Intelligent question answering method and device based on large model, storage medium and server

Abstract

The invention discloses an intelligent question answering method and device based on a large model, a storage medium and a server, and relates to the field of information retrieval. The method comprises the steps of receiving a to-be-answered question input by a user; according to formats indicated by a plurality of preset retrieval algorithms, converting the question to be answered into query statements respectively corresponding to the plurality of retrieval algorithms; according to the generated multiple query statements, retrieval is performed in a knowledge base, and candidate document sets corresponding to the multiple query statements are obtained; fusing the candidate document sets according to a preset fusion rule to obtain a fused document set; extracting a target answer from the fusion document set according to a pre-trained generation large model; according to the method and the device, the candidate document sets generated by fusing different algorithms are balanced between accuracy and diversity, so that the answer is enabled to be related to the question, and rich information is provided.

Description

Technical Field

[0001] The present application relates to the field of information retrieval, and in particular to a large-model-based intelligent question-answering method, device, storage medium, and server. Background Art

[0002] As businesses continue to expand and their operations become increasingly complex, knowledge management becomes increasingly crucial. As one of a company's core assets, effective management and utilization of knowledge plays a key role in enhancing competitiveness and promoting innovation. Against this backdrop, many companies are deploying intelligent knowledge assistants to meet employees' daily research needs.

[0003] As a convenient knowledge acquisition tool, the Intelligent Knowledge Assistant plays a vital role in employees' daily work. It helps employees quickly find work-related information, such as business processes, policies and regulations, and technical documentation. This improves their efficiency in acquiring knowledge and promotes the circulation and sharing of corporate knowledge.

[0004] However, the search methods currently used by intelligent knowledge assistants deployed by enterprises mainly rely on keyword matching. This traditional search method has exposed a series of serious problems in actual application, greatly affecting the efficiency and quality of employees' knowledge acquisition, as follows:

[0005] (1) The query process is cumbersome: In keyword-matching search, users often need to enter multiple keywords for combined queries to obtain accurate information. Furthermore, due to differences in how different users phrase their questions, accurate keyword selection is difficult. This results in users spending a significant amount of time filtering and locating information, repeatedly adjusting query keywords, and constantly trying different query combinations, increasing query complexity and time costs.

[0006] (2) Poor relevance of search results: Keyword matching searches are primarily based on superficial vocabulary matching, failing to deeply understand the semantics and intent of user queries. This results in search results containing a large amount of information that is irrelevant to the user's question, making it difficult for users to quickly filter out the information that best matches their query.

[0007] (3) Inability to provide accurate answers and low efficiency: Existing search methods cannot accurately understand the underlying meaning of user queries, resulting in an inability to provide users with accurate answers. This results in a large number of queries requiring manual referral and review, increasing the workload. After obtaining information, employees also need to spend time judging and filtering it, further reducing work efficiency. Moreover, due to inaccurate search results, employees may need to search multiple times to obtain the required information, which affects the user experience and reduces their trust in intelligent knowledge assistants.

[0008] To sum up, the keyword matching search method used by existing enterprise intelligent knowledge assistants can no longer meet the growing knowledge management needs of enterprises. A new technical solution is urgently needed to solve the above problems, improve the accuracy and efficiency of knowledge retrieval, and enhance employee work efficiency and user experience. Summary of the Invention

[0009] The present invention provides a large-scale model-based intelligent question-answering method, device, storage medium, and server, which can solve the problems of low knowledge retrieval accuracy and efficiency in existing question-answering systems. The technical solution is as follows:

[0010] In a first aspect, an embodiment of the present application provides an intelligent question-answering method based on a large model, the method comprising:

[0011] Receive questions to be answered from users;

[0012] Converting the question to be answered into query statements corresponding to the respective retrieval algorithms according to the formats indicated by the preset retrieval algorithms;

[0013] Searching the knowledge base according to the generated multiple query statements to obtain candidate document sets corresponding to the multiple query statements;

[0014] According to the preset fusion rules, each candidate document set is fused to obtain a fused document set;

[0015] Extracting target answers from the fused document collection based on a pre-trained generative model;

[0016] The target answer is pushed to the user.

[0017] In a second aspect, an embodiment of the present application provides an intelligent question-answering device based on a large model, the device comprising:

[0018] A receiving unit, configured to receive a question to be answered input by a user;

[0019] a conversion unit, configured to convert the question to be answered into query statements corresponding to the plurality of retrieval algorithms according to the formats indicated by the plurality of preset retrieval algorithms;

[0020] A retrieval unit, configured to search the knowledge base according to the generated multiple query statements, and obtain candidate document sets corresponding to the multiple query statements;

[0021] A fusion unit, configured to fuse the candidate document sets according to a preset fusion rule to obtain a fused document set;

[0022] An extraction unit, configured to extract a target answer from the fused document set based on a pre-trained generative model;

[0023] A feedback unit is used to push the target answer to the user.

[0024] In a third aspect, an embodiment of the present application provides a computer storage medium, wherein the computer storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the above-mentioned method steps.

[0025] In a fourth aspect, an embodiment of the present application provides a server, which may include: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the above-mentioned method steps.

[0026] The beneficial effects of the technical solutions provided by some embodiments of the present application include at least:

[0027] This application breaks through the limitations of a single search path, analyzes queries and searches from multiple angles, generates multiple sets of candidate document collections, greatly improves the recall rate, and ensures that more relevant information is retrieved. At the same time, multiple algorithms complement each other, can correct algorithm deviations, and enhance the robustness of the results. In addition, this application achieves a balance between accuracy and diversity by integrating candidate document collections generated by different algorithms, ensuring that the answers are relevant to the questions while providing rich information. Ultimately, it provides high-quality input for the downstream generation of large models, optimizes the accuracy and reliability of its extraction of target answers, and comprehensively improves the performance of downstream generation tasks of the intelligent question-answering system, bringing users a better question-answering experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 This is a schematic diagram of the network architecture provided by the embodiment of the present application;

[0030] Figure 2 This is a flow chart of the intelligent question-answering method based on a large model provided in an embodiment of the present application;

[0031] Figure 3 This is a flowchart of synonym replacement provided by an embodiment of the present application;

[0032] Figure 4 This is a schematic diagram of the structure of an intelligent question-answering device based on a large model provided by this application;

[0033] Figure 5 is a schematic diagram of a computer storage medium provided in an embodiment of the present application;

[0034] Figure 6 This is a structural diagram of a server provided by this application. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions and advantages of the present application clearer, the embodiments of the present application will be described in further detail below with reference to the accompanying drawings.

[0036] It should be noted that the intelligent question-answering method based on a large model provided in this application is generally executed by a server, and accordingly, the intelligent question-answering device based on a large model is generally set in the server.

[0037] Figure 1 An exemplary network architecture of the large-model-based intelligent question-answering method or large-model-based intelligent question-answering device that can be applied to the present application is shown.

[0038] like Figure 1 As shown, the network architecture may include: a terminal device 101 and a server 102. The terminal device 101 and the server 102 may communicate via a network, which is used as a medium for providing communication links between the above-mentioned units. The network may include various types of wired communication links or wireless communication links, for example: wired communication links include optical fibers, twisted pairs, or coaxial cables, and wireless communication links include Bluetooth communication links, wireless fidelity (Wi-Fi) communication links, or microwave communication links.

[0039] Among them, an intelligent question-answering system is deployed in the server 102, which receives questions to be answered input by users through the client or web page of the terminal device 101, generates query statements based on the questions to be answered and sends them to the server 102, and the server 102 pushes the corresponding target answers to the terminal device 101.

[0040] It should be noted that the terminal device 101 and the server 102 can be hardware or software. When the terminal device 101 and the server 102 are hardware, they can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the terminal device 101 and the server 102 are software, they can be implemented as multiple software or software modules (for example, to provide distributed services), or as a single software or software module, without specific limitation herein.

[0041] Various communication client applications can be installed on the terminal device of this application, such as: video recording applications, video playback applications, voice interaction applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0042] The terminal device can be either hardware or software. When the terminal device is hardware, it can be various terminal devices with a display screen, including but not limited to smartphones, tablet computers, laptop computers, and desktop computers. When the terminal device is software, it can be installed in the terminal devices listed above. It can be implemented as multiple software or software modules (for example, to provide distributed services), or it can be implemented as a single software or software module, which is not specifically limited here.

[0043] When the terminal device is hardware, it may also be equipped with a display device and a camera. The display device can be any device capable of displaying a video, and the camera is used to capture video streams. For example, the display device may be a cathode ray tube display (CR), a light-emitting diode display (LED), an electronic ink screen, a liquid crystal display (LCD), or a plasma display panel (PDP). Users can use the display device on the terminal device to view displayed text, images, videos, and other information.

[0044] It should be understood that Figure 1 The number of terminal devices, networks, and servers in the embodiment is only illustrative. Any number of terminal devices, networks, and servers may be used as needed.

[0045] The following will be combined with the Figure 2 , the intelligent question answering method based on a large model provided by the embodiment of the present application is described in detail. Among them, the intelligent question answering device based on a large model in the embodiment of the present application can be Figure 1 The terminal device shown.

[0046] See Figure 2 , provides a flow chart of an intelligent question-answering method based on a large model for the embodiment of this application. Figure 2 As shown, the method of the embodiment of the present application may include the following steps:

[0047] S201: Receive a question to be answered input by a user.

[0048] The server monitors user requests in real time via a network interface (e.g., HTTP). When a user enters a question on their device through the Q&A system's client interface and clicks "Submit," the request is transmitted to the server via the network in the form of a data packet. Upon receiving the data packet, the server parses it and extracts the text of the question.

[0049] For example, employee Xiao Li needs to understand how to quickly handle unusual packages during work. He enters the question "What is the process for handling unusual packages?" into the client interface of the company's internal Q&A system and clicks Submit. After receiving this request, the server interprets the question as "What is the process for handling unusual packages?"

[0050] S202: According to the statement formats required by the preset multiple search algorithms, convert the question to be answered into query statements corresponding to the multiple search algorithms respectively.

[0051] The server pre-stores a variety of search algorithms, each of which has specific requirements for the format of query statements. The server analyzes the semantics, structure, and keywords of the question to be answered, and then converts the original question according to the requirements of different search algorithms. For example, for a search algorithm based on keyword matching, the server extracts the keywords in the question and organizes them into query statements in the format required by the algorithm; for a search algorithm based on semantic understanding, the server may perform semantic analysis on the question and convert it into a query statement that can express semantic information. The server has powerful natural language processing capabilities to ensure that the converted query statement can accurately reflect the intention of the original question and meet the requirements of different search algorithms.

[0052] For example, when Xiao Li asks, "What is the process for handling abnormal packages?", the server converts the query using a pre-set search algorithm. For the keyword-matching BM25 algorithm, the server extracts the keywords "abnormal package" and "process" and formulates them into the query "abnormal package processing process." For algorithms based on semantic understanding, the server might convert the question into a semantically expressive query, such as "Get the steps for handling abnormal packages."

[0053] In some possible embodiments of the present application, multiple retrieval algorithms include: retrieving candidate documents in a knowledge base based on a keyword matching algorithm, or retrieving candidate documents in a knowledge base based on a vector similarity algorithm.

[0054] Among them, multiple search algorithms include those based on keyword matching (such as the classic sparse search algorithm BM25) and those based on vector similarity (which falls under the category of dense search). This multi-algorithm design that combines sparse and dense search can leverage the advantages of both, forming a powerful complementary effect.

[0055] Keyword matching algorithms in sparse search, such as BM25, can quickly locate documents containing specific keywords thanks to their precise capture and efficient matching capabilities. When dealing with straightforward questions with clear keywords, these algorithms can quickly filter out highly relevant documents, ensuring high recall and avoiding missing information directly relevant to the question.

[0056] The vector similarity algorithm used in dense search possesses powerful semantic generalization capabilities. By mapping text into a high-dimensional vector space, it can capture semantic similarities between texts, going beyond superficial keyword matching. Even if the question and document are expressed differently but share similar semantics, the vector similarity algorithm can accurately identify them, thereby uncovering potentially relevant documents and expanding the scope and depth of search.

[0057] This application organically combines sparse and dense search, leveraging both the efficiency and accuracy of sparse search in keyword matching and the flexibility and generalization of dense search in semantic understanding. In the multi-dimensional search process, the two complement each other, comprehensively analyzing queries from different perspectives to generate a richer and more comprehensive set of candidate documents, further optimizing the quality and diversity of search results and laying the foundation for the subsequent generation of more accurate and higher-quality target answers.

[0058] For example, candidate documents are retrieved from the knowledge base based on a keyword matching algorithm. The specific process may include:

[0059] Word segmentation of questions to be answered and generation of query statements: After receiving the questions to be answered, the server will perform word segmentation on them. The word segmentation process uses professional word segmentation tools or algorithms to split the questions to be answered into meaningful vocabulary units based on the specific vocabulary rules and grammatical structures in the logistics field. For example, for the question to be answered "What is the abnormal package processing process", the server will segment it into "abnormal package", "processing process", "is", and "what". Afterwards, according to the sentence format required by the preset keyword matching algorithm rules, the word segmentation results are screened and reorganized to generate query statements suitable for the algorithm, such as "abnormal package processing process" (removing some auxiliary words that have little effect on keyword matching, etc.).

[0060] The server pre-builds an inverted index for the documents in the knowledge base. The inverted index uses keywords that appear in the documents as keys and records a list of documents containing the keywords, along with information such as the location and frequency of the keywords within the documents. When the server receives a query based on a keyword matching algorithm, it uses the inverted index to quickly locate documents containing those keywords. For example, if the query is "abnormal package handling process," the server will use the inverted index to find documents containing both keywords.

[0061] For documents found that contain query keywords, the server will calculate the weight of the keywords in the document based on factors such as the frequency and position of the keywords in the document. Generally speaking, the more frequently a keyword appears in a document and the more important its position is (such as the title, the beginning of a paragraph, etc.), the greater its weight. The server will comprehensively calculate the weights of all query keywords in the document to obtain the relevance score between the document and the query statement. Finally, the server will sort the documents in descending order according to the relevance score and select documents with higher relevance scores as the candidate document set. For example, in document A, "abnormal package" appears 3 times and "processing process" appears once, and both appear at the beginning of a paragraph; in document B, "abnormal package" appears once and "processing process" appears twice, but the appearance position is relatively late. After calculation, the server believes that document A has a higher relevance to the query statement and ranks it before document B.

[0062] Furthermore, after obtaining a preliminary set of candidate documents, the server may perform some optimization operations. For example, it may remove duplicate documents to prevent the same document from appearing multiple times in the candidate set; it may also perform denoising on the documents to remove content irrelevant to the query and improve the quality of the candidate documents.

[0063] For example, candidate documents are retrieved from the knowledge base based on a vector similarity algorithm. The specific process may include:

[0064] Word segmentation and query vector calculation of the question to be answered: After the server receives the question to be answered, it performs word segmentation on it and breaks it into meaningful vocabulary units. For example, for the question to be answered, "What is the process of handling abnormal packages?", the word segmentation results are "abnormal packages", "handling process", "is", and "what". After that, the server calculates the query vector of the question to be answered. The bag-of-words model, TF-IDF model, word embedding model (such as Word2Vec, GloVe), and other methods can be used to convert the text after word segmentation into a vector representation. These models can encode the semantic information of the question to be answered in the form of a vector, so that questions with similar semantics are closer in the vector space. For example, using the Word2Vec model, each word after word segmentation is converted into a corresponding word vector, and then the query vector of the question to be answered is obtained by weighted averaging or other aggregation methods.

[0065] The server also uses the aforementioned vector representation method to convert documents in the knowledge base into vector representations. For each word in the document, the same word vector model is used to convert it into a word vector. Then, through aggregation methods (such as weighted averaging), the document vector is obtained. This way, each document in the knowledge base has a corresponding vector representation that reflects the document's semantic information.

[0066] After obtaining the document vector and query vector, the server calculates the similarity between them. Commonly used methods for calculating vector similarity include cosine similarity and Euclidean distance. Cosine similarity measures the similarity between two vectors by calculating the cosine of the angle between them. The closer the value is to 1, the more similar the two vectors are. Euclidean distance calculates the straight-line distance between two vectors. The smaller the value, the more similar the two vectors are. The server calculates the similarity between all document vectors and query vectors in the knowledge base to obtain a similarity score between each document and the question to be answered. The server then sorts the documents according to the similarity score and selects documents with higher similarity scores as the candidate document set. The server can set a similarity threshold based on actual needs. Only documents with a similarity score higher than the threshold will be selected into the candidate document set. For example, the server sets the similarity threshold to 0.7 and only selects documents with a similarity score greater than 0.7 as candidate documents.

[0067] In some possible embodiments of the present application, see Figure 3 The flowchart of synonym replacement shown in the figure includes the following steps:

[0068] S301. Decompose the question to be answered into multiple words.

[0069] S302: Identify industry-specific vocabulary among the multiple vocabulary, and determine synonyms of the industry-specific vocabulary.

[0070] S303: After replacing the industry-specific vocabulary in the question to be answered with corresponding synonyms, at least one similar question associated with the question to be answered is generated.

[0071] S304: Generate corresponding query statements based on the question to be answered and at least one related similar question.

[0072] Among them, after the server receives the question to be answered, it will use professional word segmentation technology to split it. In the field of logistics, there are a large number of specific professional terms and compound words, and ordinary word segmentation tools may not be able to accurately split them. Therefore, the server will use a word segmentation model trained based on the logistics field corpus. The model has learned the vocabulary distribution and grammatical structure of the text in the logistics field. For example, for the question to be answered "How to deal with customs clearance delays in cross-border logistics", the server will split it into multiple words such as "how", "handle", "cross-border logistics", "in", "of", "customs clearance delays", and "situation". This word segmentation method not only takes into account the literal meaning of the words, but also combines the professional knowledge in the logistics field to ensure the accuracy of the segmentation.

[0073] After word segmentation is complete, the server verifies the results. Using the vocabulary and grammatical rules of the logistics field, the server checks whether each word conforms to the conventional expression conventions in the field. If a word is found to be improperly segmented, the server adjusts the word segmentation. For example, for the phrase "cold chain transport vehicle dispatch," if the initial word segmentation is "cold chain," "transport vehicle," and "dispatching," the server will detect that "cold chain transport" is a more commonly used professional word combination and adjust it to "cold chain transport" and "vehicle dispatch."

[0074] The server builds a specialized vocabulary database for the logistics sector, containing commonly used professional terms, concepts, and names within the industry. After word segmentation, the server matches the vocabulary list against the specialized vocabulary database to identify industry-specific terms. For example, in the word segmentation result "How to optimize warehouse layout to improve efficiency," the server identifies "warehouse layout" as a specialized term within the logistics sector.

[0075] For the identified industry-specific terms, the server uses a variety of methods to determine their synonyms. On the one hand, it refers to professional dictionaries and literature in the logistics field, which usually record synonyms or near-synonyms of professional terms. On the other hand, the server analyzes a large amount of text data in the logistics field to count the frequency and contextual relationships of other words with similar semantics to professional terms. For example, for "warehouse layout", the server found synonyms such as "warehouse planning" and "inventory arrangement" in professional dictionaries and literature. At the same time, through text data analysis, it was found that these words often appear in contexts related to "warehouse layout".

[0076] The server replaces industry-specific terms in the question with corresponding synonyms. During the replacement process, the server considers the semantic similarity and applicability of the synonyms to the original term. For example, for the question "How to improve the efficiency of cold chain transportation," replacing "cold chain transportation" with the synonym "refrigerated transportation" generates the similar question "How to improve the efficiency of refrigerated transportation."

[0077] For certain industry-specific terms, there may be multiple synonyms. The server combines these synonyms to generate multiple, similar questions. For example, for the term "warehouse layout," the server replaces it with "warehouse planning" and "inventory arrangement," generating similar questions like "How to optimize warehouse planning to improve efficiency" and "How to optimize inventory arrangement to improve efficiency." The server also considers partial replacements, replacing only some of the specialized terms in a question to generate a wider variety of similar questions.

[0078] The server generates rules based on pre-set query statements, converting the question to be answered and related, similar questions into query statements. These rules take into account the search methods and requirements of the logistics knowledge base. For example, for the question to be answered, "How to deal with customs clearance delays in cross-border logistics," the server removes auxiliary words and function words that are not meaningful to the search, such as "how," "in," "of," and "situation," and generates the query statement "cross-border logistics customs clearance delay handling." For the similar question, "How to solve cross-border logistics customs clearance delays," the server also removes irrelevant words and generates the query statement "cross-border logistics customs clearance delay solution."

[0079] After generating a query, the server optimizes it. It checks whether the terms in the query have a high degree of match in the knowledge base. If any terms are found to be potentially impacting search results, the server replaces or adjusts them. For example, for the query "the last mile problem in logistics delivery," the server might replace "problem" with the more common "problem," generating the optimized query "the last mile problem in logistics delivery" to improve search accuracy and efficiency within the knowledge base.

[0080] Furthermore, natural language processing (NLP) technology (such as WordNet), a synonym dictionary or a word vector model based on deep learning (such as Word2Vec, BERT, etc.) is used to determine at least one synonym corresponding to the industry-specific vocabulary.

[0081] The server first performs word segmentation on the question to be answered and identifies the industry-specific vocabulary. Then, the server searches WordNet based on the part of speech of the professional vocabulary. If the professional vocabulary is a noun, the server will search in the synonym set of the noun; if it is a verb, it will search in the synonym set of the verb. In the field of logistics, for the professional vocabulary "warehousing", after the server determines that it is a noun, it searches in the noun synonym set of WordNet and may get synonyms such as "storage" and "stowage". The server will filter these synonyms, remove some words that are not commonly used in the logistics context or have large semantic differences, and retain the synonyms that are closest in semantics to the original professional vocabulary in the logistics business.

[0082] After identifying the industry-specific terms in the question to be answered, the server performs an exact match query in a synonym dictionary within the logistics field. The server compares the terms against the entries in the synonym dictionary one by one. If a match is found, it retrieves a list of synonyms for that term. For the term "loading and unloading," a query in the synonym dictionary might yield synonyms such as "cargo handling involving loading and unloading operations." The server evaluates and selects these synonyms based on the actual logistics context and semantic similarity, determining the most appropriate synonym for subsequent question processing.

[0083] When the server uses the trained Word2Vec model to determine synonyms, it converts industry-specific vocabulary into corresponding word vectors. The server then calculates the cosine similarity between this word vector and other word vectors in the vocabulary. The closer the cosine similarity is to 1, the closer the directions of the two word vectors in the semantic space are, that is, the higher the semantic similarity. The server sorts the words in descending order of cosine similarity and selects the words with the highest similarity ranking as synonyms. For the professional vocabulary "port", after calculating it through the Word2Vec model, the server may obtain synonyms such as "wharf" and "seaport".

[0084] When the server uses the BERT model to identify synonyms, it places the industry-specific term in a context sentence and then uses the BERT model to obtain a vector representation of the term in that context. The server uses the hidden layer vector output by the BERT model as the semantic representation of the term. Similar to Word2Vec, the server calculates the similarity (e.g., cosine similarity) between the target term vector and other term vectors in the vocabulary and selects synonyms based on the similarity ranking. For example, for the specialized term "logistics information system," the server places it in a context sentence related to the logistics business, such as "Logistics information systems play a vital role in the operations of logistics companies." The server then uses the BERT model to obtain its vector representation and calculates similarity with other terms. This might yield synonyms such as "logistics information platform" and "logistics management information system." The server then performs further semantic analysis and screening on these synonyms to ensure that their semantics are highly consistent with the original term in the context of the logistics business.

[0085] Furthermore, in some embodiments of the present application, after receiving the question to be answered, the server uses professional text analysis technology to scan and identify the words in the question one by one. A professional vocabulary detection rule base based on the knowledge system of the logistics field is constructed. The rule base covers various common professional terms, concepts and names in the logistics industry, such as "intermodal transport", "cold chain logistics", "bonded warehouse", etc. The server will match each word in the question to be answered with the professional vocabulary in the rule base, and determine whether the question contains industry professional vocabulary through various methods such as string matching and semantic similarity calculation. For example, for the question to be answered "How to optimize transportation connections in intermodal transport", the server detects the industry professional vocabulary "intermodal transport" through matching.

[0086] The server pre-acquires a set of specialized terms associated with the pre-trained generative model. This set is derived from a large collection of logistics literature, industry reports, and expert knowledge, and aligns with the expertise used during the training of the generative model. When the server detects an industry-specific term in a question, it compares it with the associated set of specialized terms. If the term is found in the set, it indicates that it falls within the knowledge scope of the generative model, and the server can proceed directly to the subsequent question-solving process.

[0087] If the detected industry-specific vocabulary is not in the set of professional vocabulary associated with the pre-trained large language model for generation, the server will activate the online search function. The server will call a professional search engine interface and submit this industry-specific vocabulary as a search keyword to the search engine. During the search process, the server will set some search parameters, such as limiting the search scope to authoritative websites, industry forums, professional databases, etc. in the logistics field, to improve the accuracy and relevance of search results. For example, for the relatively new or less common industry-specific vocabulary "intelligent warehousing sorting robot" that appears in the question to be answered, after the server fails to find a match in the professional vocabulary set, it will search in the logistics-related fields through the search engine.

[0088] After the search engine returns a large number of search results, the server will screen and analyze these results. It will judge which texts can accurately and comprehensively explain this industry-specific vocabulary based on information such as the titles, abstracts, and body contents of the search results. The server will preferentially select the explanatory content published by authoritative institutions (such as logistics industry associations, official websites of well-known logistics enterprises, professional academic journals, etc.). For example, for "intelligent warehousing sorting robot", the server may screen out the detailed introduction texts about the technical principles, application scenarios, advantages, etc. of this robot published by the logistics industry association. After screening out the appropriate explanatory content text, the server will add it to the question to be answered. The addition method can be presented in the form of annotations, supplementary explanations, etc. directly after the original question, or the explanatory content can be integrated with the original question to form a new question containing a detailed explanation of the professional vocabulary. In this way, the large language model for generation can better understand the meaning of this industry-specific vocabulary when processing the question, thereby improving the accuracy and comprehensiveness of question answering.

[0089] It should be noted that the knowledge base includes a vector database to retrieve candidate documents by means of vector similarity. The generation process of the vector database is described below.

[0090] The original knowledge base documents may contain a lot of noise such as duplicate information, advertising content, meaningless characters or special symbols, etc. The server uses text analysis algorithms to identify and remove these irrelevant contents, and only retains the core information valuable to the question answering system. For example, in the knowledge documents in the logistics field, remove the advertising slogans irrelevant to the business, duplicate copyright statements, etc., to ensure the purity of the text content.

[0091] Normalize the cleaned text, including unifying the case, removing stop words, and cleaning punctuation marks. Unifying the case helps to avoid semantic confusion caused by case differences; removing stop words (such as "de", "shi", "zai", etc.) can reduce unnecessary interference and highlight the key information in the text; cleaning punctuation marks can further simplify the text and make subsequent processing more efficient.

[0092] For long documents, the server uses a text segmentation algorithm to break them into shorter sentences or paragraphs. This step not only improves the efficiency of subsequent vectorization processing but also enhances retrieval accuracy. For example, a long document on logistics and warehousing management can be segmented into multiple paragraphs covering aspects such as warehouse layout, inventory control, and equipment maintenance. Each paragraph is vectorized as an independent unit of information, facilitating more accurate matching of query requirements during retrieval.

[0093] Terminology standardization: In specific fields like logistics, specialized terms are numerous and often synonymous. The server uses a domain-specific vocabulary to standardize terminology within documents, unifying different synonyms or terms into a standardized form. For example, "cold chain logistics," "refrigerated transportation," and "low-temperature logistics" can be standardized into "cold chain logistics." This enhances the semantic consistency of the text and improves the accuracy of subsequent vectorization and retrieval.

[0094] Disambiguation: Polysemous and ambiguous words are common in text. The server uses contextual understanding technologies (such as deep learning models like BERT) to disambiguate. By analyzing the semantics of words in context, the server accurately determines their meaning, ensuring the accuracy of the text's semantics. For example, in a logistics context, the word "car" could refer to a transport vehicle or a train car. The server determines its specific meaning based on the context, avoiding misunderstandings in subsequent searches.

[0095] Vectorization is the process of converting preprocessed text into digital representations for similarity calculation and retrieval.

[0096] Traditional word embedding methods, such as Word2Vec and GloVe, represent text by mapping each word into a fixed-dimensional vector. These methods are suitable for traditional low-dimensional vector models and offer high computational efficiency, but they have limitations in processing contextual information and semantic relationships, failing to accurately capture the semantic variations of words in different contexts.

[0097] Transformer-based models, such as BERT, RoBERTa, and Sentence-BERT, are able to better capture contextual information and the deeper semantics of sentences. Pre-trained on large-scale corpora, they acquire rich linguistic knowledge and semantic representations, generating high-quality semantic vector representations for each sentence or document. They are widely used in text semantic search and question-answering systems. In the vectorization of logistics knowledge bases, Transformer-based models can more accurately understand specialized terminology and complex semantics, improving retrieval accuracy.

[0098] Document vectorization process: Preprocessed documents (such as paragraphs or sentences) are embedded using the selected pre-trained model. For Transformer-based models, documents can be directly input into the model, which outputs the document's embedding vector. For traditional word vector methods, the document embedding vector can be generated by calculating the average of all word vectors in a sentence or paragraph. For example, the Sentence-BERT model is used to vectorize a paragraph in a logistics document, generating a fixed-dimensional vector that accurately reflects the semantic information of the paragraph.

[0099] The embedding vectors for all documents are stored in a vector database. This database uses specialized data structures and storage methods to efficiently store and retrieve vector data. During storage, the server encodes and compresses the vectors to reduce storage space while ensuring retrieval performance. For example, using a vector database tool like FAISS (Facebook AI Similarity Search) to store vector data for logistics documents in the database is crucial for subsequent retrieval.

[0100] Index construction is to improve the efficiency of vector retrieval. By establishing an index structure, document vectors similar to the query vector can be quickly located.

[0101] Choose an appropriate index type based on the scale of your vector database and your search requirements. Common index types include FlatIndex, HierarchicalNavigableSmallWorld (HNSW), and Locality-Sensitive Hashing (LSH). FlatIndex has a simple structure and is suitable for small-scale data. HNSW indexes can efficiently process large-scale data and offer good search performance. LSH indexes are suitable for approximate search scenarios, improving search speed while maintaining a certain level of search accuracy.

[0102] The server creates an index for the vector data in the vector database based on the selected index type. During the index creation process, the server extracts and analyzes the vector features and maps them to the index structure. For example, when using the HNSW index, the server constructs a hierarchical clustering structure, grouping similar vectors together to form an efficient retrieval network.

[0103] S203: Search the knowledge base using a corresponding search algorithm based on the generated multiple query statements to obtain candidate document sets corresponding to the multiple query statements.

[0104] The server maintains a knowledge base that stores a vast amount of logistics-related knowledge. This knowledge base may contain knowledge in various forms, such as documents and database records. The server sends the generated query statements to the knowledge base's retrieval module. The retrieval module matches and filters the knowledge base based on the different query statements. For query statements based on keyword matching, the retrieval module searches for documents containing these keywords within the document. For query statements based on semantic understanding, the retrieval module analyzes the document's semantic information to identify documents that are semantically relevant to the query statement. Ultimately, the server obtains a set of candidate documents corresponding to each query statement. The documents in these sets are all related to each other, but there may be duplication or significant differences in the retrieval results of different algorithms.

[0105] For example, the server sends two queries for Xiao Li's question to the knowledge base for search. For the query "Abnormal Package Handling Process," the search module finds multiple documents containing these two keywords in the knowledge base, such as "Abnormal Package Handling Operation Manual" and "Logistics Abnormal Situation Response Guide." For the semantic query "Obtaining the Steps for Handling Abnormal Packages," the search module also finds some semantically relevant documents, such as the section describing the steps in "Logistics Abnormal Handling Case Analysis." The server combines these documents into two candidate document sets.

[0106] In some embodiments of the present application, the server utilizes the operating system's multi-threading mechanism to create a separate thread for each search path. Each thread executes the corresponding search algorithm logic in its own memory space, running in parallel with other threads. The advantage of multi-threaded parallel execution is that it can fully utilize the server's multi-core processor resources and improve computing efficiency. For example, on a server with a quad-core processor, four threads can be started simultaneously, each executing a different search path, allowing the search task to be completed in a shorter time.

[0107] For large-scale knowledge bases and complex search tasks, servers may employ distributed computing frameworks (such as Hadoop and Spark) to parallelize multi-way searches. Distributed computing frameworks break down search tasks into multiple subtasks and distribute these subtasks to different compute nodes for parallel execution. Each compute node processes a portion of the knowledge base data and search algorithm logic, then aggregates the results to the master node. This approach leverages the advantages of cluster computing to handle massive amounts of data and complex search algorithms, further improving the response speed and scalability of multi-way searches.

[0108] Furthermore, in multi-way search, different search tasks may have different execution times due to factors such as algorithm complexity and data size. If synchronous execution is used, where one search task completes before the next, the entire multi-way search process will be dominated by the most time-consuming task, resulting in longer system response times and a poor user experience.

[0109] The asynchronous execution strategy allows different retrieval tasks to be executed independently, without relying on the completion of other tasks. After each retrieval task is started, it will perform data processing and retrieval operations according to its own execution process. The server can process other tasks or respond to requests from other users at the same time. When a retrieval task is completed, the server will obtain its results in a timely manner and integrate them with the results of other retrieval tasks. For example, in a retrieval task based on a keyword matching algorithm, after the server quickly locates some relevant documents, it can save these results first and continue to execute other retrieval tasks without having to wait for all retrieval tasks to be completed before integrating the results.

[0110] S204: Fusing the candidate document sets according to a preset fusion rule to obtain a fused document set.

[0111] The server pre-sets fusion rules for integrating multiple candidate document sets. These fusion rules may consider factors such as a document's relevance score, source reliability, and document content duplication. The server performs a relevance assessment on the documents in each candidate document set and calculates a relevance score. Then, based on the fusion rules, the server filters and merges documents from different candidate document sets. For example, duplicate documents are removed, while documents with high relevance scores and reliable sources are retained. This process allows the server to obtain a more comprehensive and accurate fused document set, the documents in which can better answer user questions.

[0112] For example, the server merges the two candidate document sets corresponding to Xiao Li's question. During the fusion process, the server finds that the "Abnormal Package Handling Operation Manual" exists in both candidate document sets and has a high relevance score, so it retains this document. Simultaneously, based on a source reliability assessment, the server retains the "Logistics Abnormal Situation Response Guide," published by an authoritative organization in the logistics industry, while removing some related documents from unidentified sources. Ultimately, the server obtains a fused document set that contains comprehensive information about the abnormal package handling process.

[0113] In some possible embodiments of the present application, fusing the candidate document sets according to a preset fusion rule to obtain a fused document set includes:

[0114] Calculate the text similarity between the question to be answered and the candidate documents retrieved by multiple retrieval algorithms, sort them in descending order according to the text similarity, and then select several candidate documents in the front as the fusion document set; or

[0115] Calculate the semantic similarity between the question to be answered and the candidate documents retrieved by multiple retrieval algorithms, sort them in descending order according to the calculation results, and then select several candidate documents in the front as the fusion document set; or

[0116] Deduplicate the candidate documents retrieved by multiple retrieval algorithms according to the content similarity; calculate the feature vectors of each candidate document after deduplication; calculate the feature vectors of each candidate document according to the pre-trained ranking model, and output a score; sort each candidate document in descending order according to the score, and select several documents with the highest scores as the fusion document set.

[0117] Among them, the fusion method based on text similarity: After the server receives the question to be answered and the set of candidate documents retrieved by multiple retrieval algorithms (such as keyword matching algorithm, vector similarity algorithm, etc.), it will calculate the text similarity between each candidate document and the question to be answered. The text similarity calculation is mainly based on the matching degree at the string level. The server will preprocess the question to be answered and the candidate documents, such as removing stop words (such as words like "of", "is", "in", etc. that have little impact on the text meaning semantically), performing stemming (restoring different morphological words to their basic forms), etc. Then, methods such as edit distance and Jaccard similarity coefficient are used to measure the text similarity between the two. Edit distance refers to the minimum number of edit operations (including inserting, deleting, replacing characters) required to convert one string into another string. The smaller the edit distance, the higher the text similarity; the Jaccard similarity coefficient measures the similarity by calculating the ratio of the size of the intersection of two sets to the size of the union. Here, the question to be answered and the candidate documents can be regarded as sets composed of words respectively.

[0118] Sorting and screening: After the server calculates the text similarity between each candidate document and the question to be answered, it will sort all candidate documents in descending order according to the similarity score. The higher the similarity score, the closer the candidate document is to the question to be answered at the text level. Then, the server will select several candidate documents in the front as the fusion document set according to the preset threshold or quantity requirement. For example, the server can set to only select the top 10 candidate documents with the highest similarity scores. These documents have a high degree of matching with the question to be answered at the text level and are more likely to contain the information required by the user.

[0119] For example, in the question-and-answer system of a logistics company, the question to be answered is "Process for handling damaged express parcels". The candidate documents retrieved based on the keyword matching algorithm include "Standards for Handling Damaged Express Parts" and "Manual for Handling Abnormal Package Situations", etc., and the candidate documents retrieved based on the vector similarity algorithm include "Logistics Abnormality Handling Process (Including Damaged Package Section)" and "Measures for Handling Damaged Packages in Transport". The server calculated the text similarity between these candidate documents and the question to be answered, and found that the words "damaged" and "handling" that are closely related to the question to be answered appeared many times in the "Standards for Handling Damaged Express Parts", and its text similarity score was relatively high. After sorting, the server selected the top 8 candidate documents with the highest text similarity scores to form a fused document set. These documents are closest to the question to be answered in terms of text.

[0120] Fusion method based on semantic similarity: Unlike text similarity, semantic similarity pays more attention to the semantic connotation expressed by the text. The server will use pre-trained semantic models (such as BERT, GPT, etc.) to convert the questions to be answered and candidate documents into semantic vector representations. These models have been trained on large-scale corpora and can capture the semantic information in the text. Then, the server calculates indicators such as cosine similarity and Euclidean distance between the vector of the question to be answered and the vector of the candidate document to measure the semantic similarity. The closer the cosine similarity is to 1, the closer the directions of the two vectors in the semantic space are, that is, the higher the semantic similarity; the smaller the Euclidean distance, the closer the distance between the two vectors in the semantic space is, and the higher the semantic similarity.

[0121] The server sorts all candidate documents in descending order based on the calculated semantic similarity scores. Candidate documents with high semantic similarity scores are more semantically consistent with the question being answered and can better answer the user's question. The server then follows pre-defined rules, such as selecting the top N candidate documents by semantic similarity score as the fused document set. For example, the server might set N to 12, thus selecting the 12 candidate documents with the highest semantic similarity scores as part of the fused document set.

[0122] For example, in a logistics enterprise question-and-answer scenario, the question to be answered is "How to deal with express delivery delays caused by bad weather." The server uses a semantic model to convert the question to be answered and each candidate document into a vector and then calculates the semantic similarity. It was found that the "Logistics Distribution Adjustment Plan under Bad Weather" detailed how to adjust the delivery plan in bad weather to deal with possible express delivery delays, and its semantic vector had a high cosine similarity with the question to be answered vector. After sorting all candidate documents in descending order of semantic similarity, the server selected the top 12 candidate documents to form a fused document set. These documents are semantically highly relevant to the question to be answered.

[0123] Based on the fusion of feature vectors and ranking models: The server first deduplicates candidate documents retrieved by multiple retrieval algorithms based on content similarity. Content similarity can be calculated using a variety of methods, such as the TF-IDF method based on the bag-of-words model, which represents documents as weighted vectors of vocabulary and then calculates the similarity between vectors. Alternatively, a word embedding model can be used to convert documents into vectors and then calculate similarity. The server sets a similarity threshold. When the content similarity between two documents exceeds this threshold, they are considered duplicate or highly similar documents, and only one of them is retained. This prevents the presence of a large number of duplicate or highly similar documents in the fused document collection, improving the quality and diversity of the document collection.

[0124] For each candidate document after deduplication, the server calculates its feature vector. This feature vector can contain information from multiple dimensions, such as document length (measured by the number of characters or words), keyword frequency (counting the number of times keywords related to the question appear in the document), and document timeliness (calculating a timeliness score based on the document's publication or update time). The server normalizes these features to make them comparable and then combines them into a feature vector, which represents the attributes of each candidate document.

[0125] The server has pre-trained a ranking model, which can be built based on machine learning algorithms (such as logistic regression, decision trees, neural networks, etc.). The ranking model learns how to predict the relevance of a document to the question to be answered based on the feature vectors of the candidate documents by learning a large amount of labeled data (that is, data on the relevance of known documents to the query questions). The server inputs the feature vectors of each candidate document after deduplication processing into the pre-trained ranking model, and the model outputs a score, which represents the relevance score of the candidate document to the question to be answered. The higher the score, the more relevant the candidate document. The server sorts the candidate documents in descending order according to the score, and then selects several documents with the highest scores as the fused document set. For example, the server can be set to select the top 15 candidate documents.

[0126] For example, in a logistics enterprise question-and-answer system, multiple retrieval algorithms returned multiple candidate documents. The server first performed content similarity deduplication processing to remove documents with highly similar content. Then, the feature vectors of the remaining candidate documents were calculated. For example, the feature vector of document A contained information such as a length of 500 words, a frequency of 3 occurrences of the keyword "logistics delay", and a timeliness score of 0.8. The server input these feature vectors into the pre-trained ranking model, and the model outputted a score of 0.92 for document A. After performing the same operation on all candidate documents, the server sorted them in descending order by score and selected the top 15 candidate documents to form a fused document set. These documents have better comprehensive performance in multiple feature dimensions and are more likely to meet user needs.

[0127] S205 , extracting target answers from the fused document collection based on the pre-trained generative model.

[0128] The server pre-trains a large generative model with powerful text understanding and generation capabilities. The server takes the fused document collection as input and sends it to the large generative model. The large generative model analyzes and understands the text in the fused document collection, extracting key information relevant to the user's question. Based on this key information, the model then generates an accurate and concise target answer.

[0129] For example, the server sends the fused document collection corresponding to Xiao Li's question to the generative model. After analyzing the document content, the model extracts the key steps in the abnormal package handling process, such as "Step 1: Receive abnormal package information; Step 2: Verify the package anomaly; Step 3: Take appropriate action based on the type of anomaly." The model then generates the target answer: "The abnormal package handling process is as follows: Step 1: Receive abnormal package information; Step 2: Verify the package anomaly; Step 3: Take appropriate action based on the type of anomaly, such as return or reshipment."

[0130] S206: Push the target answer to the user.

[0131] After generating the target answer, the server sends it back to the user's client interface via a network interface in the required data format (such as JSON or XML). The client interface parses and displays the received answer, allowing the user to easily view and understand the answer content. The server must ensure the timeliness and accuracy of answer push to provide a good user experience.

[0132] For example, the server pushes the target answer to Xiao Li's question, "The abnormal package handling process is as follows: Step 1, receive abnormal package information; Step 2, verify the package abnormality; Step 3, take appropriate measures based on the abnormality type, such as return or reshipment," to the client interface used by Xiao Li. After seeing this answer on the client interface, Xiao Li can clearly understand the abnormal package handling process and better complete his work.

[0133] This application has the following beneficial effects:

[0134] Using multi-channel recall fusion technology, based on multiple preset retrieval algorithms, the question to be answered is converted into query statements corresponding to different algorithms. These are then searched separately within the knowledge base to generate multiple sets of candidate documents. This multi-dimensional retrieval approach analyzes queries from different perspectives, greatly increasing the likelihood of retrieving relevant documents and effectively improving recall rates. This ensures that more relevant information is retrieved, providing richer material for subsequent extraction of the target answer.

[0135] Multi-recall fusion technology utilizes multiple retrieval algorithms to conduct searches from multiple perspectives. Each algorithm has its own unique advantages and applicable scenarios. When one algorithm exhibits deviations or poor retrieval performance in a specific situation, the results of other algorithms can complement and correct the results, making the resulting fused document collection more comprehensive and accurate. This allows the target answer to be extracted from a more reliable and stable document collection, thereby enhancing the robustness of the overall intelligent question-answering system and reducing the probability of incorrect answers due to algorithmic bias.

[0136] The downstream generative model needs to extract the target answer from a collection of candidate documents. The quality of the candidate document collection directly affects the performance of the generative model. Multi-channel recall fusion technology optimizes the quality and diversity of retrieval results, providing higher-quality and more comprehensive input data for the generative model. This enables the generative model to analyze and judge the target answer based on richer and more accurate information, thereby improving the accuracy and reliability of the generated target answer. This further enhances the performance of downstream generative tasks in the entire intelligent question-answering system, providing users with higher-quality and more satisfying question-answering services.

[0137] The following are device embodiments of the present application, which can be used to implement the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the method embodiments of the present application.

[0138] See Figure 4, which shows a schematic diagram of the structure of a large-model-based intelligent question-answering device provided by an exemplary embodiment of the present application, hereinafter referred to as device 4. Device 4 can be implemented as all or part of a server through software, hardware, or a combination of both. Device 4 includes: a receiving unit 401, a conversion unit 402, a retrieval unit 403, a fusion unit 404, an extraction unit 405, and a feedback unit 406.

[0139] Receiving unit 401, for receiving a question to be answered input by a user;

[0140] A conversion unit 402 is configured to convert the question to be answered into query statements corresponding to the plurality of retrieval algorithms according to the formats indicated by the plurality of preset retrieval algorithms;

[0141] A retrieval unit 403 is configured to search the knowledge base according to the generated multiple query statements to obtain candidate document sets corresponding to the multiple query statements.

[0142] A fusion unit 404 is configured to fuse the candidate document sets according to a preset fusion rule to obtain a fused document set;

[0143] An extraction unit 405 is configured to extract a target answer from the fused document set based on a pre-trained generative model;

[0144] The feedback unit 406 is configured to push the target answer to the user.

[0145] In one or more possible embodiments, multiple query statements are used to initiate multiple query tasks in parallel, and candidate documents retrieved using each query statement are obtained asynchronously.

[0146] In one or more possible embodiments, fusing the candidate document sets according to a preset fusion rule to obtain a fused document set includes:

[0147] Calculate the text similarity between the question to be answered and the candidate documents retrieved by multiple retrieval algorithms, sort them in descending order based on the text similarity, and then select the top candidate documents as the fused document set; or

[0148] Calculate the semantic similarity between the question to be answered and the candidate documents retrieved by multiple retrieval algorithms, sort them in descending order based on the calculated results, and then select the top candidate documents as the fused document set; or

[0149] De-duplicate candidate documents retrieved by multiple retrieval algorithms based on content similarity; calculate the feature vectors of each candidate document after de-duplication; calculate the feature vectors of each candidate document based on a pre-trained ranking model and output a score; sort each candidate document in descending order based on the score, and select the documents with the highest scores as the fused document set.

[0150] In one or more possible embodiments, the multiple retrieval algorithms include: a keyword matching algorithm and a vector similarity algorithm.

[0151] In one or more possible embodiments, converting the question to be answered into query statements corresponding to each of the multiple search algorithms includes:

[0152] Break down the question to be answered into multiple words;

[0153] Identify industry-specific vocabulary among multiple vocabularies and determine synonyms of the industry-specific vocabulary;

[0154] After replacing the industry-specific vocabulary in the question to be answered with corresponding synonyms, at least one similar question related to the question to be answered is generated;

[0155] A corresponding query statement is generated according to the question to be answered and at least one related similar question.

[0156] In one or more possible embodiments, determining synonyms of industry-specific vocabulary includes:

[0157] Convert industry-specific vocabulary into corresponding word vectors;

[0158] Calculate the cosine similarity between the word vector and other word vectors in the vocabulary;

[0159] Sort the words in descending order of cosine similarity;

[0160] Select several words with the highest similarity ranking as synonyms.

[0161] In one or more possible embodiments, it further includes:

[0162] An annotation unit is used to obtain a set of professional vocabulary associated with the pre-trained generation model if the question to be answered includes industry professional vocabulary; if the industry professional vocabulary is not in the professional vocabulary set, search for the explanatory content text of the industry professional vocabulary in an online manner, and add the explanatory content text to the question to be answered.

[0163] It should be noted that the device 4 provided in the above embodiment only uses the division of the above functional modules as an example when executing the intelligent question-answering method based on a large model. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the above functions. In addition, the intelligent question-answering device based on a large model provided in the above embodiment and the embodiment of the intelligent question-answering method based on a large model are of the same concept. The implementation process is detailed in the method embodiment and will not be repeated here.

[0164] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0165] See also Figure 5 FIG. 1 is a schematic diagram of a computer storage medium provided in an embodiment of the present application, wherein the computer storage medium can store multiple instructions (i.e. Figure 5 The computer program shown in FIG. 1 is suitable for being loaded by a processor and executed as described above. Figure 2 The method steps of the embodiment shown, the specific execution process can be found in Figure 2 The detailed description of the illustrated embodiment will not be repeated here.

[0166] The present application also provides a computer program product, which stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the large model-based intelligent question-answering method as described in the above embodiments.

[0167] See Figure 6 , provides a structural diagram of a server according to an embodiment of the present application. Figure 6 As shown, the server 600 may include: at least one processor 601 , at least one network interface 604 , a user interface 603 , a memory 605 , and at least one communication bus 602 .

[0168] The communication bus 602 is used to implement the connection and communication between these components.

[0169] The user interface 603 may include input units such as a mouse and a keyboard.

[0170] The network interface 604 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0171] The processor 601 may include one or more processing cores. The processor 601 utilizes various interfaces and lines to connect various components within the server 600. By running or executing instructions, programs, code sets, or instruction sets stored in the memory 605, and calling data stored in the memory 605, the processor 601 executes various functions of the server 600 and processes data. Optionally, the processor 601 may be implemented using at least one hardware form selected from the group consisting of digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 601 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; and the modem is responsible for handling wireless communications. It is understood that the modem may not be integrated into the processor 601 and may be implemented separately on a single chip.

[0172] Among them, the memory 605 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read-Only Memory). Optionally, the memory 605 includes a non-transitory computer-readable storage medium. The memory 605 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 605 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 605 may also be optionally at least one storage device located away from the aforementioned processor 601. As Figure 6 As shown, the memory 605 as a computer storage medium may include an operating system, a network communication module, a user interface module, and an application program.

[0173] exist Figure 6In the server 600 shown in FIG. 1 , the user interface 603 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 601 can be used to call the application stored in the memory 605 and specifically execute the following operations: Figure 2 The specific process can be referred to the method shown in Figure 2 As shown, no further details are given here.

[0174] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0175] The above disclosure is only a preferred embodiment of the present application, and certainly cannot be used to limit the scope of rights of the present application. Therefore, equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. An intelligent question answering method based on a large model, characterized in that: include: Receive questions to be answered from users; Converting the question to be answered into query statements corresponding to the respective retrieval algorithms according to the formats indicated by the preset retrieval algorithms; Searching the knowledge base according to the generated multiple query statements to obtain candidate document sets corresponding to the multiple query statements; According to the preset fusion rules, each candidate document set is fused to obtain a fused document set; Extracting target answers from the fused document collection based on a pre-trained generative model; The target answer is pushed to the user.

2. The method according to claim 1, characterized in that Multiple query statements are used in parallel to initiate multiple query tasks, and candidate documents retrieved by each query statement are obtained asynchronously.

3. The method according to claim 1 or 2, characterized in that The step of fusing the candidate document sets according to the preset fusion rules to obtain the fused document set includes: Calculate the text similarity between the question to be answered and the candidate documents retrieved by multiple retrieval algorithms, sort them in descending order based on the text similarity, and then select the top candidate documents as the fused document set; or Calculate the semantic similarity between the question to be answered and the candidate documents retrieved by multiple retrieval algorithms, sort them in descending order based on the calculated results, and then select the top candidate documents as the fused document set; or De-duplicate candidate documents retrieved by multiple retrieval algorithms based on content similarity; calculate the feature vectors of each candidate document after de-duplication; calculate the feature vectors of each candidate document based on a pre-trained ranking model and output a score; sort each candidate document in descending order based on the score, and select the documents with the highest scores as the fused document set.

4. The method according to claim 3, characterized in that The multiple search algorithms include: a keyword matching algorithm and a vector similarity algorithm.

5. The method according to claim 1, 2 or 4, characterized in that: The step of converting the question to be answered into query statements corresponding to the multiple search algorithms includes: Break down the question to be answered into multiple words; Identify industry-specific vocabulary among multiple vocabularies and determine synonyms of industry-specific vocabulary; Replace the industry-specific vocabulary in the question to be answered with corresponding synonyms, and generate at least one similar question related to the question to be answered; A corresponding query statement is generated according to the question to be answered and at least one related similar question.

6. The method according to claim 5, characterized in that The synonyms of the industry-specific vocabulary described above include: Convert industry-specific vocabulary into corresponding word vectors; Calculate the cosine similarity between the word vector and other word vectors in the vocabulary; Sort the words in descending order of cosine similarity; Select several words with the highest similarity ranking as synonyms.

7. The method according to claim 1 or 2 or 4 or 6, characterized in that Also includes: If the question to be answered includes industry-specific vocabulary, obtain the professional vocabulary set associated with the pre-trained generation model. If the industry-specific vocabulary is not in the professional vocabulary set, search for the explanatory content text of the industry-specific vocabulary in an online manner, and add the explanatory content text to the question to be answered.

8. An intelligent question-answering device, characterized in that: include: A receiving unit, configured to receive a question to be answered input by a user; a conversion unit, configured to convert the question to be answered into query statements corresponding to the plurality of retrieval algorithms according to the formats indicated by the plurality of preset retrieval algorithms; A retrieval unit, configured to search the knowledge base according to the generated multiple query statements, and obtain candidate document sets corresponding to the multiple query statements; A fusion unit, configured to fuse the candidate document sets according to a preset fusion rule to obtain a fused document set; An extraction unit, configured to extract a target answer from the fused document set based on a pre-trained generative model; A feedback unit is used to push the target answer to the user.

9. A computer storage medium, characterized in that The computer storage medium stores a plurality of instructions, which are suitable for being loaded by a processor and executing the method steps according to any one of claims 1 to 7.

10. A server, characterized in that: include: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method steps according to any one of claims 1 to 7.

Images