Intelligent question and answer method and system

Abstract

The invention provides an intelligent question and answer method and system, and the method comprises the steps: S1, receiving a document, and carrying out the preprocessing of the document, including cleaning, keyword scanning extraction, classification and metadata scoring; and S2, in response to questions of the user, performing analysis, vector calculation and similarity calculation on the questions, and outputting a result. According to the method, online business changes can be followed in time, the problem that iteration of a local knowledge base is not timely is solved, in addition, documents are directionally processed through the industry model, the semantic recognition capability of the synonyms is improved, the questioning intention of a user is more accurately grasped, professional business knowledge is returned, and the answering accuracy is greatly improved.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and in particular to an intelligent question-answering method and system. Background Technology

[0002] For industries with rapid business changes, such as gaming, the game publishing and launch process is lengthy and the content integration is complex. The knowledge of various media, distribution, and social networking channels in the industry iterates rapidly, resulting in high learning and usage costs for business users and untimely knowledge updates. Therefore, operations personnel need to frequently operate the game publishing backend, which involves complex operation configuration and check items, as well as complex supporting documentation.

[0003] The existing technological solution is to use a RAG (Retrieval-Augmented Generation) document knowledge base and combine it with an AI Agent robot to answer questions, thereby helping to improve the efficiency of operations integration. At the same time, the knowledge base is self-iterated through R&D documents, which improves the efficiency of daily system maintenance and helps business personnel quickly get started with the business management system.

[0004] Conventional AI agent question-answering robots combine information retrieval and generative modeling techniques. RAGs aim to retrieve relevant information from documents and generate accurate, context-sensitive answers. Existing question-answering robots of this type typically employ key technologies such as vector search, keyword search and hybrid search, and metadata indexing. When combined with the LLM reasoning capabilities of AI agents, these RAGs can plan multi-step tasks, handle complex queries, and optimize answer quality through multiple rounds of retrieval and reflection (such as single-step or multi-step reflection).

[0005] However, most of the document processing for these conventional RAGs is based on length-based block processing, resulting in relatively low accuracy in question-and-answer retrieval, making it difficult to adapt to the complex scenarios of game publishing.

[0006] The existing technology has the following specific problems:

[0007] First, there are numerous media channels to connect with, and the business scenarios are complex: the product functions, business processes, document structures, and application scenarios supported by each media channel are quite different, and the business meanings are also different. There are also cases of overlapping key values ​​for the same access parameter.

[0008] Secondly, all business lines need to connect to a unified management backend, and terminology and keywords need to be processed uniformly to ensure business consistency.

[0009] Third, channel knowledge iterates rapidly, resulting in inconsistencies in version documentation and untimely updates to the local business knowledge base. Summary of the Invention

[0010] To address the above problems, this invention provides an intelligent question-answering method, comprising the following steps:

[0011] S1: Receive documents and preprocess them, including cleaning, keyword scanning and extraction, classification and metadata scoring;

[0012] S2: Respond to user queries, analyze the queries, perform vector calculations and similarity calculations, and output the results.

[0013] Furthermore, in step S1, the cleaning includes: dividing the acquired document into blocks, completing basic cleaning preprocessing, and deleting meaningless words.

[0014] Furthermore, in step S1, keyword scanning and extraction includes matching metadata association similarity across channels, businesses, and applications.

[0015] Furthermore, in step S1, the metadata scoring includes:

[0016] The metadata scoring vector μ is calculated using the metadata scoring matrix δ and the business feature vector θ, and the formula is as follows:

[0017]

[0018] in:

[0019]

[0020] δ is the metadata scoring matrix, n is the number of business line scenarios associated with the metadata, k is the business line relevance score, b is the business field meaning similarity score, rts is the metadata similarity score under the document's business line, and rs is the application business scenario applicability score.

[0021] Furthermore, the metadata scoring also includes calculating the total metadata score (RTS):

[0022]

[0023] Where n is the number of scenarios in which metadata is matched, q is the weight of the application scenario in document parsing, and rts is the similarity score of metadata in the business line of the document.

[0024] Furthermore, step S2 specifically includes:

[0025] S21: For the consultation questions input by business personnel, use the Jieba library to perform word segmentation to obtain the input word vector ω1;

[0026] S22: Generate vocabulary primitive vector ω2 from metadata keywords using the BERT model.

[0027] S23: Perform a dot product operation between the input vocabulary vector ω1, the BERT base vector ω2, and the metadata score vector μ;

[0028] S24: Calculate vector score similarity.

[0029] Furthermore, the similarity is calculated as cosine similarity, using the following formula:

[0030]

[0031] Where: ω1 is the input vocabulary vector, ω2 is the BERT base vector, and μ is the metadata score vector.

[0032] Furthermore, the methods for outputting results include:

[0033] The top K related words were selected by sorting and comparing them according to the similarity score sim.

[0034] Calculate the Spearman relevance coefficient to obtain the final similarity score ρ for word association.

[0035] Based on the similarity score ρ, feedback words are selected, combined and concatenated one by one to generate the final answer statement and output it.

[0036] Furthermore, the method for receiving the document in step S1 includes:

[0037] It receives documents uploaded manually, as well as documents retrieved online via web crawlers;

[0038] Document groups were identified based on three dimensions: channels, business, and applications, and the latest historical versions were obtained.

[0039] Compare and analyze the changes between the latest historical versions and the latest entered versions.

[0040] This application also provides an intelligent question-answering system, including a document preprocessing unit and an information processing unit, wherein:

[0041] The document preprocessing unit receives documents and preprocesses them, including cleaning, keyword scanning and extraction, classification, and metadata scoring.

[0042] The information processing unit responds to user queries, analyzes the queries, performs vector calculations and similarity calculations, and outputs the results.

[0043] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor runs the computer program, the intelligent question-answering method described above is implemented.

[0044] To address the issues of low RAG document processing accuracy, poor business adaptability, untimely knowledge base updates, and inconsistent terminology in existing technologies, this application provides a RAG optimization AI Agent based on a fusion game publishing business feature model. This agent comprises two parts: preprocessing of access channel document metadata scoring and real-time retrieval based on scoring vectorization optimization. It can promptly keep up with online business changes, solving the problem of untimely iteration of the local knowledge base. Furthermore, by using industry-specific models to process documents, it improves the ability to recognize synonyms, more accurately grasps the user's question intent, and returns professional business knowledge, significantly improving the accuracy of the answers. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the intelligent question-answering method according to an embodiment of this application. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0048] This application involves some technical terms, the meanings and concepts of which are as follows:

[0049] AI Agent: An AI Agent is an intelligent system capable of autonomously performing tasks, interacting with its environment, and making decisions based on objectives. It combines artificial intelligence technologies (such as machine learning, natural language processing, and reinforcement learning) to perceive its environment, reason, plan, and take action. It is typically used to automate complex tasks or provide intelligent user interaction. An AI Agent can be a software program, a robot, or other form of intelligent entity, possessing a degree of autonomy and adaptability.

[0050] RAG (Retrievable-Augmented Generation) is an AI framework that combines information retrieval and generative models to improve the performance of Large Language Models (LLMs) when handling complex queries. It retrieves relevant information from external knowledge bases and combines this information with a generative model to output answers, providing more accurate and context-sensitive responses. This significantly improves AI performance when handling complex, domain-specific, or up-to-date information queries, and is widely used in intelligent search, enterprise knowledge management, customer service systems, and other fields.

[0051] Metadata is data that describes data; it's structured data used to provide information about other data. It doesn't contain the actual content of the data, but rather describes its attributes, structure, source, or context, helping users and systems understand, manage, and use the data. Simply put, metadata is like a catalog card in a library, recording information such as the book's title, author, and publication date, but not the book's content. Metadata is typically stored in key-value pairs, tags, or structured formats (such as JSON and XML) and is widely used in databases, file systems, websites, and digital media. Metadata plays a crucial role in the RAG retrieval phase.

[0052] To address the issues of low RAG document processing accuracy, poor business adaptability, untimely knowledge base updates, and inconsistent terminology in existing technologies, this application provides a technical solution for RAG optimization that integrates game publishing business characteristics. This solution improves the retrieval accuracy and business relevance of the AI ​​Agent. Specifically, it is a RAG-optimized AI Agent based on a model that integrates game publishing business characteristics. This agent comprises two parts: preprocessing of access channel document metadata scoring and real-time retrieval based on score-based vectorization optimization. The specific technical solution is as follows:

[0053] S1: Receive documents and preprocess them, including cleaning, keyword scanning and extraction, classification and metadata scoring;

[0054] S2: Respond to user queries, analyze the queries, perform vector calculations and similarity calculations, and output the results.

[0055] Please see Figure 1 The technical solution of this application will be described in conjunction with various preferred embodiments.

[0056] S1: Receive documents and preprocess them, including cleaning, keyword scanning and extraction, classification and metadata scoring.

[0057] The method for receiving documents in step S1 includes:

[0058] It receives documents uploaded manually, as well as documents retrieved online via web crawlers;

[0059] Build a web crawler system that, based on the channel access documents related to the business lines, confirms the online document links. For example, automatically pull the online access documents of various game distribution channels every day and store them in the document database.

[0060] Business personnel can upload offline documents through the backend.

[0061] The system identifies document groups based on three dimensions: channel, business, and application, and retrieves the latest historical versions. It automatically recognizes the channel, business, and application dimensions to which a document belongs and assigns it to the corresponding document group.

[0062] The system compares and analyzes changes between the latest historical versions and the latest entered versions. It regularly checks the latest historical versions and new versions of each document group, and makes a preliminary judgment on whether changes have been made by comparing file size and MD5 value. For documents with changes, a detailed content difference analysis is performed.

[0063] In step S1, the cleaning process includes: dividing the acquired document into blocks, completing basic cleaning preprocessing, and deleting meaningless words, such as removing spaces, garbled characters, and other meaningless content.

[0064] The specific method for segmentation is to segment according to a specified identifier, which can be combined with the actual document format to confirm the identifier to be applied.

[0065] After cleaning, the documents are scanned using a keyword extraction algorithm, and the extracted keywords are matched with the metadata dictionary to calculate the similarity in terms of channel, business, and application dimensions.

[0066] Considering the actual business situation, due to the complexity of the distribution business, there are significant differences in the naming of parameters defined on the distribution side in various external channel documents, and there are cases of overlapping naming of parameter keys. Therefore, it is necessary to unify the distribution docking model and perform unified metadata processing for parameters of various channels to facilitate vectorized search and matching during document preprocessing and RAG embedding.

[0067] This application proposes a preferred implementation method for calculating vector score similarity. First, a scoring matrix metadata score δ is constructed. Then, a metadata score vector μ and a total score RTS are calculated using a metadata evaluation model θ. Finally, the document is segmented according to the RTS value to generate business relevance tags. Wherein:

[0068] The formula for the rating matrix metadata rating δ is as follows:

[0069]

[0070] Where: δ is the metadata scoring matrix, n is the number of business line scenarios associated with the metadata, k is the business line relevance score, b is the business field meaning similarity score, and rs is the application business scenario applicability score.

[0071] The metadata evaluation dimension vectorization model is as follows:

[0072]

[0073] As can be seen from the above formula, this application is divided into three dimensions: the business line where the metadata is located (K), the meaning of the metadata explanation (B), and the application scenario of the business line corresponding to the metadata (RS).

[0074] The metadata scoring vector μ is calculated using the metadata scoring matrix δ and the business feature vector θ, and the formula is as follows:

[0075]

[0076] The final calculation formula for the total metadata score RTS is as follows:

[0077]

[0078] Where n is the number of scenarios for metadata matching applications, q is the weight of the application scenario in document parsing, and the specific value of the weight can be dynamically determined according to the coverage of business scope, and rts is the similarity score of metadata under the business line of the document.

[0079] S2: Respond to user queries, analyze the queries, perform vector calculations and similarity calculations, and output the results.

[0080] After completing the RAG document parsing and processing described above, when business users use the AI ​​Agent to ask questions, the AI ​​Agent can optimize the retrieval based on the user's input question, combining the RAG with the metadata scoring vector μ, thereby returning the best answer. The specific steps include:

[0081] S21: For the consultation questions input by business personnel, the Jieba library is used for analysis to obtain the input word vector ω1;

[0082] S22: Generate vocabulary basis vector ω2 from metadata keywords using the BERT model;

[0083] S23: Perform a dot product operation between the input vocabulary vector ω1, the BERT base vector ω2, and the metadata score vector μ;

[0084] S24: Calculate the vector score similarity. The similarity is calculated using cosine similarity, and the formula is:

[0085]

[0086] After obtaining the vector score similarity, the results are output. Specific methods for outputting the results include:

[0087] The top K related words were selected by sorting and comparing them according to the similarity score sim.

[0088] Calculate the Spearman relevance coefficient to obtain the final similarity score ρ for word association.

[0089] Based on the similarity score ρ, feedback words are selected, combined and concatenated one by one to generate the final answer statement and output it.

[0090] Therefore, the major technical effects brought about by this application are as follows:

[0091] 1. High timeliness of knowledge base updates: The system retrieves all related channel data changes on a daily basis, keeping up with online business changes in a timely manner and solving the problem of untimely iteration of the local knowledge base;

[0092] 2. High recognition rate of professional terms: By using industry models to process documents in a targeted manner, the semantic recognition of synonyms is improved, the intent of user questions is grasped more accurately, and professional business knowledge is returned;

[0093] 3. Low training cost of RAG knowledge base: Relying on the vectorized reference of professional models, the data volume of business documents can be controlled, reducing the cost of knowledge base construction and maintenance;

[0094] 4. Strong business adaptability: By grouping documents in three dimensions and unifying metadata processing, problems such as inconsistent terminology and overlapping parameter naming are solved, ensuring business consistency.

[0095] This application also provides an intelligent question-answering system, including a document preprocessing unit and an information processing unit, wherein:

[0096] The document preprocessing unit receives documents and preprocesses them, including cleaning, keyword scanning and extraction, classification, and metadata scoring.

[0097] The information processing unit responds to user queries, analyzes the queries, performs vector calculations and similarity calculations, and outputs the results.

[0098] This application also provides a computer storage medium that can store an executable program, wherein when the executable program is run on a computer, the computer executes the intelligent question-answering method.

[0099] This application also provides an electronic device, which includes a processor and a memory, the memory for storing an executable program, and the processor for executing the executable program to implement the intelligent question-answering method.

[0100] It should be noted that those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a computer program instructing related hardware. The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent question-answering method, characterized in that, Includes the following steps: S1: Receive documents and preprocess them, including cleaning, keyword scanning and extraction, classification and metadata scoring; S2: Respond to user queries, analyze the queries, perform vector calculations and similarity calculations, and output the results.

2. The intelligent question-answering method as described in claim 1, characterized in that, In step S1, the cleaning process includes: dividing the acquired documents into blocks, completing basic cleaning preprocessing, and deleting meaningless words.

3. The intelligent question-answering method as described in claim 1, characterized in that, In step S1, keyword scanning and extraction includes matching metadata association similarity across channels, businesses, and applications.

4. The intelligent question-answering method as described in claim 1, characterized in that, In step S1, the metadata scoring includes: The metadata scoring vector μ is calculated using the metadata scoring matrix δ and the business feature vector θ, and the formula is as follows: in: δ is the metadata scoring matrix, n is the number of business line scenarios associated with the metadata, k is the business line relevance score, b is the business field meaning similarity score, rts is the metadata similarity score under the document's business line, and rs is the application business scenario applicability score.

5. The intelligent question-answering method as described in claim 4, characterized in that, The metadata scoring also includes calculating the total metadata score (RTS): Where n is the number of scenarios for metadata matching applications, and q is the weight of the application scenario in document parsing.

6. The intelligent question-answering method as described in claim 1, characterized in that, Step S2 specifically includes: S21: For the consultation questions input by business personnel, use the Jieba library to perform word segmentation to obtain the input word vector ω1; S22: Generate vocabulary primitive vector ω2 from metadata keywords using the BERT model. S23: Perform a dot product operation between the input vocabulary vector ω1, the BERT base vector ω2, and the metadata score vector μ; S24: Calculate vector score similarity.

7. The intelligent question-answering method as described in claim 6, characterized in that, The similarity is calculated using cosine similarity, and the formula is: Where: ω1 is the input vocabulary vector, ω2 is the BERT base vector, and μ is the metadata score vector.

8. The intelligent question-answering method as described in claim 1, characterized in that, Methods for outputting results include: The top K related words were selected by sorting and comparing them according to the similarity score sim. Calculate the Spearman relevance coefficient to obtain the final similarity score ρ for word association. Based on the similarity score ρ, feedback words are selected, combined and concatenated one by one to generate the final answer statement and output it.

9. The intelligent question-answering method as described in claim 1, characterized in that, The method for receiving documents in step S1 includes: It receives documents uploaded manually, as well as documents retrieved online via web crawlers; Document groups were identified based on three dimensions: channels, business, and applications, and the latest historical versions were obtained. Compare and analyze the changes between the latest historical versions and the latest entered versions.

10. An intelligent question-answering system, characterized in that, It includes a document preprocessing unit and an information processing unit, wherein: The document preprocessing unit receives documents and preprocesses them, including cleaning, keyword scanning and extraction, classification, and metadata scoring. The information processing unit responds to user queries, analyzes the queries, performs vector calculations and similarity calculations, and outputs the results.

Images