Intelligent question and answer method and device, equipment, storage medium and program product

Abstract

The invention provides an intelligent question answering method and device, equipment, a storage medium and a program product, and relates to the field of financial science and technology. The method comprises the following steps: performing retrieval in a preset bank field document library to obtain a retrieval result list corresponding to each query in a question set generated based on user input; wherein the retrieval result list comprises a plurality of documents sorted according to similarity; according to document sorting and similarity information of each retrieval result list, dynamically optimizing fusion parameters of a fusion sorting algorithm; fusing and reordering documents in each retrieval result list based on a fusion ordering algorithm by adopting the dynamically optimized fusion parameters to obtain a reordered target retrieval result list; and generating a target answer corresponding to the user input based on the target retrieval result list. According to the method, the intelligent question and answer effect can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of financial technology, and in particular to an intelligent question-answering method and device, equipment, a storage medium and a program product. BACKGROUND

[0002] There are a large number of business systems, policy documents and operation specifications in the daily operation of banks and other financial institutions. When serving customers, branch staff need to frequently query these documents to answer inquiries and process procedures. With the complexity of business and the diversification of customer needs, traditional manual searching methods are inefficient and difficult to quickly respond to customer needs.

[0003] In this context, intelligent question-answering systems have gradually become an important tool for improving the service efficiency of bank branches. Currently, retrieval-augmented generation (RAG) technology is widely used in intelligent question-answering systems. It combines large models with external knowledge bases to retrieve relevant knowledge and assist in generating answers after receiving user queries. Specifically, the system retrieves relevant documents from multiple paths in the external knowledge base based on user queries, obtains multiple sets of retrieval results, then simply merges or sorts the results according to fixed rules, and inputs the sorted results and user queries into a large model to generate the final answer.

[0004] However, the above intelligent question-answering system has the defect of being difficult to ensure the accuracy of the results after fusion when fusing multiple source retrieval results, thereby affecting the effect of intelligent question-answering. SUMMARY

[0005] The present application provides an intelligent question-answering method, device, equipment, storage medium and program product to improve the accuracy of result fusion and thereby improve the effect of intelligent question-answering.

[0006] In a first aspect, the present application provides an intelligent question-answering method, which comprises:

[0007] Retrieving in a pre-set bank domain document library to obtain a retrieval result list corresponding to each query in the question set generated based on user input; wherein the retrieval result list comprises multiple documents sorted by similarity;

[0008] According to the document sorting and similarity information of each retrieval result list, dynamically optimizing the fusion parameters of the fusion sorting algorithm;

[0009] Using the dynamically optimized fusion parameters, the documents in each retrieval result list are fused and re-ordered based on the fusion sorting algorithm to obtain a target retrieval result list after re-ordering;

[0010] generate a target answer corresponding to the user input based on the target retrieval result list.

[0011] In a second aspect, the present application provides an intelligent question-answering device, which comprises:

[0012] a retrieval module configured to perform retrieval in a preset bank domain document library based on vector similarity to obtain a retrieval result list corresponding to each query in a question set generated based on user input; wherein the retrieval result list comprises a plurality of documents sorted by similarity

[0013] an optimization module configured to dynamically optimize fusion parameters of a fusion sorting algorithm according to document sorting and similarity information of each retrieval result list;

[0014] a fusion module configured to perform fusion and re-sorting of documents in each retrieval result list based on the fusion sorting algorithm using the dynamically optimized fusion parameters to obtain a target retrieval result list after re-sorting;

[0015] a generation module configured to generate a target answer corresponding to the user input based on the target retrieval result list.

[0016] In a third aspect, the present application provides an electronic device, which comprises at least one processor and a memory connected in communication with the processor.

[0017] The memory stores computer-executable instructions.

[0018] The at least one processor executes the computer-executable instructions stored in the memory to implement the method according to any one of the first aspect.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium, which stores computer-executable instructions, and the computer-executable instructions are executed by a processor to implement the method according to any one of the first aspect.

[0020] In a fifth aspect, the present application provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the method according to any one of the first aspect.

[0021] The intelligent question answering method provided in the application is executed by any electronic device, and is used for realizing intelligent question answering service in a bank point scene, and assisting staff to efficiently answer customer inquiries. Specifically, the electronic device receives user input, performs retrieval in a preset bank domain document library based on vector similarity, and obtains a retrieval result list corresponding to each query in a question set generated based on the user input. Then, the fusion parameter of the fusion sorting algorithm is dynamically optimized according to the document sorting and similarity information of each retrieval result list, and the target retrieval result list is obtained by fusing and reordering the documents in each retrieval result list based on the fusion sorting algorithm using the dynamically optimized fusion parameter. Finally, the target answer corresponding to the user input is generated based on the target retrieval result list. In this process, since the question set is generated based on the user input, the retrieval dimension is widened, more potential intentions are covered, and the comprehensiveness of the retrieval result is improved. At the same time, the fusion parameter is dynamically optimized according to the document sorting and similarity information in the retrieval result list, the rationality of the weight distribution of each retrieval result list during fusion is ensured, the target retrieval result list after final reordering is more accurate, and the intelligent question answering effect is improved. BRIEF DESCRIPTION OF DRAWINGS

[0022] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the application and, together with the description, serve to explain the principles of the application.

[0023] Figure 1 An application scene schematic diagram of an intelligent question answering method provided by an embodiment of the application;

[0024] Figure 2 A flowchart of an intelligent question answering method provided by an embodiment of the application Figure 1 ;

[0025] Figure 3 A flowchart of an intelligent question answering method provided by an embodiment of the application Figure 2 ;

[0026] Figure 4 A structure schematic diagram of an intelligent question answering device provided by an embodiment of the application;

[0027] Figure 5 A structure schematic diagram of an electronic device provided by an embodiment of the application.

[0028] Through the above drawings, the specific embodiments of the application have been shown, and will be described in more detail hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of the application by any means, but to illustrate the concept of the application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0029] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0030] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0031] It should be noted that the intelligent question-answering method, device, equipment, storage medium and program products provided in this application can be used in the field of fintech, or in any field other than fintech. The application fields of the intelligent question-answering method, device, equipment, storage medium and program products in this application are not limited.

[0032] In the daily operations of banks and other financial institutions, there are numerous documents such as business regulations, policy documents, and operational procedures. Staff members need to spend a significant amount of time searching, reading, abstracting, and summarizing these documents. Furthermore, in bank branch settings, branch staff frequently need to consult these documents to answer business inquiries and process business procedures when providing services to customers.

[0033] As banking operations become more complex and customer needs more diverse, traditional methods relying on manual document retrieval are inefficient and fail to meet the demands for rapid customer response. Against this backdrop, intelligent question-and-answer systems are gradually becoming an important tool for improving the efficiency of bank branch services. These systems provide branch staff with auxiliary capabilities for customer service, policy retrieval, and understanding of business rules, shortening knowledge retrieval time and reducing passive waiting time for users. This, in turn, improves the efficiency and compliance of frontline staff in customer service, leading to higher-quality service.

[0034] Currently, Retrieval-Augmented Generation (RAG) technology is widely used in intelligent question-answering systems. It combines a large model with an external knowledge base to retrieve relevant knowledge and assist in generating an answer after receiving a user query. Specifically, a RAG-based intelligent question-answering system retrieves relevant documents from multiple paths in an external knowledge base based on the user query, obtaining multiple sets of search results. These results are then merged and reordered, and the reordered results are input into a large model along with the user query to generate the final answer.

[0035] However, the aforementioned intelligent question-answering system has the drawback of difficulty in ensuring the accuracy of the fused results when integrating multi-source retrieval results. In addition, when the customer's input question is unclear or incomplete, or when a question depends on multiple sub-questions, or when it is too colloquial, it will also lead to inaccurate results, thereby affecting the effectiveness of intelligent question answering.

[0036] Therefore, this application provides an intelligent question-answering method, apparatus, device, storage medium, and program product, aiming to solve the above-mentioned technical problems of known technologies. Specifically, the intelligent question-answering method of this application proposes to first generate multiple relevant queries based on user input to form a question set; then, for each query, search in a bank document database to obtain a corresponding sorted document list; then, dynamically adjust the parameters of the fusion sorting based on the sorting and similarity of the documents in these lists; finally, merge and re-sort all lists using the adjusted parameters to obtain the most relevant document list, and then generate an answer based on this list.

[0037] It should be understood that the intelligent question-answering method of this application can be used in any scenario that requires intelligent question-answering based on a domain document library, as long as the domain document library can be changed. For example, Figure 1 This is a schematic diagram illustrating an application scenario of an intelligent question-answering method provided in an embodiment of this application. For example... Figure 1 As shown, the method of this application can be used in the service scenarios of bank branches.

[0038] Specifically, in this scenario, the electronic device is the terminal equipment at a bank branch, and the banking-related document library is stored on a cloud server. When the terminal equipment receives user input, it queries the banking-related document library, searching through banking-related documents based on vector similarity, and obtains a list of search results corresponding to each query in the question set generated based on the user input. Subsequently, the terminal equipment dynamically optimizes the fusion parameters used in the fusion ranking algorithm based on the document order and similarity information of each search result list, and re-ranks the documents in each search result list using the fusion ranking algorithm based on these parameters to obtain the target search result list. Finally, based on this target search result list, the target answer is fed back to the user.

[0039] In the above process, the ability to quickly and accurately provide business consultation answers to bank branch staff reduces the time spent manually searching documents and improves customer service efficiency. At the same time, the dynamic optimization of the fusion and sorting ensures the accuracy of the answers and reduces business processing errors, thus effectively improving the intelligent question-and-answer effect.

[0040] In this embodiment, the aforementioned electronic device can also be an intelligent customer robot for user terminals. Those skilled in the art will understand that the aforementioned electronic device can also be an independent physical server, or a virtualization platform or cloud platform composed of multiple servers. Essentially, it can be any device with computing and network communication capabilities, and this application does not limit it to any particular type. Furthermore, the banking document library can also be stored on a local server or in a distributed storage system, and this embodiment does not limit it to any particular type either.

[0041] It should be understood that, in addition to the bank branch service scenario mentioned above, the method of this application can also be used in any other scenario involving intelligent question answering based on a domain document library, and this application does not limit this to such scenarios.

[0042] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings, taking any electronic device as the execution subject as an example.

[0043] This application provides an intelligent question-answering method. Figure 2 A flowchart illustrating an intelligent question-answering method provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method in this embodiment includes:

[0044] S201, Search the preset banking domain document library to obtain a list of search results corresponding to each query in the question set generated based on user input.

[0045] The search results list includes multiple documents sorted by similarity.

[0046] Specifically, in this embodiment, the electronic device converts the original documents in the banking domain document library into document feature vectors through an embedding model to construct a vector database; and converts each query in the question set into a query feature vector through an embedding model.

[0047] Based on this, for each query feature vector, the vector similarity between the query feature vector and the feature vectors of each document is calculated in parallel in the vector database; the calculated vector similarity is then used to sort the results and generate a corresponding list of search results for each query.

[0048] More specifically, in this embodiment, the embedding model is a specialized semantic encoding model adapted for the banking domain. It is trained on corpora from areas such as banking policy documents and business terminology databases to map text sequences to a fixed-dimensional semantic vector space, achieving accurate representation of banking business rules, professional terminology, and contextual semantics. Vector similarity is calculated using the cosine similarity algorithm, which quantifies the degree of semantic association by the ratio of the inner product of two vectors to the product of their magnitudes. The ranking process is based on descending order of similarity values, forming a list of search results sorted from high to low relevance.

[0049] It should be understood that in practical applications, the method of obtaining feature vectors can also be combined with word embedding models optimized for the banking domain (such as Word2Vec) and text aggregation strategies, or with knowledge-enhanced embedding models incorporating banking entity knowledge. When performing vector-based retrieval, efficient indexing mechanisms of vector databases (such as inverted file indexes, hierarchical navigation small-world graph indexes, etc.) can be used to improve retrieval efficiency. The sorting process can also be an ascending order based on vector similarity values ​​(selecting documents with lower relevance only as references in specific scenarios), or a descending order after filtering documents with similarity values ​​higher than a preset threshold. In this embodiment, no restrictions are placed on the specific model type for obtaining feature vectors, the indexing mechanism of the vector database, or the sorting method.

[0050] In addition, in practical applications, the document library in the banking field can also be retrieved by keyword matching, or by statistical language model retrieval (such as the BM25 algorithm), or by entity linking and path reasoning retrieval that integrates the banking knowledge graph. This embodiment does not limit this method.

[0051] In this embodiment, the domain knowledge base retrieval method based on vector similarity, by converting text into semantic vectors, breaks through the limitations of traditional retrieval based on literal matching. It can effectively capture deep semantic relationships in the banking field, such as synonyms (e.g., "dead term" and "fixed term", "check bank statements" and "print transaction details") and business process associations (e.g., "report lost card" and "replace card"), reducing retrieval errors caused by differences in expression. The vectors generated by the unified embedding model ensure the comparability of queries and documents in the same semantic space, improving the retrieval accuracy of banking professional content. The parallel similarity calculation and ranking mechanism ensures retrieval efficiency in multi-query scenarios and provides high-quality basic data support for subsequent fusion ranking.

[0052] As a preferred example, before vector transformation, the original documents in the banking domain document library and the user input are preprocessed. The user input is then rewritten and expanded to improve the matching accuracy between the query and the knowledge base, cover the user's potential intent, and provide a more accurate foundation for subsequent retrieval. The specific process is as follows:

[0053] When an electronic device receives user input, it first preprocesses the user input to obtain a preprocessed user input query. Then, it uses a pre-trained large model to rewrite the user input query, generating at least one rewritten query, and constructs a question set based on the user input query and the rewritten query.

[0054] In this embodiment, user input includes at least one of voice, text, and image. Based on this, preprocessing includes: performing error correction processing on the user input based on at least one of general error correction knowledge, a banking-specific error correction dictionary, constraints embedded in banking business rules, user profiles, and dialogue context; and / or performing multimodal fusion encoding on the text, voice, and image in the user input, and performing semantic alignment and intent clarification on the user input based on the encoded cross-modal information.

[0055] Specifically, the bank-specific error correction dictionary is built based on real conversation data from bank branches, including the mapping relationship between high-frequency customer expressions and standard banking terminology; the constraints embedded in banking business rules include at least one of account number format verification, business operation compliance verification, and context inference rules.

[0056] More specifically, in this embodiment, the bank-specific error correction dictionary is built based on a massive amount of real conversation data from bank branches (such as customer recordings, teller records, and historical work orders). It includes standard terms (such as "report lost", "fixed-term to current account", "off-counter", and "online banking"), and also focuses on including high-frequency customer expression mappings (such as "fixed term" -> "fixed term", "current account" -> "current account deposit", "card swallowed" -> "ATM swallowed card", "check transaction history" -> "transaction details print", "change password" -> "reset password".

[0057] In this embodiment, the constraints of banking business rules are embedded in error correction logic to ensure that the semantics conform to the business logic. The constraints of banking business rules include: ① Account number format verification, used to verify the validity of data format, range, or logic, such as a fixed length of 19 digits for account numbers; ② Context inference rules, used to limit error correction options based on context and infer implicit constraints based on known information, such as the "report lost" operation only applying to bank cards in a "normal" status, or the automatic completion of the "purpose of foreign exchange purchase" field required by foreign exchange controls if a user mentions "cross-border transfer"; ③ Business operation compliance verification, used to enforce compliance with regulatory requirements, such as anti-money laundering requirements, where a single transfer exceeding 50,000 yuan requires manual review.

[0058] In this embodiment, when the electronic device uses the dialogue context for error correction, it includes the following steps: 1) Constructing a user profile tag system, which includes key features that affect language expression, and updating them through interactive implementation, such as age stratification, historical data, business type, etc.; 2) Prioritizing dictionary scheduling based on user profiles, for example, enabling dialect correction dictionaries for elderly users and professional terminology enhancement for high-net-worth users ("fixed income" -> "fixed income products"); 3) Refining and completing questions based on the context of multiple rounds of dialogue, for example, if a user asks "Can I cancel it?" -> combined with the previous "I just bought insurance" -> completes it as "Cancel insurance".

[0059] The preprocessing described above significantly improves the recognition rate of colloquial and incorrect expressions, making preprocessed user input queries closer to the standard expressions in the business knowledge base. This reduces irrelevant knowledge retrieval due to expression errors and improves the overall efficiency of intelligent question answering. Simultaneously, it avoids providing misleading answers due to misunderstandings (such as confusing reporting a lost item with account cancellation), reducing compliance risks and thus enhancing the user experience.

[0060] Furthermore, in this embodiment, the electronic device also employs a multimodal neural network structure, capable of simultaneously processing text (including user speech-to-text), speech (original audio prosody and emotion), and images (certificate photos taken by the user at the branch, screenshots, gestures, etc.), performing multimodal fusion encoding. Then, based on the encoded cross-modal information, semantic alignment and intent clarification are performed on the user input.

[0061] In this embodiment, the multimodal neural network structure includes a text encoder, an audio encoder, an image encoder, and a cross-modal fusion layer. Based on this, when the electronic device performs semantic alignment and intent clarification on user input based on cross-modal information, the specific process includes: the text encoder extracts keywords and contextual semantics; the audio encoder extracts speech content and emotional signals; the image encoder recognizes credential information or gestures; the cross-modal fusion layer uses a dynamic gating fusion unit to weight and fuse features from each modality, utilizes the multimodal Transformer core for cross-modal interactive reasoning, achieves semantic alignment, and outputs a clear intent based on business rules.

[0062] More specifically, in this embodiment, the text encoder adopts a fine-tuned financial BERT, and the audio encoder adopts a Wav2Vec2.0 model fine-tuned based on bank customer service voice corpus. By extracting speech features (such as phonemes, prosody, and speech rate) and emotional features (such as pitch and pause rhythm) in the audio, it can recognize the emotional tendencies (such as questions and anxiety) and semantic stress in the user's voice. The image encoder is based on the ResNet-50 visual Transformer (ViT) and is specifically trained for image types such as vouchers (such as bank cards, ID cards, and business receipts), interface screenshots (such as APP operation interfaces), and gestures (such as pointing and swiping). It can accurately identify key information (such as card number, button position, and business identifier) ​​in the image. The cross-modal fusion layer uses an architecture that combines a dynamic gating fusion unit with a multimodal Transformer. The dynamic gating unit learns a weight matrix to adaptively weight text feature vectors, audio feature vectors, and image feature vectors (e.g., increasing the weight of audio features when the voice is emotionally intense, and increasing the weight of image features when the image contains clear business identifiers). The multimodal Transformer then uses a cross-attention mechanism to achieve semantic association reasoning between modalities (e.g., associating and aligning the voice "transfer limit" with the text "daily limit of 50,000 yuan" in the screenshot), and finally outputs a unified semantic representation for clear intent.

[0063] For example, taking a user's voice saying "I can't find my passbook!" (an anxious tone, fast speech) and a photo of the passbook cover as an example, the audio encoder (a Wav2Vec2.0 model fine-tuned from bank customer service voice data) can extract anxiety features from the voice, obtaining an emotional urgency score of 0.9, and judging it as an emergency event; the text encoder (a fine-tuned financial BERT) recognizes the speech-to-text as "passbook lost"; the image encoder (a visual Transformer based on ResNet-50) extracts key information such as the passbook account number "xxxxxxxx" through OCR. The cross-modal fusion layer (a combination of dynamic gating fusion units and multimodal Transformers) enhances the weight of audio features based on the urgency of the speech. The multimodal Transformer associates text with the account, infers the need for reporting a lost account based on business rules, and outputs the result "The customer needs to urgently report the loss of their fixed-term deposit passbook, account number 'xxxxxxxx', current status is normal. Please prioritize reporting the loss via voice: 1. Verify identity information; 2. Confirm the type of loss report; 3. Set a temporary password..." The entire process achieves multimodal fusion processing, accurately and efficiently handling user needs.

[0064] In this embodiment, upon receiving user input, the electronic device simultaneously performs two preprocessing steps: error correction and multimodal fusion coding. A two-layer error correction engine corrects representational errors, while multimodal joint understanding fills in information gaps. In practical applications, the electronic device may perform only one of these preprocessing methods; this embodiment does not limit this approach.

[0065] As described above, error correction relies on banking-specific resources and business rules to accurately correct users' colloquial and erroneous expressions. It maps non-standard expressions like "dead-end" and "check bank statements" to business terms like "fixed-term" and "transaction details printout," ensuring semantic consistency between the query and the knowledge base. This reduces irrelevant knowledge retrieval due to expression discrepancies and avoids confusion between "reporting lost" and "closing an account" through compliance checks, guaranteeing the accuracy and compliance of the answer. Multimodal fusion coding breaks through the limitations of pure text interaction by combining cross-modal information such as voice emotion and image credentials. For example, it combines a vague user question like "How do I do this?" with a screenshot from mobile banking to refine it into a specific query for a particular business scenario, filling information gaps to capture the complete intent. Both methods, used in tandem or individually, improve the accuracy of user input parsing, providing a high-quality foundation for subsequent retrieval and answer generation. This effectively solves problems such as non-standard user expressions and incomplete descriptions of complex business processes in bank branch scenarios, ultimately improving the effectiveness and user experience of intelligent question answering.

[0066] In this embodiment, the pre-trained large model used by the electronic device to rewrite the user's input query is specifically a BART-large seq2seq model. Each user input query is processed by the large model to generate three more clearly and rigorously worded rewritten queries. For example, if a customer asks, "What should I do if I forget my card password?", the large model generates similar rewritten queries: 1. "How do I reset my bank card password if I forget it?"; 2. "Bank password reset service procedures"; 3. "What procedures should I follow to securely retrieve or reset my bank card password if I forget it?".

[0067] In practical applications, a query rewriting strategy based on retrieval feedback can also be adopted. This involves first obtaining relevant document fragments through preliminary retrieval, and then using the standard expressions in the documents to reverse-optimize and generate rewritten queries. Alternatively, structured rewriting can be performed using banking domain rule templates. For example, a standardized query containing elements such as "account type" and "operation type" can be automatically generated for "account-related issues." This embodiment does not impose any limitations on this approach. Furthermore, in practical applications, the number of rewritten queries can be dynamically adjusted based on the complexity of the user input. For example, two rewritten queries can be generated for simple questions, while four to six rewritten queries can be generated for complex, multi-intent questions. This embodiment also does not impose any limitations on this approach.

[0068] In the above process, preprocessing addresses the issues of expression bias and incomplete information in user input through error correction and multimodal understanding, transforming non-standard expressions into standardized queries that align with the business knowledge base. Rewriting and expansion, on the other hand, generates multiple semantically similar queries to cover potential user intent and different expression perspectives. The combination of these two approaches significantly improves the breadth and accuracy of matching queries with the banking domain document library, providing a more comprehensive retrieval basis for subsequent parallel searches, reducing the omission of relevant knowledge due to the limitations of a single query, and thus laying a solid foundation for generating accurate and comprehensive answers.

[0069] S202, dynamically optimize the fusion parameters of the fusion ranking algorithm based on the document sorting and similarity information of each search result list.

[0070] The document ranking in the search results list refers to the order in which documents in the search results for each query are arranged from highest to lowest vector similarity. For example, a document might be ranked 3rd in the search results for query A and 5th in the search results for query B. Similarity information refers to the quantified value of the vector similarity between each document and its corresponding query, such as a cosine similarity value of 0.85 or 0.72.

[0071] In this embodiment, the fusion ranking algorithm adopts the Reciprocal Rank Fusion (RRF) algorithm, which assigns a reciprocal rank score to each document (calculated using the formula: ...). In the RRF formula, set D represents a given set of documents to be sorted, R represents a set of rankings corresponding to the document set, r(d) represents the ranking of document d in set D, and k is a constant. The scores of the same document in multiple sets of search results are added together to obtain a combined score, which is used as the final ranking criterion. The fusion parameters include the constant k in the RRF formula, as well as dynamically adjustable path weights (set based on the overall relevance performance of each search result list).

[0072] In practical applications, the fusion ranking algorithm can also adopt other multi-source result fusion algorithms (such as the LambdaRank ranking model based on machine learning). Accordingly, the fusion parameters can include the weight coefficients of each retrieval path feature in the model, the influence factor of similarity features, the adjustment parameters of the ranking loss function, etc., and these parameters can still be dynamically optimized based on the document ranking of each retrieval result list (such as the difference in the ranking distribution of documents under different paths) and similarity information (such as average similarity and similarity fluctuation value). For example, higher weight can be given to the retrieval path corresponding to documents with stable ranking and high similarity, or the contribution of features in the model can be adjusted according to the similarity distribution to improve the accuracy of fusion ranking.

[0073] S203 uses dynamically optimized fusion parameters and a fusion sorting algorithm to merge and reorder documents in each search result list to obtain a reordered target search result list.

[0074] S204, Generate the target answer corresponding to the user input based on the target search result list.

[0075] In this embodiment, the electronic device uses dynamically optimized fusion parameters to perform a reciprocal ranking fusion calculation on the documents in each search result list. The reciprocal ranking scores of the same document in different lists are summed to obtain a combined score. Then, the documents are reordered from highest to lowest based on their combined scores to form the target search result list. For example, if a document ranks 1st, 2nd, 3rd, and 4th in the four search result lists, and k=60, its combined score is 1 / (60+1)+1 / (60+2)+1 / (60+3)+1 / (60+4), and its final position in the target list is determined accordingly.

[0076] Furthermore, the electronic device selects the top-ranked (e.g., Top 5) documents from the list as the knowledge base, and inputs the content of these documents along with all queries in the question set into the BART-large seq2seq model. The large model then combines the retrieved business knowledge with the user's query intent to generate a natural language answer, which is then fed back to the user.

[0077] It should be understood that in practical applications, the number of documents selected when generating the target answer can be adjusted according to the needs of the business scenario (e.g., selecting the Top 3 for simple questions and the Top 10 for complex questions). Large models can also use other pre-trained models suitable for text generation. The fusion ranking algorithm can also combine factors such as document timeliness and source authority to optimize the ranking results. This embodiment does not limit the specific implementation details.

[0078] The intelligent question-answering method provided in this embodiment achieves accurate retrieval and efficient utilization of banking domain knowledge through four core steps: retrieving a list of search results, dynamically optimizing fusion parameters, merging and sorting the search results to generate a target search result list, and finally generating an answer. The dynamically optimized fusion parameters ensure a reasonable weight allocation for each search result list, the fusion sorting improves the overall relevance of the results, and the answer generated based on the target search result list effectively combines business knowledge with user needs, thus improving the accuracy, comprehensiveness, and response efficiency of intelligent question answering in bank branches.

[0079] The method described in this embodiment solves the problems in known technologies, such as inaccurate search results due to unclear, incomplete, or colloquial user question expressions, and rigid weighting when merging multiple search result lists, making it difficult to reflect the true relevance of documents. In particular, it significantly improves the practicality and compliance of intelligent question answering, especially considering the characteristics of bank branch scenarios with many business terms and strict rules.

[0080] This application also provides an embodiment of an intelligent question-answering method, which provides a detailed description of the optimization of fusion parameters and the method for generating a list of target search results. Figure 3 A flowchart illustrating an intelligent question-answering method provided in this application embodiment. Figure 2 The following is combined with Figure 3 The implementation process and principles of this method are explained in detail. For example... Figure 3 As shown, the method in this embodiment further includes:

[0081] S301, dynamically optimize the fusion parameters of the fusion ranking algorithm based on the document sorting and similarity information of each search result list.

[0082] Specifically, in this embodiment, the fusion sorting algorithm is the RRF algorithm, and the fusion parameters include the constant k value and path weight of the fusion sorting algorithm.

[0083] For the constant k value, the electronic device optimizes it in the following way: if the similarity between the first-ranked document in the first search result list and the corresponding query is higher than the similarity between the first-ranked document in the second search result list and the corresponding query, reaching a first preset threshold, then the first search result list is assigned a lower constant k value than the second search result list, so that the first search result list obtains a higher weight in the fusion ranking; and / or, if the minimum similarity of the top N documents in the third search result list is lower than a second preset threshold, then the constant k value of the third search result list is increased, so as to reduce the weight of the third search result list in the fusion ranking.

[0084] More specifically, when optimizing the constant k value, the electronic device first obtains the similarity between the top-ranked document in each search result list and the corresponding query, as well as the minimum similarity between the top N (e.g., N=5) documents. It should be noted that the first preset threshold is the trigger threshold for the difference in similarity between the first and second documents in the two lists. That is, when the difference in similarity between the first and second documents in the first list reaches or exceeds this threshold, it indicates that the quality of the first document in the first list is significantly better. At this time, by assigning it a lower k value, the calculation logic of "1 / (k+r(d))" in the RRF algorithm (the smaller the k value, the larger the score of the last-ranked document under the same ranking) is used to reasonably increase the weight of the first list. The first preset threshold is set to 15%, the second preset threshold is 0.4, and the initial constant k defaults to 60. For example, if the document ranked first in the first search results list has a similarity of 0.9 and the document ranked first in the second search results list has a similarity of 0.7, the difference between the two is 20% (reaching and exceeding 15%), then the k value of the first search results list is reduced from 60 to 40, while the k value of the second search results list remains at 60; if the minimum similarity of the top 5 documents in the third search results list is 0.35 (below 0.4), then its k value is increased from 60 to 80 to weaken the influence of low-quality documents in that list.

[0085] It should be understood that in practical applications, the first preset threshold can be flexibly set according to the accuracy requirements of the business scenario (e.g., 20% for high-risk business scenarios and 10% for general consultations), and the second preset threshold can also be dynamically adjusted according to the overall quality of the search results (e.g., for business areas with frequent document updates, the threshold can be increased to 0.5 to strictly screen high-quality results; for niche businesses with scarce documents, it can be reduced to 0.3 to expand the scope of results coverage). This embodiment does not limit this.

[0086] For path weights, the electronic device optimizes them in the following way: for each list of search results, the semantic complexity of the user input query and the average semantic similarity between each rewritten query and the user input query are determined; based on the semantic complexity and the average semantic similarity, the path weights of the search results list are determined.

[0087] More specifically, when optimizing path weights, electronic devices first calculate the semantic complexity of the user input query based on the number of intents and entity density. The number of intents needs to be normalized: the number of independent business intents included in the query (e.g., "How to calculate and deduct personal income tax for early loan repayment penalties" contains 2 intents) is counted and mapped to a normalized interval of [1.0, 2.0] (1 intent corresponds to 1.0, 2 intents to 1.3, 3 or more intents to 1.5, with a maximum of 2.0), avoiding excessive weights due to directly rounding the number of intents to integers. Simultaneously, the mean cosine similarity between each rewritten query and the user input query is calculated (e.g., the similarities of 3 rewritten queries are 0.8, 0.7, and 0.9, with a mean of 0.8). The weights are then determined based on the formula "Path Weight = Base Weight (default 1.0) × Semantic Complexity × Mean Semantic Similarity". For example, if a user enters the query "How to calculate and deduct the penalty for early loan repayment from personal income tax" (containing 2 intents, with a normalized semantic complexity of 1.3), and the mean semantic similarity between a rewritten query and the original query is 0.85, then the path weight of the search result list corresponding to the rewritten query is 1.0 × 1.3 × 0.85 = 1.105. Compared to the search result list with a semantic complexity of 1.0 and a mean similarity of 0.6 (path weight = 0.6), the rewritten query has a higher weight in the fusion ranking.

[0088] In practical applications, the optimization of the constant k can also be based on the overall similarity variance of the search result list (the smaller the variance, the more stable the result, and the value of k can be reduced); the optimization of path weight can be combined with user profile (such as appropriately increasing the path weight corresponding to the rewritten query of elderly users). This embodiment does not limit the specific optimization dimensions.

[0089] It should be understood that the similarity mentioned above refers to semantic similarity. This means that after converting the query and document into vectors using a pre-trained semantic embedding model (such as a fine-tuned BERT model for the financial domain), the cosine similarity of the vectors is calculated. This measure measures the degree of semantic association between the two (such as the matching degree between query intent and document content), rather than a simple literal matching degree. For example, although "reporting a lost savings card" and "how to report a lost debit card" have different literal expressions, their semantic similarity will approach 1.0. This is consistent with the semantic understanding-based retrieval logic in the technical solution, ensuring that the similarity metric used when dynamically optimizing the fusion parameters truly reflects the quality of the semantic association between the query and the document.

[0090] In the above process, by dynamically adjusting the constant k through similarity differences, the weight of high-quality search result lists can be accurately strengthened and the influence of low-quality lists can be weakened, avoiding the rigidity of weight caused by a single k value. The path weight is optimized based on semantic complexity and mean similarity, so that the search results corresponding to the rewritten query that are closer to the original query intent can obtain higher weights, thereby improving the targeting and accuracy of the fusion ranking. The combination of the two achieves refined dynamic optimization of the fusion parameters.

[0091] S302 uses dynamically optimized fusion parameters to calculate the reciprocal ranking score of each document in multiple search result lists and sums them to obtain the combined score.

[0092] Specifically, for each document, in each search result list in which it appears, the individual score is calculated according to the document's rank r(d) in the list and the corresponding list optimization constant k, using the formula "reciprocal rank score = 1 / (k + r(d))". All individual scores are then summed to obtain the combined score. For example, if a document ranks first in the list with k=40 (score = 1 / (40+1) = 1 / 41) and second in the list with k=60 (score = 1 / (60+2) = 1 / 62), then the combined score = 1 / 41 + 1 / 62 ≈ 0.024 + 0.016 = 0.04.

[0093] S303, multiply the combined score by the timeliness factor and / or the document source authority coefficient to obtain the final score.

[0094] In this embodiment, due to the time-sensitive nature of banking policies, interest rates, and other information, and the varying credibility of documents from different sources, the final score in this embodiment is calculated as: combined score × timeliness factor × document source authority coefficient. The timeliness factor is calculated based on the number of days between the document's publication date and the current time (e.g., factor within 30 days = 1.0, factor between 30 and 90 days = 0.8). The document source authority coefficient is determined using a preset matrix (e.g., central bank document = 1.5, branch manual = 1.2).

[0095] In practical applications, only the timeliness factor (such as when querying "latest deposit interest rate") or only the authority coefficient of the document source (such as when querying "core business system") may be introduced. This embodiment does not limit this.

[0096] S304. Sort the documents based on the final scores to obtain a list of target search results.

[0097] Specifically, the documents are sorted in descending order of their final scores. The resulting list of documents, ranked by relevance, timeliness, and authority, is the target search results list. For example, if document A has a final score of 0.04 × 1.0 × 1.5 = 0.06 and document B has a final score of 0.03 × 0.8 × 1.2 = 0.028, then document A ranks higher than document B in the target search results list.

[0098] In this embodiment, the method ensures a reasonable weight allocation of multi-source retrieval results by calculating the combined score through dynamically optimized fusion parameters. By superimposing timeliness factors and document source authority coefficients, the final ranking takes into account semantic relevance, information timeliness, and source credibility. This solves the problem of traditional RRF relying solely on ranking and ignoring the document's own attributes. The generated target retrieval result list is more in line with the business needs of the banking sector, providing higher-quality knowledge basis for generating accurate answers in the future.

[0099] As described above, this application preprocesses user input in multiple dimensions (combining a banking-specific error correction dictionary, business rule constraints, user profiles, and dialogue context to achieve accurate error correction and semantic standardization, while utilizing multimodal neural networks to process text, speech, and image information and complete cross-modal semantic alignment) to generate a question set that conforms to business standards, including user input queries and rewritten queries. Based on the question set, a retrieval is performed to obtain multiple lists of search results. Then, the fusion parameters of the fusion ranking algorithm (such as the constant k value and path weight of the RRF algorithm) are dynamically optimized based on the document ranking and similarity information of each list. This dynamic adjustment achieves a reasonable weight allocation for multi-source results. The optimized parameters are then used to fuse and rank the search result lists, and a target search result list is generated by combining timeliness factors and document source authority coefficients. Finally, an accurate answer is generated based on this list.

[0100] The overall solution improves the standardization and multimodal understanding of user input through preprocessing, enhances the accuracy of multi-source retrieval results through dynamic optimization of fusion parameters, and balances business timeliness and authority through multi-factor ranking. It effectively solves the problems of inaccurate retrieval and biased answers caused by non-standard user expression, rigid fusion of multi-source results, and neglect of business scenario characteristics in traditional intelligent question answering. It significantly improves the accuracy, compliance and user experience of intelligent question answering, and is especially suitable for fields such as banks with high requirements for business accuracy and compliance.

[0101] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0102] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0103] The above embodiments introduce an intelligent question-answering method from the perspective of process flow. The following embodiments introduce an intelligent question-answering device from the perspective of virtual module or virtual unit. For details, please refer to the following embodiments.

[0104] This application provides an intelligent question-answering device. Figure 4 This is a schematic diagram of the structure of an intelligent question-answering device provided in an embodiment of this application, as shown below. Figure 4 As shown, the device includes:

[0105] The retrieval module 41 is used to perform a retrieval in a preset banking domain document library and obtain a list of retrieval results corresponding to each query in the question set generated based on user input; wherein, the list of retrieval results includes multiple documents sorted by similarity;

[0106] The optimization module 42 is used to dynamically optimize the fusion parameters of the fusion ranking algorithm based on the document sorting and similarity information of each search result list;

[0107] The fusion module 43 is used to merge and re-sort the documents in each search result list using dynamically optimized fusion parameters and based on the fusion sorting algorithm to obtain the re-sorted target search result list.

[0108] The generation module 44 is used to generate the target answer corresponding to the user input based on the target retrieval result list.

[0109] In another possible implementation of this application embodiment, the fusion parameters include the constant k value of the fusion sorting algorithm; the optimization module 42 is specifically used for:

[0110] If the similarity between the first-ranked document in the first search result list and the corresponding query is higher than the similarity between the first-ranked document in the second search result list and the corresponding query, reaching a first preset threshold, then a lower constant k value is assigned to the first search result list compared to the second search result list, so that the first search result list can obtain a higher weight in the fusion ranking.

[0111] And / or, if the minimum similarity of the top N documents in the third search results list is lower than a second preset threshold, then the constant k value of the third search results list is increased to reduce the weight of the third search results list in the fusion ranking.

[0112] Another possible implementation of this application embodiment includes a problem set comprising user-input queries and rewritten queries, and fusion parameters including path weights; the optimization module 42 is specifically used for:

[0113] For each list of search results, determine the semantic complexity of the user-input query and the average semantic similarity between each rewritten query and the user-input query;

[0114] The path weights of the search results list are determined based on semantic complexity and the mean of semantic similarity.

[0115] In another possible implementation of this application embodiment, the fusion module 43 is specifically used for:

[0116] Using dynamically optimized fusion parameters, the reciprocal ranking score of each document in multiple search result lists is calculated and summed to obtain the combined score;

[0117] The final score is obtained by multiplying the combined score by the timeliness factor and / or the document source authority coefficient.

[0118] The documents are sorted based on their final scores to obtain a list of the target search results.

[0119] In another possible implementation of this application embodiment, the retrieval module 41 is further configured to:

[0120] Upon receiving user input, the user input is preprocessed to obtain the preprocessed user input query.

[0121] A pre-trained large model is used to rewrite the user input query, generating at least one rewritten query, and a question set is constructed based on the user input query and the rewritten query.

[0122] Another possible implementation of this application embodiment includes preprocessing as follows:

[0123] Based on at least one of the following: general error correction knowledge, banking-specific error correction dictionary, constraints embedded in banking business rules, user profiles, and dialogue context, error correction processing is performed on user input;

[0124] And / or, perform multimodal fusion encoding on the text, speech, and images in the user input, and perform semantic alignment and intent clarification on the user input based on the encoded cross-modal information.

[0125] Another possible implementation of this application embodiment is that the bank-specific error correction dictionary is constructed based on real conversation data from bank branches, including the mapping relationship between high-frequency customer expressions and standard banking terminology; the constraints embedded in the banking business rules include at least one of account number format verification, business operation compliance verification, and context inference rules.

[0126] In another possible implementation of this application embodiment, the retrieval module 41 is specifically used for:

[0127] The original documents in the banking document library are converted into document feature vectors through an embedding model to build a vector database;

[0128] Each query in the question set is transformed into a query feature vector through an embedding model.

[0129] In another possible implementation of this application embodiment, the retrieval module 41 is specifically used for:

[0130] For each query feature vector, the vector similarity between the query feature vector and the feature vectors of each document is calculated in parallel in the vector database;

[0131] The results are sorted based on the calculated vector similarity, and a corresponding list of search results is generated for each query.

[0132] The intelligent question-answering device provided in this application embodiment is applicable to the embodiments of the above-described intelligent question-answering method, and will not be described again here.

[0133] This application provides an electronic device. Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 5 As shown, Figure 5 The illustrated electronic device includes at least one processor 51 and a memory 52. ​​The processor 51 and the memory 52 are connected, for example, via a bus 53. Optionally, the electronic device may also include a transceiver 54. It should be noted that in practical applications, the transceiver 54 is not limited to one, and the structure of this electronic device does not constitute a limitation on the embodiments of this application.

[0134] Processor 51 may be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 51 may also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0135] Bus 53 may include a pathway for transmitting information between the aforementioned components. Bus 53 may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Bus 53 may be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus 53 is represented by a single thick line, but this does not mean that there is only one bus 53 or only one type of bus 53.

[0136] The memory 52 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0137] The memory 52 is used to store computer execution instructions for implementing the scheme of this application, and the execution is controlled by the processor 51. The processor 51 is used to execute the computer execution instructions stored in the memory 52 to implement the content shown in the foregoing method embodiments.

[0138] This application also provides a computer-readable storage medium, which may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk. Specifically, the computer-readable storage medium stores computer-executable instructions, which are used to implement the methods in the above embodiments.

[0139] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the technical solution of the above method embodiments. Its implementation principle and technical effects are similar, and will not be repeated here.

[0140] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0141] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An intelligent question-answering method, characterized in that, The method includes: A search is performed in a pre-defined banking document library to obtain a list of search results corresponding to each query in a question set generated based on user input; wherein, the list of search results includes multiple documents sorted by similarity. Based on the document sorting and similarity information of each of the search result lists, the fusion parameters of the fusion sorting algorithm are dynamically optimized. Using dynamically optimized fusion parameters, the documents in each of the search result lists are merged and reordered based on the fusion sorting algorithm to obtain a reordered target search result list. The target answer corresponding to the user input is generated based on the target search result list.

2. The method according to claim 1, characterized in that, The fusion parameters include the constant k value of the fusion ranking algorithm; the step of dynamically optimizing the fusion parameters of the fusion ranking algorithm based on the document ranking and similarity information of each of the search result lists includes: If the similarity between the first-ranked document in the first search result list and the corresponding query is higher than the similarity between the first-ranked document in the second search result list and the corresponding query, reaching a first preset threshold, then a lower constant k value is assigned to the first search result list compared to the second search result list, so that the first search result list obtains a higher weight in the fusion ranking. And / or, if the minimum similarity of the top N documents in the third search result list is lower than a second preset threshold, then the constant k value of the third search result list is increased to reduce the weight of the third search result list in the fusion ranking.

3. The method according to claim 1 or 2, characterized in that, The question set includes user-input queries and rewritten queries, and the fusion parameters include path weights; the dynamic optimization of the fusion ranking algorithm parameters based on the document ranking and similarity information of each of the search result lists includes: For each list of search results, determine the semantic complexity of the user input query and the average semantic similarity between each rewritten query and the user input query; The path weights of the retrieval result list are determined based on the semantic complexity and the mean of each semantic similarity.

4. The method according to claim 1 or 2, characterized in that, The process involves using dynamically optimized fusion parameters and, based on the fusion ranking algorithm, fusing and re-ranking the documents in each of the search result lists to obtain a re-ranked target search result list, including: Using dynamically optimized fusion parameters, the reciprocal ranking score of each document in multiple search result lists is calculated and summed to obtain the combined score; The combined score is multiplied by the timeliness factor and / or the document source authority coefficient to obtain the final score; The documents are sorted based on the final scores to obtain the target retrieval result list.

5. The method according to claim 1 or 2, characterized in that, The method further includes: Upon receiving the user input, the user input is preprocessed to obtain a preprocessed user input query. The user input query is rewritten using a pre-trained large model to generate at least one rewritten query, and the question set is constructed based on the user input query and the rewritten query.

6. The method according to claim 5, characterized in that, The preprocessing includes: The user input is corrected based on at least one of the following: general error correction knowledge, a banking-specific error correction dictionary, constraints embedded in banking business rules, user profiles, and dialogue context. And / or, perform multimodal fusion encoding on the text, speech, and images in the user input, and perform semantic alignment and intent clarification on the user input based on the encoded cross-modal information.

7. The method according to claim 6, characterized in that, The bank-specific error correction dictionary is constructed based on real conversation data from bank branches, including the mapping relationship between high-frequency customer expressions and standard banking terminology; the constraints embedded in the banking business rules include at least one of account number format verification, business operation compliance verification, and context inference rules.

8. The method according to claim 1 or 2, characterized in that, The process involves searching a pre-defined banking document library to obtain a list of search results corresponding to each query in a question set generated based on user input, including: The original documents in the banking domain document library are converted into document feature vectors through an embedding model to construct a vector database; Each query in the question set is converted into a query feature vector using the embedding model.

9. The method according to claim 8, characterized in that, The process involves searching a pre-defined banking document library to obtain a list of search results corresponding to each query in a question set generated based on user input, including: For each query feature vector, the vector similarity between the query feature vector and each document feature vector is calculated in parallel in the vector database; The results are sorted based on the calculated vector similarity, and a corresponding list of search results is generated for each query.

10. An intelligent question-and-answer device, characterized in that, The device includes: The retrieval module is used to perform searches in a pre-defined banking-related document database based on vector similarity, and to obtain a list of retrieval results corresponding to each query in a question set generated based on user input; wherein, the list of retrieval results includes multiple documents sorted by similarity. The optimization module is used to dynamically optimize the fusion parameters of the fusion ranking algorithm based on the document sorting and similarity information of each of the search result lists. The fusion module is used to merge and reorder the documents in each of the search result lists based on the fusion sorting algorithm using dynamically optimized fusion parameters, so as to obtain a reordered target search result list. The generation module is used to generate the target answer corresponding to the user input based on the target retrieval result list.

11. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The at least one processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-9.

13. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-9.

Images