Financial field information retrieval method and system based on hybrid model

Abstract

The invention is suitable for the technical field of information retrieval, and provides a financial field information retrieval method and system based on a hybrid model. The method comprises: receiving a question and answer request; arranging and filtering by adopting a pre-training sorting model; analyzing important information by using a local algorithm or a small model; identifying the value information by using the middle model; generating a natural language answer through a large language model; and reasoning verification and final output are carried out. Cooperative division of labor can be performed on small and medium models and large models which can be quickly updated, newest financial data and laws and regulations are effectively managed and timely incorporated, so that real-time effectiveness of question and answer content is guaranteed, dependence on frequent training of the large models is reduced by utilizing the characteristics of simple structure and flexible updating of the small and medium models, and the question and answer efficiency is improved. The method greatly improves the absorption efficiency of new knowledge and new data, greatly improves the precision and reliability of processing complex financial problems through the mixed architecture of the model, and reduces the overall operation cost.

Description

Technical Field

[0001] The present invention belongs to the field of information retrieval technology, and in particular relates to a financial field information retrieval method and system based on a hybrid model. Background Art

[0002] With the rapid growth of data and knowledge in the financial industry, financial institutions have accumulated huge knowledge bases and document materials. When users consult financial issues, they often hope to get accurate answers quickly. Current financial question-answering technologies mainly fall into the following categories: 1. Question-answering based on pre-set FAQs or rule bases. This method can provide accurate responses to fixed questions, but is prone to matching failures or giving irrelevant answers to questions with slightly different wording or complex, comprehensive questions, lacking flexibility. 2. Answering questions using a single machine learning or deep learning model. For example, directly using a pre-trained large language model to generate answers. Large language models have powerful language generation and reasoning capabilities, and can understand complex questions and provide coherent answers. However, relying solely on large language models also has obvious shortcomings: due to the lack of real-time external knowledge retrieval, they may provide answers that are inconsistent with the facts. 3. Attempts to combine information retrieval with intelligent question-answering. For example, before a large language model answers a question, relevant content is retrieved from a knowledge base for reference. This retrieval-augmented generation approach improves the accuracy and reliability of answers to a certain extent. However, in practical applications, efficiently finding the most relevant information from massive financial data and fully leveraging the advantages of different types of AI models remains a huge challenge.

[0003] Therefore, the existing technology generally has the following problems:

[0004] (1) Insufficient data freshness and accuracy: In the financial field, market data and policies and regulations are updated very frequently, and the existing RAG system is difficult to achieve both real-time and accuracy.

[0005] (2) Inefficient knowledge updating: Traditional RAG solutions often require retraining of large models or a large number of parameter adjustments, resulting in a long and costly update cycle;

[0006] (3) Insufficient reasoning ability in financial scenarios: Complex financial problems often require multi-step logical reasoning and professional domain knowledge (such as portfolio analysis, risk assessment, etc.);

[0007] (4) Unable to balance system efficiency and cost: Over-reliance on large pre-trained models will result in extremely high computing costs and slow response speeds. Summary of the Invention

[0008] The purpose of the embodiments of the present invention is to provide a financial information retrieval method and system based on a hybrid model, aiming to solve the technical problems existing in the existing technology mentioned in the background technology.

[0009] The embodiment of the present invention is implemented as follows:

[0010] A hybrid model-based financial information retrieval method, the method specifically comprising the following steps:

[0011] Receive user Q&A requests in the financial field, retrieve multiple data sources, obtain multiple candidate information, and aggregate and merge them to obtain an information list;

[0012] Use a pre-trained sorting model to sort and filter multiple candidate information in the information list to generate important information;

[0013] Analyze the important information using a preset local algorithm or small model, extract key elements or potential answers, and extract valuable information from the important information;

[0014] Using a preset model, the value information is identified and key knowledge points or alternative answers are extracted;

[0015] Inputting the value information, the key knowledge points and / or the alternative answers into a preset large language model to generate a natural language answer;

[0016] The natural language answer is reasoned, verified and finally outputted.

[0017] As a further limitation of the technical solution of the embodiment of the present invention, receiving a user's question and answer request regarding the financial field, performing multi-channel data recall through multiple data sources, obtaining multiple candidate information, and aggregating and merging to obtain an information list specifically includes the following steps:

[0018] Receive user Q&A requests regarding the financial field;

[0019] Perform multi-channel data recall processing to obtain multiple candidate information from multiple structured data sources and multiple unstructured data sources;

[0020] A plurality of candidate information are aggregated and merged to generate an information list.

[0021] As a further limitation of the technical solution of the embodiment of the present invention, the aggregating and merging of the plurality of candidate information to generate an information list specifically includes the following steps:

[0022] Obtaining ranking positions of the plurality of candidate information in the corresponding search channel;

[0023] According to the plurality of ranking positions, assigning a reciprocal score to a plurality of corresponding candidate information;

[0024] Accumulating multiple inverse scores corresponding to the multiple candidate information, and comprehensively ranking the multiple candidate information;

[0025] Use weighted methods or machine learning-based ranking models to fuse multiple candidate information for comprehensive ranking and generate an information list.

[0026] As a further limitation of the technical solution of the embodiment of the present invention, the use of a pre-trained sorting model to arrange and filter multiple candidate information in the information list to generate important information specifically includes the following steps:

[0027] Using a pre-trained ranking model, score the relevance and importance of multiple candidate information in the information list and obtain scoring data;

[0028] Arrange information according to the scoring data to generate arrangement information;

[0029] According to the scoring data, the arrangement information is filtered to remove noise or redundant content to generate important information.

[0030] As a further limitation of the technical solution of the embodiment of the present invention, the method of analyzing the important information using a preset local algorithm or small model to extract key elements or potential answers and extracting valuable information from the important information specifically includes the following steps:

[0031] Analyze the important information using a preset local algorithm or small model to extract key elements, wherein the key elements include currency, date and / or company name;

[0032] Based on the key elements, identify the question intent;

[0033] If the question is intended to be a formulatable one, then the rule base is directly called or calculations are performed to provide potential answers;

[0034] Extract valuable information from the important information.

[0035] As a further limitation of the technical solution of the embodiment of the present invention, the use of the preset middle model to identify the value information and extract key knowledge points or alternative answers specifically includes the following steps:

[0036] Extracting key information from the question-and-answer request using a preset entity and relationship extraction model, the key information including company name, fund name, and / or financial indicators;

[0037] By using a preset candidate answer extraction / reading comprehension model, according to the key information, a text paragraph or numerical value that answers the user's question is extracted from the information list to obtain key knowledge points;

[0038] Through a preset sorting and relevance model, the knowledge points are sorted based on semantic similarity or content quality to generate alternative answers.

[0039] As a further limitation of the technical solution of the embodiment of the present invention, inputting the value information, the key knowledge points and / or the alternative answers into a preset large language model to generate a natural language answer specifically includes the following steps:

[0040] Inputting the value information, the key knowledge points and / or the alternative answers into a preset large language model;

[0041] Perform multi-round dialogue fusion processing, store and manage dialogue history, and obtain historical context information;

[0042] Inputting the historical context information into a preset large language model;

[0043] Through the large language model, internal and external knowledge are integrated to generate natural language answers.

[0044] As a further limitation of the technical solution of the embodiment of the present invention, the reasoning verification and final output of the natural language answer specifically include the following steps:

[0045] Based on knowledge rules, the natural language answers are compared and verified with financial industry-specific rules, financial formulas, and regulatory provisions;

[0046] Using a natural language inference model, verify the consistency and contradiction of the natural language answer;

[0047] Obtaining a verification result, and adjusting the natural language answer to generate an adjusted answer;

[0048] The adjusted answer is finally outputted.

[0049] A hybrid model-based information retrieval system for the financial field includes a multi-channel data recall module, an information arrangement and filtering module, a local algorithm processing module, a medium-scale model module, a large language model module, and an answer reasoning and verification module, wherein:

[0050] The multi-channel data retrieval module is used to receive user Q&A requests in the financial field, perform multi-channel data retrieval through multiple data sources, obtain multiple candidate information, and aggregate and merge them to obtain an information list;

[0051] The information arrangement and filtering module is used to arrange and filter multiple candidate information in the information list using a pre-trained sorting model to generate important information;

[0052] A local algorithm processing module is used to analyze the important information using a preset local algorithm or small model, extract key elements or potential answers, and extract valuable information from the important information;

[0053] A medium-scale model module is used to identify the value information using a preset medium-scale model and extract key knowledge points or alternative answers;

[0054] A large language model module, configured to input the value information, the key knowledge points and / or the alternative answers into a preset large language model to generate a natural language answer;

[0055] The answer reasoning and verification module is used to perform reasoning verification and final output on the natural language answer.

[0056] As a further limitation of the technical solution of the embodiment of the present invention, the operation and maintenance verification execution module specifically includes:

[0057] A question-and-answer request receiving unit, configured to receive a user's question-and-answer request regarding the financial field;

[0058] A multi-channel data recall unit is used to perform multi-channel data recall processing and obtain multiple candidate information from multiple structured data sources and multiple unstructured data sources respectively;

[0059] The aggregation and merging unit is used to aggregate and merge multiple candidate information to generate an information list.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The embodiment of the present invention receives question-and-answer requests; employs a pre-trained ranking model for sorting and filtering; uses local algorithms or small models to analyze important information; utilizes a medium model to identify valuable information; generates natural language answers through a large language model; and performs reasoning verification and final output. It is able to collaboratively divide the work between rapidly updateable small and medium-sized models and the large model, effectively managing and promptly incorporating the latest financial data and regulatory developments to ensure the real-time effectiveness of the question-and-answer content. By leveraging the simple structure and flexible updating capabilities of small and medium-sized models, it reduces reliance on frequent large model training, significantly improving the efficiency of absorbing new knowledge and data. Furthermore, the hybrid model architecture significantly enhances the accuracy and reliability of handling complex financial problems and reduces overall operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1A flowchart of a hybrid model-based financial information retrieval method provided by an embodiment of the present invention is shown;

[0063] Figure 2 The application architecture diagram of the financial field information retrieval system based on the hybrid model provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0065] Understandably, existing technologies often have the following problems: (1) Insufficient data freshness and accuracy: In the financial field, market data and policies and regulations are updated very frequently, and existing RAG systems find it difficult to balance real-time performance and accuracy; (2) Inefficient knowledge updating: Traditional RAG solutions often require retraining of large models or a large number of parameter adjustments, resulting in long and costly update cycles; (3) Insufficient reasoning capabilities in financial scenarios: Complex financial problems often require multi-step logical reasoning and professional domain knowledge (such as portfolio analysis, risk assessment, etc.); (4) Inability to balance system efficiency and cost: Over-reliance on large pre-trained models will result in extremely high computing costs and slow response speeds.

[0066] To solve the above problems, an embodiment of the present invention discloses a financial field information retrieval method and system based on a hybrid model. The method receives user question and answer requests about the financial field, performs multi-channel data recall through multiple data sources, obtains multiple candidate information, and aggregates and merges them to obtain an information list; uses a pre-trained sorting model to arrange and filter multiple candidate information in the information list to generate important information; uses a preset local algorithm or small model to analyze the important information, extract key elements or potential answers, and extract valuable information from the important information; uses a preset medium model to identify valuable information and extract key knowledge points or alternative answers; inputs the valuable information, key knowledge points and / or alternative answers into a preset large language model to generate natural language answers; and performs reasoning verification and final output on the natural language answers. It is able to collaborate and divide the work between small and medium-sized models that can be quickly updated and large models, effectively manage and promptly incorporate the latest financial data and regulatory developments, thereby ensuring the real-time effectiveness of question and answer content. By utilizing the simple structure and flexible updating characteristics of small and medium-sized models, it reduces dependence on frequent training of large models and greatly improves the efficiency of absorbing new knowledge and new data. The hybrid architecture of the model greatly improves the accuracy and reliability of handling complex financial problems and reduces overall operating costs.

[0067] Specifically, Figure 1The flowchart of the financial field information retrieval method based on the hybrid model provided by an embodiment of the present invention is shown.

[0068] In a preferred embodiment of the present invention, a financial information retrieval method based on a hybrid model comprises the following steps:

[0069] Step S101: Receive a user's question and answer request regarding the financial field, perform multi-channel data recall through multiple data sources, obtain multiple candidate information, and aggregate and merge them to obtain an information list.

[0070] In an embodiment of the present invention, a user's question and answer request regarding the financial field is received, and multi-channel data recall processing is performed. Multiple candidate information is obtained from multiple structured data sources and multiple unstructured data sources, respectively, and the ranking positions of the multiple candidate information in the corresponding retrieval channels are obtained. According to the multiple ranking positions, the multiple corresponding candidate information are assigned reciprocal scores, and then the multiple reciprocal scores corresponding to the multiple candidate information are accumulated, and the multiple candidate information are comprehensively ranked. Thereafter, a weighted method or a ranking model based on machine learning is used to fuse the multiple candidate information that have been comprehensively ranked to generate an information list.

[0071] It's understandable that multi-channel data retrieval involves applying multiple search methods simultaneously to user queries, obtaining candidate results from different data sources or index types in parallel, and then fusing the results. This approach leverages the complementary strengths of different search technologies to address their respective shortcomings, resulting in more comprehensive and accurate search results. For example, when an intelligent question-answering system is faced with a user's question, it can concurrently query a web search engine, its internal Elasticsearch (keyword search), and a vector database (semantic search) to obtain multiple candidate answers.

[0072] It is understandable that in order to ensure the efficiency of multi-channel retrieval, each channel of retrieval is usually executed in parallel, and parallel retrieval avoids the high latency that may be caused by serial channel-by-channel queries. After the retrieval is completed, the multiple results need to be deduplicated and merged and sorted to generate a unified result list. Commonly used fusion sorting methods include simple and efficient algorithms such as Reciprocal Rank Fusion (RRF). Reciprocal Rank Fusion assigns a reciprocal score based on the ranking position of the results in each retrieval channel (such as 1 point for the first place, 0.5 points for the second place, etc.), and then accumulates the scores of multiple channels for comprehensive ranking. Simple weighting methods (setting weights for results of different channels) or machine learning-based ranking models can also be used to optimize the results during fusion.

[0073] Step S102: using a pre-trained sorting model, arranging and filtering multiple candidate information in the information list to generate important information.

[0074] In an embodiment of the present invention, a pre-trained sorting model is used to score the relevance and importance of multiple candidate information in the information list, obtain scoring data, and then arrange the information according to the scoring data to generate arrangement information. Then, based on the scoring data, the arrangement information is filtered for noise or redundant content to generate important information.

[0075] Step S103: Analyze the important information using a preset local algorithm or small model, extract key elements or potential answers, and extract valuable information from the important information.

[0076] In an embodiment of the present invention, a preset local algorithm or small model is used to analyze important information, extract key elements including currency, date and / or company name, and then identify the question intent based on the key elements, and judge the question intent. When the question intent is a formulatable question, the rule library is directly called or calculations are performed to provide potential answers. At the same time, valuable information is extracted from the important information.

[0077] It is understandable that problems that can be formulated may include: interest rate calculation, financial indicator definition, etc.

[0078] Step S104: using the preset middle model, identify the value information and extract key knowledge points or alternative answers.

[0079] In an embodiment of the present invention, key information such as company name, fund name and / or financial indicators are extracted from the question and answer request through a preset entity and relationship extraction model, and text paragraphs or numerical values ​​that answer the user's question are extracted from the information list according to the key information through a preset candidate answer extraction / reading comprehension model to obtain key knowledge points. The knowledge points are then sorted based on semantic similarity or content quality through a preset sorting and relevance model to generate alternative answers.

[0080] It is understandable that before the large language model is used, screening, refinement and intelligent processing are carried out, and small and medium-sized models significantly improve the accuracy and efficiency of answers. If a reliable answer has been obtained, it can be directly adopted or submitted to the large language model as a reference factor.

[0081] Step S105: input the value information, the key knowledge points and / or the alternative answers into a preset large language model to generate a natural language answer.

[0082] In an embodiment of the present invention, value information, key knowledge points and / or alternative answers are input into a preset large language model, and multi-round dialogue fusion processing is performed, dialogue history is stored and managed, historical context information is obtained, and the historical context information is input into the preset large language model. Through the large language model, internal and external knowledge is integrated to generate natural language answers.

[0083] It's understandable that multi-turn conversation fusion supports multi-turn interactions between users and the system, ensuring contextual consistency by effectively storing and managing conversation history. Common approaches include using sliding window mechanisms or summary memory mechanisms to limit the length of conversation context. For example, a simple sliding window strategy retains only the most recent few turns of conversation for each reply, truncating or summarizing earlier conversations to avoid exceeding the context window limit of large language models. Frameworks such as LangChain provide implementations such as conversation buffer memory and conversation summary memory. Buffer memory retains the complete history but may exceed the token limit, while summary memory significantly reduces the context length occupied by the history by compressing old conversations. There's also window memory, which only retains records from the most recent k turns of conversation. These mechanisms ensure that the model can "remember" previous user questions without disrupting the relevance of responses due to an excessively long history. In multi-turn conversations, users often omit or incomplete their questions. For example, in a financial scenario, a user might first ask, "What is Apple's current stock price?" and then immediately follow up with, "What about the price-to-earnings ratio?" The "P / E ratio" in the second question refers to "Apple" in the previous sentence, omitting the subject. To correctly understand this type of question, the system needs to integrate the conversation history to complete the context. One technical approach is to introduce a question rewriting module, which feeds the current question and the conversation history into a rewriting model to generate a self-contained, complete question. For example, in the above conversation, the system could rewrite "What about the P / E ratio?" to "What is Apple's P / E ratio?" before feeding it into the subsequent question-answering process. Research has shown that training a specialized rewriting model using conversation history can significantly improve context understanding performance. Chai et al. trained a rewriting model in the QURG system to explicitly complete the dependencies between the current question and the context, enabling the model to more fully understand multi-turn queries. This "rewrite first, then answer" framework addresses the problems of omissions and ambiguous references in multi-turn knowledge base question-answering, transforming each question into a context-independent format before being processed by a single-turn question-answering model. At the same time, to enhance knowledge integration in conversations, large-scale language models are often integrated with external knowledge bases or retrieval components (retrieval-augmented generation, or RAG). Specifically, during each round of user questions, the system retrieves relevant information from a financial knowledge base (such as a securities market database, financial news documents, or a company knowledge graph) based on the question content. The retrieval results, along with the conversation history, are then provided to the language model for reference. This allows the model to not only rely on the knowledge contained in its own parameters but also access the latest factual data, alleviating model hallucinations and knowledge timeliness issues.For example, if a user asks, "Is the current P / E ratio of a certain stock higher than its historical average?" the system can query the stock's current P / E ratio and historical average data in real time and append this search evidence to the prompt, allowing the large model to provide an answer based on authoritative data. This knowledge base enhancement makes conversational responses more objective and accurate.

[0084] Understandably, technology plays a key role in improving search recall and comprehension accuracy in financial question-and-answering. Natural language questions from users may be inaccurately worded, overly brief, or lack the proper terminology, preventing the search module from finding relevant answers. To address this issue, before submitting the user's question to the search engine or knowledge base, the question can be rewritten or expanded to make it more semantically specific and cover more synonyms. One specific rewriting method is semantic parsing. Specifically, natural language understanding technology is used to convert the user's question into a structured semantic representation (such as intent and slots, logical form, and SQL query), and then a polished natural language question is generated. For example, for the question "How has this fund performed recently?" the system can parse the subject "this fund" to refer to the name of a fund, the indicator "performance" to refer to yield or net asset value changes, and the timeframe "recently" to refer to the past month. The system then uses a template to rewrite the question into "What is the yield of [Fund Name] in the past month?" for more precise retrieval. Template matching is also a common technique, defining rewriting templates based on common question patterns. For example, if a sentence uses ambiguous terms like "recently," it can be uniformly expanded to include a specific time range. Another example is rewriting "how is the performance" into a more specific question like "increase or decrease" or "yield of return." This rule-based template rewriting leverages domain prior knowledge to align user language with standard questioning, improving retrieval performance. For more open-ended questions, reinforcement learning-based rewriting strategies can automatically learn how to generate retrieval-friendly queries. The "Active Question Answering" (Active QA) framework proposed by Buck et al. inserts an intelligent agent between the user and the question-answering system, repeatedly rewriting the user's question to obtain better answers. This agent interacts with the downstream QA system to learn which rewrites improve answer quality and optimizes the rewriting strategy using policy gradient reinforcement learning. Experiments show that the rewrites learned by this agent are often not simple synonym conversions, but rather automatically adjust keyword weights or word forms, similar to the TF-IDF weighting and stemming techniques used in traditional information retrieval. This means that the RL agent may tend to repeat or emphasize key information terms in the question or use more general synonyms to increase retrieval hit rate. This approach significantly improves the system's ability to recall relevant evidence on complex question-answering datasets (such as the Jeopardy question bank). In addition to RL training, efficient implementation can also be achieved with the help of a small trainable rewriting model. In practice, some studies have used a large pre-trained model (such as GPT-3) to generate high-quality rewriting as a teacher, and then used this as pseudo-labeled data to fine-tune a smaller model (such as T5) to learn the rewriting task. This small model is inserted into the retrieval process as a rewriter: for each user query, a rewritten version of the query is first generated and then sent to the back-end retrieval module. In this way, the rewriting quality of the large model can be achieved without frequently calling the large model.Rewrite models can even be optimized using reinforcement learning (such as rewards based on the relevance of the final search results) to further improve recall. For example, the result sets retrieved from the original query and the rewritten query can be compared. If the rewritten query yields more relevant results, the rewriting model is rewarded positively, thereby training it to generate rewrites that improve recall. Expansion is particularly important in the retrieval phase of financial question-answering. By adding synonyms, industry terms, or hyponyms, we can cover different wordings of the user's question and improve recall. For example, if a user asks, "How are bank stocks performing this year?" a direct search might miss articles containing only specific bank names. Expansion to, "How will the banking sector (e.g., ICBC, CCB, etc.) perform in 2023?" would retrieve more information on the performance of specific bank stocks. Many RAG systems rewrite user queries before retrieval to increase semantic information and improve matching. For example, in the Rewrite-Retrieve-Read process, query rewriting is explicitly included as the first step, and experiments have shown that this improves the relevance of the generated answers. In short, question rewriting technology improves the probability of the retrieval module finding the correct answer in financial question answering by enriching and refining user queries, while also helping the language model understand the question intent more accurately.

[0085] Step S106: perform reasoning verification and final output on the natural language answer.

[0086] In an embodiment of the present invention, based on knowledge rules, the natural language answers are compared and verified with the specific rules, financial formulas and regulatory provisions of the financial industry, and the natural language reasoning model is used to verify the consistency and contradiction of the natural language answers. By obtaining the verification results and adjusting the natural language answers, an adjusted answer is generated, and the adjusted answer is finally output.

[0087] It is understandable that reasoning verification is required to conduct logical reasoning and result review on the preliminary answers generated by large language models to ensure the accuracy and professionalism of the final response. Among them, verification based on knowledge rules: comparing the answers with specific rules, financial formulas, regulatory provisions, etc. of the financial industry to identify and correct errors or non-compliances; verification based on lightweight models: using models such as natural language inference (NLI) to determine whether the "LLM answer + supporting evidence" are consistent with each other. If the confidence level is too low or there is an obvious contradiction, corrections will be made or feedback will be sent back to the upstream module for a new answer, thereby significantly reducing the possibility of factual deviations or logical conflicts in the answers, and the answers can be formatted and optimized in terms of wording as necessary according to industry standards.

[0088] Further, Figure 2 The application architecture diagram of the financial field information retrieval system based on the hybrid model provided by an embodiment of the present invention is shown.

[0089] Specifically, in another preferred embodiment of the present invention, a financial information retrieval system based on a hybrid model includes:

[0090] The multi-channel data recall module 101 is used to receive user question and answer requests about the financial field, perform multi-channel data recall through multiple data sources, obtain multiple candidate information, and aggregate and merge them to obtain an information list.

[0091] In an embodiment of the present invention, the multi-channel data recall module 101 receives a user's question and answer request about the financial field, performs multi-channel data recall processing, obtains multiple candidate information from multiple structured data sources and multiple unstructured data sources, and obtains the ranking positions of the multiple candidate information in the corresponding retrieval channel, assigns reciprocal scores to the multiple corresponding candidate information according to the multiple ranking positions, and then accumulates the multiple reciprocal scores corresponding to the multiple candidate information, and comprehensively sorts the multiple candidate information. Afterwards, a weighted method or a sorting model based on machine learning is used to fuse the multiple candidate information that have been comprehensively sorted to generate an information list.

[0092] The information arrangement and filtering module 102 is used to arrange and filter multiple candidate information in the information list using a pre-trained arrangement model to generate important information.

[0093] In an embodiment of the present invention, the information arrangement and filtering module 102 uses a pre-trained sorting model to score the relevance and importance of multiple candidate information in the information list, obtain scoring data, and then arrange the information according to the scoring data to generate arrangement information, and then filter the arrangement information for noise or redundant content based on the scoring data to generate important information.

[0094] The local algorithm processing module 103 is used to analyze the important information using a preset local algorithm or small model, extract key elements or potential answers, and extract valuable information from the important information.

[0095] In an embodiment of the present invention, the local algorithm processing module 103 uses a preset local algorithm or small model to analyze important information, extract key elements including currency, date and / or company name, and then identify the question intent based on the key elements, and judge the question intent. When the question intent is a formulatable question, it directly calls the rule library or performs calculations to provide potential answers, and at the same time, extracts valuable information from the important information.

[0096] The medium-scale model module 104 is used to identify the value information using a preset medium-scale model and extract key knowledge points or alternative answers.

[0097] In an embodiment of the present invention, the medium-scale model module 104 extracts key information such as company name, fund name and / or financial indicators from the question and answer request through a preset entity and relationship extraction model, and extracts text paragraphs or numerical values ​​that answer the user's question from the information list according to the key information through a preset candidate answer extraction / reading comprehension model to obtain key knowledge points, and then sorts the key knowledge points based on semantic similarity or content quality through a preset sorting and relevance model to generate alternative answers.

[0098] The large language model module 105 is used to input the value information, the key knowledge points and / or the alternative answers into a preset large language model to generate a natural language answer.

[0099] In an embodiment of the present invention, the large language model module 105 inputs value information, key knowledge points and / or alternative answers into a preset large language model, performs multi-round dialogue fusion processing, stores and manages dialogue history, obtains historical context information, and inputs the historical context information into the preset large language model. Through the large language model, internal and external knowledge is integrated to generate natural language answers.

[0100] The answer reasoning and verification module 106 is used to perform reasoning and verification on the natural language answer and to output it in the end.

[0101] In an embodiment of the present invention, the answer reasoning and verification module 106 compares and verifies the natural language answers with the specific rules, financial formulas and regulatory provisions of the financial industry based on knowledge rules, and uses the natural language reasoning model to verify the consistency and contradiction of the natural language answers. By obtaining the verification results and adjusting the natural language answers, an adjusted answer is generated, and the adjusted answer is finally output.

[0102] It should be understood that, although the various steps in the flow chart of each embodiment of the present invention are shown in sequence according to the indication of the arrows, these steps are not necessarily performed in sequence according to the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0103] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0104] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A financial information retrieval method based on a hybrid model, characterized in that: The method specifically comprises the following steps: Receive user Q&A requests in the financial field, retrieve multiple data sources, obtain multiple candidate information, and aggregate and merge them to obtain an information list; Use a pre-trained sorting model to sort and filter multiple candidate information in the information list to generate important information; Analyze the important information using a preset local algorithm or small model, extract key elements or potential answers, and extract valuable information from the important information; Using a preset model, the value information is identified and key knowledge points or alternative answers are extracted; Inputting the value information, the key knowledge points and / or the alternative answers into a preset large language model to generate a natural language answer; The natural language answer is reasoned, verified and finally outputted.

2. The financial information retrieval method based on hybrid model according to claim 1 is characterized in that: The receiving of a user's question-and-answer request regarding the financial field, performing multi-channel data recall through multiple data sources, obtaining multiple candidate information, and aggregating and merging to obtain an information list specifically includes the following steps: Receive user Q&A requests regarding the financial field; Perform multi-channel data recall processing to obtain multiple candidate information from multiple structured data sources and multiple unstructured data sources; A plurality of candidate information are aggregated and merged to generate an information list.

3. The financial information retrieval method based on hybrid model according to claim 2 is characterized in that: The step of aggregating and merging the plurality of candidate information to generate an information list specifically includes the following steps: Obtaining ranking positions of the plurality of candidate information in the corresponding search channel; According to the plurality of ranking positions, assigning a reciprocal score to a plurality of corresponding candidate information; Accumulating multiple inverse scores corresponding to the multiple candidate information, and comprehensively ranking the multiple candidate information; Use weighted methods or machine learning-based ranking models to fuse multiple candidate information for comprehensive ranking and generate an information list.

4. The financial information retrieval method based on hybrid model according to claim 1 is characterized in that: The method of using a pre-trained sorting model to arrange and filter multiple candidate information in the information list to generate important information specifically includes the following steps: Using a pre-trained ranking model, score the relevance and importance of multiple candidate information in the information list and obtain scoring data; Arrange information according to the scoring data to generate arrangement information; According to the scoring data, the arrangement information is filtered to remove noise or redundant content to generate important information.

5. The financial information retrieval method based on hybrid model according to claim 1 is characterized in that: The method of analyzing the important information using a preset local algorithm or small model, extracting key elements or potential answers, and extracting valuable information from the important information specifically includes the following steps: Analyze the important information using a preset local algorithm or small model to extract key elements, wherein the key elements include currency, date and / or company name; Based on the key elements, identify the question intent; If the question is intended to be a formulatable one, then the rule base is directly called or calculations are performed to provide potential answers; Extract valuable information from the important information.

6. The financial information retrieval method based on hybrid model according to claim 1 is characterized in that: The method of using the preset model to identify the value information and extract key knowledge points or alternative answers specifically includes the following steps: Extracting key information from the question-and-answer request using a preset entity and relationship extraction model, the key information including company name, fund name, and / or financial indicators; By using a preset candidate answer extraction / reading comprehension model, according to the key information, a text paragraph or numerical value that answers the user's question is extracted from the information list to obtain key knowledge points; Through a preset sorting and relevance model, the knowledge points are sorted based on semantic similarity or content quality to generate alternative answers.

7. The financial information retrieval method based on hybrid model according to claim 1 is characterized in that: Inputting the value information, the key knowledge points and / or the alternative answers into a preset large language model to generate a natural language answer specifically includes the following steps: Inputting the value information, the key knowledge points and / or the alternative answers into a preset large language model; Perform multi-round dialogue fusion processing, store and manage dialogue history, and obtain historical context information; Inputting the historical context information into a preset large language model; Through the large language model, internal and external knowledge are integrated to generate natural language answers.

8. The financial information retrieval method based on hybrid model according to claim 1 is characterized in that: The reasoning verification and final output of the natural language answer specifically includes the following steps: Based on knowledge rules, the natural language answers are compared and verified with financial industry-specific rules, financial formulas, and regulatory provisions; Using a natural language inference model, verify the consistency and contradiction of the natural language answer; Obtaining a verification result, and adjusting the natural language answer to generate an adjusted answer; The adjusted answer is finally outputted.

9. A financial information retrieval system based on a hybrid model, characterized in that: The system includes a multi-channel data recall module, an information arrangement and filtering module, a local algorithm processing module, a medium-scale model module, a large language model module, and an answer reasoning and verification module, wherein: The multi-channel data retrieval module is used to receive user Q&A requests in the financial field, perform multi-channel data retrieval through multiple data sources, obtain multiple candidate information, and aggregate and merge them to obtain an information list; The information arrangement and filtering module is used to arrange and filter multiple candidate information in the information list using a pre-trained sorting model to generate important information; A local algorithm processing module is used to analyze the important information using a preset local algorithm or small model, extract key elements or potential answers, and extract valuable information from the important information; A medium-scale model module is used to identify the value information using a preset medium-scale model and extract key knowledge points or alternative answers; A large language model module, configured to input the value information, the key knowledge points and / or the alternative answers into a preset large language model to generate a natural language answer; The answer reasoning and verification module is used to perform reasoning verification and final output on the natural language answer.

10. The hybrid model-based financial information retrieval system according to claim 9, characterized in that: The operation and maintenance verification execution module specifically includes: A question-and-answer request receiving unit, configured to receive a user's question-and-answer request regarding the financial field; A multi-channel data recall unit is used to perform multi-channel data recall processing and obtain multiple candidate information from multiple structured data sources and multiple unstructured data sources respectively; The aggregation and merging unit is used to aggregate and merge multiple candidate information to generate an information list.

Images