RAG large model intelligent question answering method and system based on credit investigation tag system enhancement

Abstract

The invention belongs to the technical field of data processing, and particularly relates to an enhanced RAG large model intelligent question-answering method and system based on a credit investigation label system, which can more accurately capture professional vocabularies and semantic relationships in the credit investigation field by introducing a specially designed credit investigation word segmentation model and a credit investigation label system. And the features are converted into vector representations to be stored in a vector database. When a user puts forward a question, the system can quickly position the most relevant text block by matching the question of the user with the tag title vector in the database, so that high-precision information retrieval and answer generation are realized. Therefore, the response speed and accuracy of the intelligent question answering system are improved, the professionality and reliability of answers are ensured, and the problem that intelligent question answering in the credit investigation field is not accurate enough in the prior art is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically relates to a RAG large model intelligent question answering method and system based on an enhanced credit label system. Background Art

[0002] In the current generative AI big model technology, although these models have demonstrated strong language understanding and generation capabilities, they are usually trained on a wide range of general data sets, which leads to obvious limitations in their application in specific professional fields. Especially for fields that require a high degree of expertise and precision, such as the credit reporting industry, existing AI big models often cannot effectively integrate professional knowledge and have difficulty providing accurate and professional answers. This is because credit reporting data has a high degree of timeliness and the interpretation of regulations requires extremely high professionalism, and most products on the market lack optimization for such characteristics.

[0003] A significant technical problem is that the generative AI models in existing technologies cannot accurately understand and handle the special business needs in the credit reporting field. For example, when faced with complex credit inquiries or regulatory interpretations, these models may misunderstand or incompletely understand the terminology in the field, resulting in inaccurate information retrieval or misleading interpretations, thus failing to meet the specific needs of users and even bringing additional risks to customers. Summary of the invention

[0004] The purpose of the present invention is to provide a RAG large-model intelligent question-answering method and system based on an enhanced credit label system, which can provide financial institutions and investors with a full range of regulation interpretation services, corporate information scanning, corporate credit data analysis and credit risk assessment to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention provides, on one hand, a RAG large model intelligent question answering method based on a credit label system enhancement, comprising the following steps: Preprocess large text data obtained from various channels, convert non-text data into text format, remove special characters and irrelevant information, and split long documents into multiple text blocks; Use word segmentation models based on rules and deep learning to extract professional vocabulary, high-frequency vocabulary, core words, and contextual relationships from text blocks; Introduce a credit labeling system that includes professional vocabulary in the credit field, specified characteristic tags, complex semantic tagging, and part-of-speech tagging, and label text blocks with basic tags and credit labels; Based on the results of word segmentation and tagging, a label title is formed, and the text block and label title are converted into vector representation through the sentence embedded model and stored in the vector database.

[0006] Preferably, the method further includes: after obtaining the user's question, using a credit word segmentation model and a credit label system to segment and label the user's question, and associating the label title.

[0007] Preferably, after obtaining the user's question, the user's question is segmented and labeled using a credit word segmentation model and a credit tag system, including: Decompose more complex problems, identify the entities, attributes and relationships therein, generate contextual relationship labels, and store them in the knowledge graph.

[0008] Preferably, the decomposing of the problem with higher complexity includes: The decomposed questions are segmented and labeled twice, and the decomposed questions and their labeled results are vectorized.

[0009] Preferably, the step of performing secondary word segmentation and tagging on the decomposed questions includes: The search vector database is first used to preliminarily screen the matching text block content through keyword matching, and then the most relevant text block is found through semantic similarity calculation.

[0010] Preferably, finding the most relevant text block by calculating semantic similarity includes: The query contents are obtained separately, and the relationships are annotated again according to the possible contextual relationships in the label titles to ensure the consistency and accuracy of the output content.

[0011] Preferably, the further relationship annotation according to the possible contextual relationship in the tag title includes: Integrate and generate the final output content, and correct or delete erroneous information points based on the contextual relationship generated by the label title of the sub-query content.

[0012] Preferably, the correcting or deleting the erroneous information point according to the contextual relationship generated by the tag title of the sub-query content includes: Provide related questions for users to choose from to prevent users from failing to ask the questions they actually want to ask, or guide users to explore related issues in depth.

[0013] On the other hand, the present invention proposes a RAG large model intelligent question answering system based on the credit label system enhancement, comprising: A text preprocessing module, used to perform data preprocessing; A word segmentation and tagging module, used to perform the word segmentation and tagging process; Vector database module, used to store and retrieve vectorized text blocks and label titles; The user interaction module is used to receive user questions and provide the final integrated output content and related questions.

[0014] Technical effects and advantages of the present invention: Compared with the prior art, the present invention proposes a RAG large-model intelligent question-answering method and system based on the enhanced credit label system, which has the following advantages: The present invention can more accurately capture professional vocabulary and semantic relationships in the field of credit investigation by introducing a specially designed credit investigation word segmentation model and credit investigation label system, and convert these features into vector representations and store them in a vector database. When a user asks a question, the system can quickly locate the most relevant text block by matching the user's question with the label title vector in the database, thereby achieving high-precision information retrieval and answer generation. Therefore, the present invention not only improves the response speed and accuracy of the intelligent question-and-answer system, but also ensures the professionalism and reliability of the answer, effectively solving the problem of insufficient precision of intelligent question-and-answer in the field of credit investigation in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a RAG large model intelligent question answering method based on the credit label system enhancement of the present invention; Figure 2 This is a block diagram of a RAG large-model intelligent question-answering system based on an enhanced credit label system according to the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0017] The present invention provides a RAG big model intelligent question-answering method based on the credit label system enhancement, aiming to provide financial institutions and investors with a full range of regulation interpretation services, enterprise information scanning, enterprise credit data analysis and credit risk assessment. Users can communicate with the big model at any time and discuss various topics in the credit field, including credit knowledge answers, credit inquiry services, credit repair opinions, credit application consultation, regulation interpretation, etc. The details are as follows: like Figure 1 As shown, in this embodiment, a RAG large model intelligent question answering method based on the credit label system enhancement includes the following steps: Step 1: Preprocess the large text data obtained from various channels, convert non-text data into text format, remove special characters and irrelevant information, and split long documents into multiple text blocks; Specifically, the large text data obtained from various channels are preprocessed, and the non-text data is converted into text format for processing using ORC, PDF extraction, etc., special characters and irrelevant information are removed, the text format is standardized, the accuracy and cleanliness of the data are ensured, and long documents are divided into multiple manageable text blocks for more efficient processing and retrieval of information.

[0018] In step 1, the large text data obtained from various channels is preprocessed. Specifically, non-text data is converted into text format using technical means such as OCR (optical character recognition) and PDF extraction, and special characters and irrelevant information are removed, while the text format is standardized. This process ensures the accuracy and cleanliness of the original data and improves the quality of subsequent processing. In addition, splitting long documents into multiple manageable text blocks allows information to be processed and retrieved more efficiently.

[0019] The main technical effects of this pretreatment method include: Improve data accuracy: By removing special characters and irrelevant information and standardizing the text format, the impact of noisy data on subsequent processing can be reduced, ensuring the purity and accuracy of the data, which is crucial for generating high-quality analysis results.

[0020] Enhanced processing efficiency: Splitting large text into multiple smaller text blocks allows the system to process these blocks in parallel, thereby speeding up processing. In addition, for a specific query or task, only the relevant text blocks need to be accessed, rather than the entire large document, greatly improving the efficiency of search and retrieval.

[0021] Optimized storage and indexing: The segmented text blocks are easier to manage and index, which helps to build an efficient vector database. This not only saves storage space, but also simplifies data management and maintenance, making it easier to quickly locate and retrieve required information.

[0022] Improved user experience: Preprocessed data is clearer and more structured, allowing users to get more accurate answers when interacting with the intelligent question-answering system. Especially in the field of credit reporting, accurate information processing is directly related to the quality of user decision-making, so this is of great significance to improving user satisfaction and trust.

[0023] In summary, through the careful preprocessing of large text data, the present invention ensures the quality of input data, enhances the processing efficiency and response speed of the system, optimizes the data storage and indexing mechanism, and ultimately realizes a more accurate and faster intelligent question-and-answer service.

[0024] Step 2: Use word segmentation models based on rules and deep learning to extract professional vocabulary, high-frequency vocabulary, core words, and contextual relationships in the text block; In step 2, a word segmentation model based on rules and deep learning is used to extract professional vocabulary, high-frequency vocabulary, core words and contextual relationships in the text block. This hybrid word segmentation method combines the advantages of two different technologies and brings significant technical effects to the intelligent question-answering system in the credit reporting field: Improve professional vocabulary recognition accuracy: Rule-based word segmentation: Through pre-built professional vocabulary and grammar rules (such as forward maximum matching method and reverse maximum matching method), professional terms in the credit reporting field can be efficiently and accurately identified. This method relies on years of accumulated business experience and domain knowledge, and can achieve high-precision word segmentation without extensive training.

[0025] Word segmentation based on deep learning: Using models such as recurrent neural networks (RNN) and long short-term memory networks (LSTM), we can capture complex semantic patterns and contextual information in texts, thereby better processing common vocabulary and sentence structures. This helps to improve the depth of text understanding and ensure that reasonable analysis can be made even when faced with new vocabulary or combinations that have never been seen.

[0026] Enhanced contextual understanding and relevance: The hybrid word segmentation method is not limited to simple word segmentation, it can also identify and extract the core words in the text block and the relationship between them. This enables the system to understand the meaning of words in a wider context, thereby supporting more accurate question matching and answer generation.

[0027] Through deep learning models, we can effectively analyze the implicit connections between words, such as causal relationships, time sequence, etc., further enriching the level of understanding of text content and improving the response quality of the question-answering system.

[0028] Improve information retrieval efficiency: The results of word segmentation can be used directly to build an index or as a feature input for subsequent natural language processing tasks. Accurate word segmentation and extraction of key elements means that relevant information fragments can be located more quickly during queries, reducing unnecessary consumption of computing resources.

[0029] For professional problems in specific fields, the optimized word segmentation model can find the most relevant document paragraphs or sentences more quickly, greatly improving the retrieval speed and relevance of the results.

[0030] Promoting cross-domain application flexibility: Rule-based word segmentation models provide a way to quickly adapt to new fields by simply updating the vocabulary and rule set; while deep learning-based models can apply existing knowledge to new scenarios through methods such as transfer learning, reducing the cost of retraining.

[0031] This flexibility makes the system suitable not only for the credit reporting field, but can also be easily expanded to other industries that require professional knowledge processing, such as legal consulting, medical diagnosis, etc.

[0032] In summary, by combining rule-based and deep learning-based word segmentation models, the present invention significantly improves the ability to recognize professional vocabulary and contextual relationships, enhances the system's contextual understanding and relevance, optimizes information retrieval efficiency, and promotes cross-domain application flexibility. These improvements work together to make the intelligent question-answering system more efficient, accurate, and reliable when dealing with complex problems in the credit reporting field.

[0033] Step 3: Introduce a credit label system that includes professional vocabulary in the credit field, specified characteristic labels, complex semantic annotations, and part-of-speech tagging, and label the text blocks with basic labels and credit labels; this process brings significant technical effects to the intelligent question-answering system: Enhance the accuracy of information retrieval: By labeling text blocks with detailed tags, including but not limited to professional vocabulary, specified features, complex semantics and parts of speech, the system can more accurately locate relevant text content when querying. This not only improves the relevance of search results, but also reduces the interference of irrelevant information.

[0034] The tag system allows the system to quickly match the most appropriate document paragraphs or sentences based on the keywords and concepts in the user's questions, thereby achieving efficient and accurate retrieval.

[0035] Improve context understanding and semantic parsing capabilities: Complex semantic annotation and part-of-speech tagging enable the system to better understand the implicit meaning in the text and the grammatical relationship between words, helping to more accurately capture the user's intentions and question context.

[0036] Especially when dealing with problems involving multiple layers of logic or specific situations, these annotations can assist the system in conducting in-depth analysis and generating answers that are more in line with actual needs.

[0037] Optimize data management and index building: After the introduction of the credit label system, each text block is given a clear label, which is convenient for subsequent data management and maintenance. Label processing simplifies the construction process of the vector database and makes the index structure clearer and easier to access.

[0038] This structured data representation not only saves storage space, but also speeds up queries and improves the overall performance of the system.

[0039] Promote the construction and application of knowledge graphs: The labeled text blocks and their associated tags can be directly used to build a knowledge graph, connecting scattered information points into an organic whole. This helps to explore the potential connections between data and provide deeper knowledge services.

[0040] In the field of credit reporting, the application of knowledge graphs can support complex credit risk assessment, corporate information monitoring and other functions, providing decision-making basis for financial institutions and investors.

[0041] Step 4: Based on the results of word segmentation and tagging, a label title is formed, and the text block and label title are converted into vector representation through the sentence embedded model and stored in the vector database.

[0042] The vector database is based on the results of word segmentation and tagging. It forms label titles for text blocks, integrates text blocks and label titles into interrelated text data, and performs data preprocessing and converts them into serialized vocabulary. It is then converted into a vector representation through a pre-trained sentence embedded model, including the embedding vector of the text block, the embedding vector of the label title, and the combined feature embedding vector of the two, which are stored in the vector database, and a vector index is constructed based on the embedding vector of the label title.

[0043] Step 5: After obtaining the user's question, use the credit word segmentation model and credit label system to segment and label the user's question and associate the label title.

[0044] Specifically, the credit investigation word segmentation model includes rule-based word segmentations such as the forward maximum matching method and the reverse maximum matching method, as well as deep learning-based word segmentations such as recurrent neural networks and long short-term memory networks.

[0045] The rule-based word segmentation is mainly formulated based on the vocabulary formed by professional vocabulary and high-frequency vocabulary involved in the credit reporting field and the corresponding grammatical rules. The forward maximum matching method or the reverse maximum matching method is used according to the actual vocabulary needs. This word segmentation method is highly professional and relies on years of accumulated credit reporting business experience. It does not require a lot of training; while the core vocabulary and contextual relationships in the text and other common vocabulary and rules are identified through deep learning-based word segmentation, which is realized through an open source pre-training model. Through this model, professional vocabulary, high-frequency vocabulary, core words, contextual relationships, etc. in the text block are extracted.

[0046] The credit reporting label system is also organized based on years of experience. It includes professional vocabulary in the credit reporting field, prescribed special labels, complex semantic annotation, part-of-speech tagging, etc. By marking text blocks with multiple basic labels and credit reporting labels, it facilitates subsequent indexing and association.

[0047] The user's questions are segmented and labeled through the credit word segmentation model and the credit label system, and the results are associated with label titles to obtain related questions. The user's questions are converted and expanded using the reasoning engine technology, and the user's questions and related questions are retrieved together to improve the recall of effective information. Specifically, it includes: identifying keywords and concepts through question splitting; deriving new conclusions or expanded queries from known facts and rules through logical reasoning rules; using data mining technology to discover implicit information or needs in user questions, thereby further expanding the scope of queries; and more accurately understanding the semantics and context of user questions, thereby generating more accurate and comprehensive query results.

[0048] On the other hand, the present invention proposes a RAG large model intelligent question answering system based on the credit label system enhancement, such as Figure 2 As shown, including: A text preprocessing module, used to perform data preprocessing steps in a RAG large model intelligent question answering method based on a credit label system enhancement; The word segmentation and tagging module is used to perform the word segmentation and tagging process in a RAG large-model intelligent question-answering method based on the enhanced credit label system; Vector database module, used to store and retrieve vectorized text blocks and label titles.

[0049] The user interaction module is used to receive user questions and provide the final integrated output content and related questions.

[0050] In addition, the above-mentioned text preprocessing module, word segmentation and tagging module, and user interaction module are also used to implement other steps of the above-mentioned RAG large model intelligent question and answer method enhanced by the credit label system during execution, which will not be elaborated here one by one.

[0051] In this embodiment, the generative AI intelligent question-answering process based on the credit label system is as follows: 1: Get user questions; 2: Segment and label user questions and match related questions; 3: Decompose the labeled questions and related questions. Specifically, decompose the more complex questions, identify the entities, attributes and relationships, generate contextual relationship tags, and store them in the knowledge graph. 4: Perform secondary word segmentation and labeling on the decomposed questions, and vectorize the decomposed questions and their labeling results; 5: Search the vector database and perform search enhancement through the two labeling results. The query content converted into a vector is searched in the index structure of the vector database for the label title vector that is most similar to the query vector, so as to quickly locate the corresponding text block. In order to improve the coverage and accuracy of the search, the labeling results are first used to match keywords and preliminarily screen the matching text block content; then the semantic similarity calculation is used to find the text block most relevant to the query. Because the query content mostly involves the credit field, the label title index after labeling based on the credit word segmentation model will be more in line with the actual needs of users in terms of similarity matching.

[0052] 6: Obtain the query content separately, and annotate the relationship again according to the possible contextual relationship in the label title to ensure the consistency and accuracy of the output content.

[0053] 7: Integrate the output content based on the contextual relationship when the question is broken down and the contextual relationship generated by the label title of the sub-query content. If any information points that are inconsistent with the context or incorrect are found in the answer, they need to be corrected or deleted.

[0054] 8: Provide related questions for users to choose from, to prevent users from failing to ask the questions they actually want to ask, or to guide users to explore related questions in depth; specifically, match the label titles according to the labeling results to form related questions, and provide users with additional choices to prevent users from failing to ask the questions they actually want to ask due to a lack of understanding when querying, or to obtain more in-depth questions based on the original questions.

[0055] In summary, the present invention, by introducing a specially designed credit word segmentation model and credit label system, can more accurately capture the professional vocabulary and semantic relationships in the credit field, and convert these features into vector representations and store them in a vector database. When a user asks a question, the system can quickly locate the most relevant text block by matching the user's question with the label title vector in the database, thereby achieving high-precision information retrieval and answer generation. Therefore, the present invention not only improves the response speed and accuracy of the intelligent question-and-answer system, but also ensures the professionalism and reliability of the answer, effectively solving the problem of insufficient precision of intelligent question-and-answer in the credit field in the prior art.

[0056] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A RAG large model intelligent question answering method based on the credit label system enhancement, characterized in that: The following steps are involved: Preprocess large text data obtained from various channels, convert non-text data into text format, remove special characters and irrelevant information, and split long documents into multiple text blocks; Use word segmentation models based on rules and deep learning to extract professional vocabulary, high-frequency vocabulary, core words, and contextual relationships from text blocks; Introduce a credit labeling system that includes professional vocabulary in the credit field, specified characteristic tags, complex semantic tagging, and part-of-speech tagging, and label text blocks with basic tags and credit labels; Based on the results of word segmentation and tagging, a label title is formed, and the text block and label title are converted into vector representation through the sentence embedded model and stored in the vector database.

2. According to claim 1, a RAG large model intelligent question answering method based on the credit label system enhancement is characterized in that: Also includes: After obtaining the user's question, use the credit word segmentation model and credit label system to segment and label the user's question and associate the label title.

3. According to claim 2, a RAG large model intelligent question answering method based on the credit label system enhancement is characterized in that: After obtaining the user's question, the credit word segmentation model and the credit label system are used to segment and label the user's question, including: Decompose more complex problems, identify the entities, attributes and relationships therein, generate contextual relationship labels, and store them in the knowledge graph.

4. According to claim 3, a RAG large model intelligent question answering method based on the credit label system enhancement is characterized in that: The above-mentioned disassembly of more complex problems includes: The decomposed questions are segmented and labeled twice, and the decomposed questions and their labeled results are vectorized.

5. According to claim 4, a RAG large model intelligent question answering method based on the credit label system enhancement is characterized in that: The second word segmentation and tagging of the decomposed questions includes: The search vector database is first used to preliminarily screen the matching text block content through keyword matching, and then the most relevant text block is found through semantic similarity calculation.

6. According to claim 5, a RAG large model intelligent question answering method based on the credit label system enhancement is characterized in that: The method of finding the most relevant text block by calculating semantic similarity includes: The query contents are obtained separately, and the relationships are annotated again according to the possible contextual relationships in the label titles to ensure the consistency and accuracy of the output content.

7. The RAG large model intelligent question answering method based on the credit label system enhancement according to claim 6 is characterized in that: The relationship annotation according to the possible contextual relationship in the tag title includes: Integrate and generate the final output content, and correct or delete erroneous information points based on the contextual relationship generated by the label title of the sub-query content.

8. The RAG large model intelligent question answering method based on the credit label system enhancement according to claim 7 is characterized in that: The correcting or deleting the erroneous information point according to the contextual relationship generated by the label title of the sub-query content includes: Provide related questions for users to choose from to prevent users from failing to ask the questions they actually want to ask, or guide users to explore related issues in depth.

9. A RAG large model intelligent question answering system based on a credit label system enhancement for implementing the method according to any one of claims 1 to 8, characterized in that: include: A text preprocessing module, used to perform data preprocessing; A word segmentation and tagging module, used to perform the word segmentation and tagging process; Vector database module, used to store and retrieve vectorized text blocks and label titles.

10. The RAG large-model intelligent question-answering system based on the credit label system enhancement according to claim 9 is characterized in that: Also includes: The user interaction module is used to receive user questions and provide the final integrated output content and related questions.

Images