Working method of college enrollment score line query system based on large model

Through a large-scale model-based college admission score line query system, TTS and problem identification technology are used to identify user problems, combined with information collection and in-depth analysis modules, the problem of difficulty in processing single and complex query functions of the existing system is solved, and efficient and individual The query response improves the user experience.

CN119088923APending Publication Date: 2024-12-06KEXUN JIALIAN INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411153313.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing college admission score line query system has a single function, making it difficult to deal with complex queries and provide personalized recommendations, and cannot quickly understand and respond to users' non-standardized problems.

Method used

采用基于大模型的查询系统,通过TTS技术转换语音咨询为文本,利用问题识别大模型判断问题类别,进入相应流程或FAQ模块进行查询和回答,结合信息收集、向量召回和RAG模块的深入分析, Provide accurate and personalized response.

Benefits of technology

It significantly improves the accuracy and query efficiency of user problem intention recognition, improves the accuracy and flexibility of information collection, improves the user experience, and can quickly respond to complex query needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119088923A_ABST
    Figure CN119088923A_ABST
Patent Text Reader

Abstract

The invention discloses a working method of a college enrollment score line query system based on a large model, relates to the technical field of information processing, and solves the technical problems of low information processing efficiency and low query result accuracy when a current score line query system is used for consulting a large number of examinees. Comprising the steps of performing question category identification on consultation questions by using a large question identification model; when the question is identified as a score line question, multiple rounds of dialogue interaction are utilized, and the information collection module is used for respectively collecting family, specialty and origin information of the customer so as to carry out database query to obtain score line information required by the user; if the identification result is not a score line question, entering an FAQ module, and carrying out vector recall on the question to obtain the most similar question and a similarity score; and for the questions of which the recall scores are lower than the preset threshold value, the questions are deeply analyzed and answered by utilizing RAG retrieval enhancement, so that the complex query requirements of the user are met, and the query efficiency is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a working method of a college admission score query system based on a large model. Background Art

[0002] With the number of college entrance examination candidates hitting record highs, parents and students have different perceptions of colleges and universities, and face difficulties in choosing schools and majors. At the same time, colleges and universities are also faced with the challenge of how to attract high-quality students. Admissions consulting has become the key to competition among colleges and universities; traditional admissions promotional materials, such as brochures and posters, provide basic information about colleges and universities and admission scores. This query method is intuitive and easy to understand, but the information is updated slowly, and it is not easy to quickly retrieve and compare data from different colleges and universities;

[0003] Modern college admissions websites provide online query services, allowing students to access the latest admission score information through the Internet; however, these systems are usually single-function, limited to score query, lack of personalized recommendations and in-depth analysis; FAQ-based systems can quickly respond to students' common questions about admissions, although this method can provide standardized answers, it usually lacks adaptability to complex queries and cannot handle non-standardized questions; based on the above shortcomings, the present invention proposes a working method for a college admissions admission score query system based on a large model. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a working method of a college admissions score query system based on a large model.

[0005] To achieve the above object, the first aspect of the present invention provides a working method of a college admission score query system based on a large model, comprising the following steps:

[0006] Step 1: Candidates or telephone consultants call in, and the system uses TTS technology to convert voice consultation questions into text; then the system uses the question recognition model to identify the question category of the converted text and determine whether the question the user is asking belongs to the "score line question";

[0007] Step 2: When it is identified as a "score line question", the system automatically enters the score line query process. The system and the customer have multiple rounds of dialogues, and the information collection module collects the customer's subject, major and place of origin information respectively; the database is queried based on the collected information to obtain the score line information required by the user;

[0008] Step 3: If the recognition result is not a "score line question", the system enters the FAQ module; in the FAQ module, the system performs vector recall on the question to obtain the most similar question and similarity score;

[0009] Step 4: For questions with a recall score higher than the preset threshold, the system directly returns the answer to the user; for questions with a recall score lower than the preset threshold, the system transfers the question to the RAG module; the RAG module uses retrieval enhancement technology combined with the generation capability of large models to conduct in-depth analysis and answer the questions.

[0010] Furthermore, the question recognition big model is obtained based on the training of the question intention recognition module; the question intention recognition module is used to collect score line related corpus information, construct prompt, organize the corpus and annotations into a big model training corpus, perform model training, and obtain the question recognition big model.

[0011] Furthermore, the specific working steps of the information collection module are as follows:

[0012] S21: Category collection: collect category-related corpus information, clean out the required type enumeration values, build prompts, organize the corpus and annotations into a large model training corpus, and perform model training;

[0013] S22: Professional collection, including:

[0014] Collect professional-related corpus information, clean out labels, build prompts, organize the corpus and annotations into a large model training corpus, and conduct model training;

[0015] Collect professional terms and the standard professional terms corresponding to them, and vectorize all professional terms and terms. After the large model extracts the professional terms, it performs professional vector recall and returns the professional standard terms for professional terms with a similarity greater than 0.9.

[0016] S23: Information collection of students’ place of origin, including:

[0017] Collect professional-related corpus information, clean out labels, build prompts, organize the corpus and annotations into a large model training corpus, and conduct model training;

[0018] Collect address-related data, fine-tune based on the pre-trained BERT model, and train the vertical domain BERT model for address extraction; extract the first four levels of addresses based on the Amap address database for address fuzzy matching and completion to build an address correction model;

[0019] Use a large model to extract the birthplace address information in the user context; use the address extraction BERT model to extract the address, and use the address correction engine to correct the address.

[0020] Furthermore, the specific working steps of the FAQ module are as follows:

[0021] Building a question-and-answer database, which includes a large number of question-and-answer pairs sorted out from business;

[0022] The questions in the question-answer pair are converted into vectors through the vector extraction model, and the question vectors are clustered through the clustering model. The category to which each question belongs and the centroid of the category are recorded respectively;

[0023] The customer's question is converted into a vector through the vector extraction model, and then the similarity between the vector and all centroid vectors is calculated, and the centroid closest to TOPK is selected;

[0024] Score the similarity between the customer question vector and all vectors of the category to which the TOPK centroid belongs, and find the question with the highest similarity score and the corresponding answer.

[0025] Furthermore, the specific working steps of the RAG module are as follows:

[0026] S31: Knowledge information collection: Using web crawler technology, collect knowledge information related to universities from official university websites, education forums, and academic paper libraries;

[0027] S32: Information preprocessing: Clean the collected knowledge information, including removing irrelevant content, correcting erroneous information, and standardizing the format; using natural language processing technology to remove duplicate information;

[0028] S33: Knowledge vectorization: Slice the cleaned knowledge information into paragraphs, convert it into vector representation using natural language processing technology, and store it in the vector database;

[0029] S34: Question Vectorization and Knowledge Retrieval:

[0030] Vectorize the questions raised by users and match them with the vectors in the knowledge base. Use vector similarity calculation to retrieve the top 10 knowledge fragments most relevant to the questions from the knowledge base.

[0031] S35: Knowledge fusion and answer generation: The question context and the recalled knowledge fragments are taken as input and sent to a pre-trained large language model; the input information is fused and inferred to finally generate the answer to the question; the large language model refers to the LLM model.

[0032] Furthermore, the method also includes: performing model distillation and parallel decoding on the problem recognition large model, the RAG large model, and the BERT address extraction model.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The system of the present invention uses a large model for problem identification to identify the problem category of the converted text, and determines whether the problem consulted by the user belongs to the "score line problem", which greatly improves the precision and accuracy of the user's problem intention identification; when it is identified as a "score line problem", the information collection module uses a large model to collect subject categories (subject categories can be enumerated); uses a large model to collect professional statements, and uses a vector library for professional standardization; uses address extraction and address information collection with address correction function to collect candidate information, which greatly improves the accuracy, flexibility and intelligence of information collection;

[0035] 2. In the present invention, if the question category identification result is not a "score line question", the system enters the FAQ module; in the FAQ module, the system performs vector recall on the question to obtain the most similar question and similarity score; for questions with a recall score higher than a preset threshold (for example, 0.95 points), the system directly returns the answer to the question to the user to provide a quick response; for questions with a recall score lower than the preset threshold, the system transfers the question to the RAG module; the RAG module uses retrieval enhancement technology, combined with the generation capability of large models, to conduct in-depth analysis and answers to the questions, so that the system can quickly understand and respond to users' complex query requirements, significantly improve query efficiency, and improve user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 The present invention is a principle block diagram of the working method of the college admissions admission score query system based on a large model.

[0038] Figure 2 It is a workflow diagram of the score line query system in the present invention. DETAILED DESCRIPTION

[0039] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] See also Figure 1 to Figure 2 The first embodiment of the present invention provides a working method of a college admission score query system based on a large model, comprising the following steps:

[0041] Step 1: Candidates or telephone consultants call in, and the system uses TTS technology to convert voice consultation questions into text; then, the system uses the question recognition model to identify the question category of the converted text and determine whether the question the user is asking belongs to the "score line question";

[0042] In this embodiment, the question recognition model is trained based on the question intention recognition module;

[0043] The question intention recognition module is used to collect score line-related corpus information, construct prompts, organize the corpus and annotations into a large model training corpus, perform model training, and obtain a large question recognition model; the present invention uses a large model to greatly improve the precision and accuracy of user question intention recognition;

[0044] Step 2: When it is identified as a "score line question", the system automatically enters the score line query process, and the system and the customer have multiple rounds of dialogue (maximum K rounds), and collect the customer's subject, major and place of origin information through the information collection module; perform a database query based on the collected information to ensure that the user can quickly obtain the required score line information; in this embodiment, K is a preset value;

[0045] In this embodiment, the information collection module uses a large model to collect subject categories (subject categories can be enumerated); uses a large model to collect professional statements, and uses a vector library to normalize professional statements; uses address extraction plus address correction function to collect candidate information, which greatly improves the accuracy, flexibility and intelligence of information collection;

[0046] The specific working steps of the information collection module are as follows:

[0047] S21: Subject collection: Collect subject-related corpus information, clean out the required type enumeration values, such as science and engineering, arts, etc., build prompts, organize the corpus and annotations into a large model training corpus, and conduct model training;

[0048] S22: Professional collection, including:

[0049] Collect professional-related corpus information, clean out labels, such as computer science, geology, etc., build prompts, organize the corpus and annotations into a large model training corpus, and conduct model training;

[0050] Collect professional terms and their corresponding standard majors, such as "geophysics: geophysics", "urban planning: urban and rural planning", etc., and vectorize all majors and professional terms;

[0051] After the big model extracts the profession, it performs professional vector recall and returns the professional standard statement for professional statements with a similarity greater than 0.9;

[0052] S23: Information collection of students’ place of origin, including:

[0053] Collect professional-related corpus information, clean out labels, such as computer science, geology, etc., build prompts, organize corpus and annotations into large model training corpus, and conduct model training; collect address-related data, fine-tune based on the pre-trained BERT model, and train the vertical domain BERT model for address extraction;

[0054] According to the Amap address database, the first four levels of addresses (province, city, district, town) are extracted for address fuzzy matching and completion to build an address correction model;

[0055] Use a large model to extract the place of origin address information in the user context, such as "Fubei Province Wuhan";

[0056] Use the address extraction BERT model to extract addresses, such as "Hubei Wuhan" -> "Province: Hubei Province; City: Wuhan";

[0057] Use the address correction engine to correct the address, such as "Province: Fubei Province; City: Wuhan" -> "Province: Hubei Province; City: Wuhan City";

[0058] The present invention uses large model technology to collect information, which can fully understand the context and semantics, and collect user-related information more intelligently and accurately; the professional recall vector library can improve the generalization of the professional collection process and collect users' professional information more flexibly; the slot correction technology can improve the pass rate of address information collection during user interaction and improve the accuracy of the collected address information;

[0059] Step 3: If the question category recognition result is not a "score line question", the system enters the FAQ module; in the FAQ module, the system performs vector recall on the question to obtain the most similar question and similarity score; in this embodiment, the specific working steps of the FAQ module are as follows:

[0060] Build a question-and-answer database that contains a large number of QA pairs sorted out by business. Use the vector extraction model to convert the questions in the question-and-answer pairs into vectors. Use the clustering model to cluster the question vectors (clustered into N categories) and record the category to which each question belongs and the centroid of the category.

[0061] The customer's question is converted into a vector through a vector extraction model (the same model as above), and then the vector is similar to all centroid vectors and the centroid closest to TOPK is selected;

[0062] Score the similarity between the customer question vector and all vectors of the category to which the TOPK centroid belongs, and find the closest (highest similarity score) question (and the corresponding answer).

[0063] Step 4: For questions with a recall score higher than a preset threshold (e.g., 0.95 points), the system directly returns the answer to the question to the user to provide a quick response;

[0064] For questions with recall scores below a preset threshold, the system forwards the question to the RAG module, which uses retrieval enhancement technology combined with the generation capabilities of large models to conduct in-depth analysis and answer the question.

[0065] In this embodiment, the specific working steps of the RAG module are as follows:

[0066] S31: Knowledge information collection: Using web crawler technology, collect knowledge information related to universities from channels such as university official websites, education forums, and academic paper libraries;

[0067] S32: Information preprocessing: Clean the collected knowledge information, including removing irrelevant content, correcting erroneous information, standardizing the format, etc.; use natural language processing technology to remove duplicate information to ensure the accuracy and uniqueness of the knowledge base;

[0068] S33: Knowledge vectorization: Slice the cleaned knowledge information into paragraphs, convert it into vector representation using natural language processing technology, and store it in a vector database for fast retrieval;

[0069] S34: Question Vectorization and Knowledge Retrieval:

[0070] Vectorize the questions raised by users to match them with the vectors in the knowledge base. Use vector similarity calculation to retrieve the top 10 knowledge fragments most relevant to the questions from the knowledge base.

[0071] S35: Knowledge fusion and answer generation: The question context and the recalled knowledge fragments are used as inputs to a pre-trained large language model; the model's deep learning capabilities are used to fuse and reason about the input information to generate accurate and comprehensive answers to the question; the large language model refers to the LLM model;

[0072] The method also includes: in order to meet the requirements of online use, model distillation and parallel decoding are performed on the large problem recognition model, the large RAG model, and the BERT address extraction model to improve the speed of reasoning;

[0073] Among them, the model distillation technology is applied: migrating large and complex model knowledge into small and more efficient models. This includes extracting key features and knowledge to reduce the complexity and number of parameters of the model while maintaining or approaching the performance of the original model.

[0074] Utilize parallel computing technology: optimize the decoding process of the model; by processing multiple decoding tasks in parallel, the waiting time can be significantly reduced and the overall processing speed can be improved.

[0075] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. 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 invention. In this specification, the schematic representation 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.

[0076] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. Working method of the college admission score query system based on the big model, It is characterized in that The steps include: Step 1: Candidates or telephone consultants call in, and the system uses TTS technology to convert voice consultation questions into text; then the system uses the question recognition model to identify the question category of the converted text and determine whether the question the user is asking belongs to the "score line question"; Step 2: When it is identified as a "score line question", the system automatically enters the score line query process. The system and the customer have multiple rounds of dialogues, and the information collection module collects the customer's subject, major and place of origin information respectively; the database is queried based on the collected information to obtain the score line information required by the user; Step 3: If the recognition result is not "score line question", the system enters the FAQ module; In the FAQ module, the system performs vector recall on questions to obtain the most similar questions and similarity scores; Step 4: For questions with a recall score higher than the preset threshold, the system directly returns the answer to the user; for questions with a recall score lower than the preset threshold, the system transfers the question to the RAG module; the RAG module uses retrieval enhancement technology combined with the generation capability of large models to conduct in-depth analysis and answer the questions.

2. According to the working method of the college admission score query system based on the large model according to claim 1, It is characterized in that The question recognition model is trained based on the question intention recognition module; The question intention recognition module is used to collect score line related corpus information, construct prompts, organize the corpus and annotations into a large model training corpus, perform model training, and obtain a large question recognition model.

3. According to the working method of the large model-based college admission score query system of claim 1, It is characterized in that The specific working steps of the information collection module are as follows: S21: Category collection: collect category-related corpus information, clean out the required type enumeration values, build prompts, organize the corpus and annotations into a large model training corpus, and perform model training; S22: Professional collection, including: Collect professional-related corpus information, clean out labels, build prompts, organize the corpus and annotations into a large model training corpus, and conduct model training; Collect professional terms and the standard professional terms corresponding to them, and vectorize all professional terms and terms. After the large model extracts the professional terms, it performs professional vector recall and returns the professional standard terms for professional terms with a similarity greater than 0.

9. S23: Information collection of students’ place of origin, including: Collect professional-related corpus information, clean out labels, build prompts, organize the corpus and annotations into a large model training corpus, and conduct model training; Collect address-related data, fine-tune based on the pre-trained BERT model, and train the vertical domain BERT model for address extraction; extract the first four levels of addresses based on the Amap address database for address fuzzy matching and completion to build an address correction model; Use a large model to extract the birthplace address information in the user context; use the address extraction BERT model to extract the address, and use the address correction engine to correct the address.

4. According to the working method of the large model-based college admission score query system of claim 1, It is characterized in that The specific working steps of the FAQ module are as follows: Building a question-and-answer database, which includes a large number of question-and-answer pairs sorted out from business; The questions in the question-answer pair are converted into vectors through the vector extraction model, and the question vectors are clustered through the clustering model. The category to which each question belongs and the centroid of the category are recorded respectively; The customer's question is converted into a vector through the vector extraction model, and then the similarity between the vector and all centroid vectors is calculated, and the centroid closest to TOPK is selected; Score the similarity between the customer question vector and all vectors of the category to which the TOPK centroid belongs, and find the question with the highest similarity score and the corresponding answer.

5. According to the working method of the college admission score query system based on the large model as described in claim 1, It is characterized in that The specific working steps of the RAG module are as follows: S31: Knowledge information collection: Using web crawler technology, collect knowledge information related to universities from official university websites, education forums, and academic paper libraries; S32: Information preprocessing: Clean the collected knowledge information, including removing irrelevant content, correcting erroneous information, and standardizing the format; Use natural language processing technology to remove duplicate information; S33: Knowledge vectorization: Slice the cleaned knowledge information into paragraphs, convert it into vector representation using natural language processing technology, and store it in the vector database; S34: Question Vectorization and Knowledge Retrieval: Vectorize the questions raised by users and match them with the vectors in the knowledge base. Use vector similarity calculation to retrieve the top 10 knowledge fragments most relevant to the questions from the knowledge base. S35: Knowledge fusion and answer generation: The question context and the recalled knowledge fragments are taken as input and sent to a pre-trained large language model; the input information is fused and inferred to finally generate the answer to the question; the large language model refers to the LLM model.

6. According to the working method of the large model-based college admission score query system of claim 3, It is characterized in that The method also includes: performing model distillation and parallel decoding on a large problem recognition model, a large RAG model, and a BERT address extraction model.