries for something that doesn't contain a specific number (such as a query for an organization or service), the system deeply integrates user profile data (such as geographical location, demographic information, and historical communication behavior characteristics) with the query entity entered by the user. For instance, when a user queries "Peking Union Medical College Hospital" and the user profile shows that their place of residence is "Xi'an," the system will intelligently prioritize resolving the query to "Peking Union Medical College Hospital in Xi'an area."

[0036] The data source for the "verified number database" is a long-term, highly reliable database. Its construction integrates multiple authoritative data channels and possesses an internal priority determination mechanism to ensure the accuracy and authority of the output results. Enterprise Number Authentication Service: By partnering with enterprises, we acquire and authenticate their official sets of numbers used for outbound calls (such as customer service hotlines, sales lines, etc.). These number information, which has been verified by the enterprises themselves, is the core component of the "authenticated number database," ensuring the official authenticity of the numbers.

[0037] Third-party platform API data sharing: It connects to the API interfaces of third-party platforms, querying and integrating information such as official contact numbers of specific merchants through real-time calls. This allows the "verified number database" to cover a wider and more dynamic range of commercial numbers verified by authoritative platforms.

[0038] User tagging data (processed): While user tags (such as "sales pitches," "harassing calls") are inherently subjective, this large volume of crowdsourced tagging data is used for statistical analysis and as supplementary input to help assess the "behavioral safety" and "public reputation" of a number. Within the internal construction of the "verified number database," this data may be aggregated, denoised, and rule-based to enrich the number's dimensional information or, in specific circumstances, influence the number's recommendation priority, but ultimately, high authority and official status remain the core. For overlapping numbers from different sources or potentially conflicting information, a pre-defined priority logic determines the final output to the user. For example, officially certified data typically has the highest priority, followed by data shared by authoritative platforms, while user tagging data serves as risk warnings or supplementary information.

[0039] This approach fundamentally solves the technical pain points of traditional search engines returning massive amounts of irrelevant results and human operator inquiry systems requiring multiple rounds of inefficient queries, providing users with a highly personalized and accurate query experience tailored to each individual.

[0040] In one optional implementation of this embodiment, dynamically generating differentiated natural language response text based on the verification result includes: if the verification is successful, packaging the numeric string with the named entity; calling the large language model, inputting the packaged data into the large language model, and outputting a preset identifier type; if the verification fails, packaging the numeric string, the named entity, and the verification failure indicator label; calling the large language model, inputting the packaged data into the large language model, and outputting a cautionary reminder message.

[0041] In this optional implementation, different response paths will be triggered based on the verification result: if the verification is successful, a friendly response with a "certified" or "safe" label will be generated; if the verification fails or there are discrepancies, a response containing a cautionary reminder, clarification, and warning of potential risks will be generated, which can guide the user to make more precise queries and help the user make safer decisions.

[0042] For example, if verification is successful, a friendly response is refined using an AI large-scale model. Specifically, if the number is successfully verified in the authentication database, the system packages the confirmed number information, official organization / unit name, and other data. Subsequently, a large-scale language model (LLM) is invoked to generate a positive, friendly response with "authenticated" or "secure" markings, clearly presented to the user.

[0043] Verification failures include: the retrieved number does not exist in the "authentication number database"; or the number exists in the "authentication number database", but the name of the corresponding authentication institution is significantly different from the name of the institution queried by the user (for example, the user queries "Peking Union Medical College Hospital", but the retrieved number is marked as "Xi'an Lianhu Peking Union Medical College Hospital" in the authentication database).

[0044] If verification fails or discrepancies exist, an AI-powered large-scale language model refines the response to be more user-friendly. Specifically, if number verification fails or a name inconsistency is found (e.g., searching for "Xiehe Hospital" yields "Xi'an Lianhu Xiehe Hospital"), the system packages the retrieved number, possible institution names, and warning labels indicating "unofficially certified" or "name discrepancy." Subsequently, a large-scale language model (LLM) is invoked to generate a user-friendly response containing prudent reminders and clarifications. This response clearly points out potential risks or ambiguities and may proactively ask the user if they need a more precise search, thus guiding them to make a safer decision.

[0045] Furthermore, regarding precise control over wording and information integrity: The wording mainly involves pre-setting "persona" information for the large model, including the "role" the large model plays in this stage (professional, gentle, objective and rational number lookup assistant / customer service), "language style (concise, organized, polite, unemotional, and humane)," and "subjectivity (Teddy's future operation of a vertical dialogue robot specializing in the field of communications and related to telephone number lookup).

[0046] Regarding the completeness of information, the response content is divided into two parts. The first part is the content that must be answered, including the results of the number lookup, the results of the verification in the Teddy information database, the reason for the judgment, and the words to guide the user to further dialogue. The second part is based on the information in the first part, the big model automatically judges the information that needs to be supplemented, and other polite replies to the user with humanistic care.

[0047] In addition, the dialog content boundaries will be set in the pre-set Prompt of the large model to enhance the accuracy of the content. That is: information that the large model does not know will not be said; information that it knows but whose authenticity cannot be determined can be said but must be accompanied by a note that is not entirely credible; sensitive information will not be said; and unnecessary words will not be said.

[0048] The core of this embodiment's closed-loop security mechanism of "search first, verify later" lies in not directly trusting the initial search results, but instead forcibly adding a "reverse verification" step based on a highly authoritative "authentication number database." More importantly, it employs a technical method to dynamically adjust and generate differentiated information content and security warning levels based on the verification results. This endows the method with crucial "judgment" and "prudent reminder" capabilities, transcending simple information display and achieving a leap from "information retrieval" to "intelligent decision support."

[0049] In the "retrieval-verification" dual-number security confirmation mechanism, after a candidate number is retrieved from the internal yellow page number database, it is immediately submitted to an independent, highly authoritative "authentication number database" for secondary cross-verification. This step aims to confirm the official status and security of the number. This design fills the gap in the verification of number authority and security in all existing technical solutions, providing users with a crucial and reliable basis for security decision-making before making calls, effectively reducing the risk of telecommunications violations. This is the core technical support point for the leap from "information retrieval" to "intelligent decision assistance" in this embodiment.

[0050] By employing a dynamic and differentiated response generation method based on verification results, the final response presented to the user is not static but dynamically generated according to the verification results of the aforementioned "authentication number database." Successful verification, failure, or discrepancies will trigger entirely different response logic and script templates. This endows the agent with the ability to "judge" and "provide prudent reminders," enabling it to generate friendly responses with an "authenticated" label, or warning responses containing prudent reminders and clarifications. It can also proactively ask the user if they need more precise queries, thereby guiding them to make safer decisions. This is a key innovation that goes beyond simple information display.

[0051] This innovation effectively fills the gap in existing technology regarding the verification of number authority and security, providing users with a crucial and reliable basis for security decision-making before making calls, and significantly reducing the risk of telecommunications violations.

[0052] As an optional implementation of this embodiment, if included, the parsing process is executed to obtain the multi-dimensional attribute information of the numeric string; the fine-tuned large language model is called for processing to obtain the natural language description containing risk warnings corresponding to the multi-dimensional attribute information.

[0053] In this optional implementation, if a numeric string is included, the number string is extracted precisely from the text. After parsing the string, multi-dimensional attribute information is obtained, and then the LLM is called to output the corresponding natural language description.

[0054] During the parsing process, the phone number string is extracted as the primary key, and query requests are sent in parallel to the three core database clusters: Basic Information Database: Query static attributes of a number, such as its operator and location.

[0055] User Tag Database: Queries and statistically analyzes number tag information (such as "delivery and takeout" and "advertising and promotion") uploaded by users of partner mobile phone manufacturers through crowdsourcing.

[0056] Authoritative and secure database: Search for lists of fraudulent and high-risk phone numbers synchronized with official agencies such as the Ministry of Public Security and the National Anti-Fraud Center.

[0057] Furthermore, the results of the above three types of queries are aggregated to form a multi-dimensional profile of the phone number information. Subsequently, a Large Language Model (LLM) is invoked to transform this structured data into a fluent and easy-to-understand natural language description, containing all key information and necessary risk warnings, and finally presented to the user.

[0058] When fine-tuning here, the LoRA (Low-Rank Adaptation: by freezing the weights of the pre-trained model and only training the parameters of the low-rank matrix) method is used to establish the corresponding relationship between the colloquial terms and the full names of more than 1,000 common entities. For example, "Xiehe (common term)" corresponds to "Xiehe Hospital (full entity name)". When the tokenizer splits out the two characters "Xiehe", and when the Transformer model determines the next character, if the probability of outputting the character "Yi" or the token "Hospital" is greater than 20%, the subsequent output text is forcibly output as "Xiehe Hospital". Such fine-tuning increases the probability of automatically completing the input of common terms into the full entity name by the user, but reduces the possibility of another type of error, that is, in the input of long natural language, misjudgment may occur because the surrounding text happens to be common terms. That is to say, when the large model is fine-tuned here, the main part of the fine-tuning is to increase the weight of the mapping relationship between the names and abbreviations of some entities, so that the large model can better recognize the names and abbreviated information of some entities. For example, the two characters "Xiehe" will strengthen the conjecture weight of "Xiehe Hospital", and combined with the context, it will enhance the accuracy of the abbreviated expressions of such common spoken languages.

[0059] In the method for the fusion and presentation of multi-source heterogeneous number data of this optional implementation manner, when the user directly queries a specific number, it can parallelly call and integrate three types of data sources with completely different natures, namely the basic information from the operator, the marked information from user crowdsourcing, and the security information from official agencies. This enables the user to comprehensively evaluate a number from three key dimensions of "official identity", "folk reputation", and "security rating", providing a multi-dimensional and comprehensive information view, far exceeding the limitations of the existing technology that only provides single and one-sided information.

[0060] As an optional implementation manner of this embodiment, the method further includes: when the classified intention category is a preset question type related to the intelligent agent, calling the large language model to match the corresponding answer from the preset knowledge base. <s

[0061] This application accurately identifies the intention of the "meta-question". For example, the user's question is about the intelligent agent itself or its service provider (such as "Who are you", "Which company provides your service") [My query history]. The LLM can distinguish such self-referential questions from ordinary queries, and match and return the standard answer from the preset official knowledge base, that is, match and return the standard answer from the preset official knowledge base through the LLM.

[0062] As an optional implementation manner of this embodiment, when the classified intention category is a preset type of unclear intention, call the large language model to generate dialogue text.

[0063] This optional implementation can identify and handle ambiguous intents. It leverages the dialogue generation capabilities of LLM to generate friendly, guiding responses, prompting users to provide more information to clarify their query needs. In other words, this implementation intelligently guides the user, rather than simply indicating "I don't understand." In one optional implementation of this embodiment, when the intent category of the classification is a clearly negative question type, a concluding text is generated.

[0064] In this optional implementation, malicious or irrelevant intents are identified and handled securely. When user conversation content clearly does not fall into the aforementioned categories and may be of a negative or inappropriate nature, such intents can be accurately identified, triggering the "Security and Compliance Module" to politely refuse to process the request. The challenge lies in how to quickly and accurately filter out inappropriate content while maintaining the politeness of the conversation and the stability of the system.

[0065] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0066] According to an embodiment of the present invention, an AI agent for number lookup is also provided, comprising a client and a server. When the client detects a user's input operation through a human-computer interaction interface, it generates a number lookup request based on the natural language text of the input operation instruction. Upon receiving the number lookup request, the server checks whether the currently logged-in account has corresponding authorization information. If authorization information exists, it obtains user profile data associated with the currently logged-in account. It then invokes a large language model to perform semantic analysis on the lookup request to classify it into a preset intent category. When the classified intent category is a preset lookup number type, it determines whether the natural language text of the lookup request contains a numeric string conforming to a preset number format. If not, it performs named entity recognition on the input natural language text to extract the query entity. It then merges the user profile data with the key query entity to generate optimized query conditions. Based on the optimized query conditions, it searches a yellow page number database to obtain at least one candidate number. The candidate number is submitted to an authentication number database for verification. Based on the verification result, it dynamically generates differentiated natural language response text.

[0067] According to embodiments of the present invention, the present invention also provides a computer program product that, when executed by a processor, can implement the methods described in any of the above embodiments.

[0068] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0069] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

Claims

1. A number inquiry method, characterized by, The method comprises the following steps: After obtaining the number query request, it is detected whether the login account indicated by the query request corresponds to authorization information. If there is authorization information, the user portrait data associated with the current login account is obtained; A large language model is called to perform semantic analysis on the query request to classify it into a preset intent category; When the classified intent category is a preset query number type, it is judged whether the natural language text of the query request contains a digital string that meets the preset number format; If the judgment result is not contained, the input natural language text is subjected to named entity recognition to extract the query entity; the user portrait data and the key query entity are fused to generate optimized query conditions; According to the optimized query conditions, the yellow page number database is searched to obtain at least one candidate number; The candidate number is submitted to the authentication number library for verification, and a natural language reply text is generated according to the verification result.

2. The number inquiry method according to claim 1, characterized by, If the judgment result is contained, a parsing process is performed to obtain multi-dimensional attribute information of the digital string; a fine-tuned large language model is called to process the multi-dimensional attribute information to obtain a natural language description containing risk prompts.

3. The number inquiry method of claim 1, wherein, The execution of the parsing process includes: querying the basic information library storing static attribute information, the user marking library storing crowd-sourced behavior information, and the security database containing risk information in parallel to obtain the multi-dimensional attribute information of the digital string.

4. The number inquiry method of claim 1, wherein, According to the verification result, a differential natural language reply text is dynamically generated, which includes: if the verification is successful, the digital string and the named entity are packaged; a large language model is called, and the packaged data is input into the large language model to output a preset identification type; If the verification fails, the digital string, the named entity, and the indication label of the failed verification are packaged; a large language model is called, and the packaged data is input into the large language model to output cautious reminder information.

5. The number inquiry method of claim 1, wherein, The method further comprises: when the classified intent category is a preset question type related to an intelligent agent, a large language model is called to match the corresponding answer from a preset knowledge base.

6. The number inquiry method of claim 1, wherein, When the classified intent category is a preset non-explicit intent type, a large language model is called to generate a dialogue text.

7. The number inquiry method of claim 1, wherein, When the classified intent category is an explicit negative question type, an ending text is generated.

8. An AI agent for number query, characterized in that, The AI intelligent agent includes a client and a server, wherein the client detects the input operation of the user through the human-computer interaction interface, and generates a number query request based on the natural language text indicated by the input operation; After the server obtains the number query request, it is detected whether the currently logged-in account corresponds to authorization information. If there is authorization information, the user portrait data associated with the current login account is obtained; a large language model is called to perform semantic analysis on the query request to classify it into a preset intent category; When the classified intent category is a preset query number type, it is judged whether the natural language text of the query request contains a digital string that meets the preset number format; If not included, named entity recognition is performed on the input natural language text to extract query entities; the user portrait data is fused with the key query entities to generate optimized query conditions; the yellow page number database is retrieved according to the optimized query conditions to obtain at least one candidate number; the candidate number is submitted to an authentication number library for verification; and a differentiated natural language reply text is dynamically generated according to the verification result.

9. A computer program product, characterised in that, The computer program, when executed by a processor, implements the method of any one of claims 1-8.

10. A computer readable storage medium characterized by, The computer readable storage medium stores computer instructions for causing the computer to execute the method of any one of claims 1-8.

Images