Question-answering method and apparatus using entity extraction, device, and medium

By using entity replacement technology and multidimensional feature similarity matching in the question-answering system, the problem of low recall rate in the financial and medical fields of traditional question-answering systems has been solved, achieving more efficient information retrieval and user interaction experience.

WO2026232782A1PCT designated stage Publication Date: 2026-11-12PING AN TECH (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/135615
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-06
Filing Date
2025-11-18
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Traditional question-answering systems struggle to comprehensively cover complex and diverse financial and medical questions, resulting in low recall rates and high maintenance costs.

Method used

The system uses entity replacement technology to identify similar questions in the corpus, replaces entities in user questions with specific referential symbols, performs multidimensional feature similarity matching in the updated corpus, obtains the target question, and returns the associated answer.

Benefits of technology

It improves the recall rate of the question-and-answer system, enabling it to better respond to complex and ever-changing user needs, provide high-quality and efficient services, and significantly enhance user experience and system performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025135615_12112026_PF_FP_ABST
    Figure CN2025135615_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A question-answering method and apparatus using entity extraction, a device, and a medium, the method comprising: determining all similar questions associated with each standard question in a first corpus, using specific referential symbols to perform replacement processing on entities in all of the similar questions and obtaining an updated second corpus; and using the specific referential symbols to perform replacement processing on entities in a user question, using the entity-replaced user question to perform question recall processing in the second corpus to acquire at least one second candidate question, and performing multi-dimensional feature similarity matching processing between the user question and each second candidate question to select a corresponding target question and return an associated answer. In this way, the retrieval range can be expanded, the most relevant similar question and the associated answer can be precisely matched from massive data, and the recall rate is effectively increased.
Need to check novelty before this filing date? Find Prior Art

Description

A question-answering method, apparatus, device, and medium applying entity extraction

[0001] This application claims priority to Chinese Patent Application No. 202510585596.4, filed on May 6, 2025, entitled "A Question-Answering Method, Apparatus, Device and Medium for Entity Extraction", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of question-answering technology in fintech and digital healthcare scenarios, and in particular to a question-answering method, apparatus, device and medium that uses entity extraction. Background Technology

[0003] In the operation of question-answering systems in the fintech and digital healthcare fields, the reliance on semantic matching technology alone presents numerous limitations, making it difficult to meet the increasingly complex needs of users. The inventors discovered that when faced with specific, standardized questions such as how to open a credit card account or how to access remote medical care, and when the related corpus is extremely broad, it is difficult for humans to exhaustively enumerate all possibilities, resulting in the system's inability to comprehensively cover the relevant content. This not only directly reduces the recall rate of the dialogue system in the information retrieval and response stages, impacting the user experience, but also significantly increases the maintenance costs for financial and healthcare operations teams. With the increasing complexity of financial and healthcare operational needs, financial and healthcare question-answering systems urgently require upgrades and iterations to adapt to the ever-growing demands. Summary of the Invention

[0004] This application provides a question-answering method, apparatus, computer device, and medium for applying entity extraction, in order to address the problem of how to overcome the limitations of traditional question-answering systems, comprehensively cover complex questions in the financial and medical fields, and thus improve the recall rate of dialogue systems.

[0005] Firstly, this application provides a question-answering method that applies entity extraction, including:

[0006] Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus;

[0007] The entities in the user question are replaced using specific referential symbols. The user question with the replaced entities is then used for question recall in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

[0008] Secondly, this application provides a question-answering device that applies entity extraction, comprising:

[0009] The corpus update module is used to identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions with specific referential symbols to obtain the updated second corpus;

[0010] The second query module is used to replace entities in the user question with specific referential symbols, and to perform question recall processing on the user question after entity replacement in the second corpus to obtain at least one second candidate question. The user question is then matched with all second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

[0011] Thirdly, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0012] Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus;

[0013] The user question is replaced with a specific symbol, and the user question with the replaced entity is retrieved in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

[0014] Fourthly, this application provides a non-volatile computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor causes the processor to perform the following operations:

[0015] Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus;

[0016] The user question is replaced with a specific symbol, and the user question with the replaced entity is retrieved in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

[0017] This application provides a question-answering method, apparatus, computer device, and storage medium for entity extraction. It pre-identifies all similar questions associated with each standard question in a first corpus, replaces entities in all similar questions using specific referential symbols to obtain an updated second corpus, and then replaces entities in user questions using the same specific referential symbols. The replaced user questions are then used for question recall within the second corpus to obtain at least one second candidate question. Finally, the user question is matched with all second candidate questions using multi-dimensional feature similarity to select the corresponding target question and return the associated answer. Through this method, in question-answering scenarios in finance and healthcare, entity replacement technology is used to analyze entity information in user questions, compare it with more data in the corpus, thereby expanding the retrieval scope to match relevant answers and effectively improving the recall rate. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 is a schematic diagram of an application environment for a question-and-answer method using entity extraction in one embodiment of this application.

[0020] Figure 2 is a flowchart illustrating a question-and-answer method using entity extraction in one embodiment of this application.

[0021] Figure 3 is a flowchart illustrating a specific implementation of step S201 in Figure 2.

[0022] Figure 4 is a schematic diagram of a specific implementation method after step S201 and before step S202.

[0023] Figure 5 is a flowchart illustrating a specific implementation of step S202 in Figure 2.

[0024] Figure 6 is a flowchart illustrating a specific implementation of step S503 in Figure 5.

[0025] Figure 7 is a structural schematic diagram of a question-and-answer device that applies entity extraction in one embodiment of this application.

[0026] Figure 8 is a structural schematic diagram of a computer device according to an embodiment of this application.

[0027] Figure 9 is another structural schematic diagram of a computer device according to one embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] The question-answering method for entity extraction provided in this application embodiment can be applied in the application environment shown in Figure 1, where the client communicates with the server via a network. The server provides operation nodes for operators to pre-confirm all similar questions associated with each standard question in the first corpus, and uses specific referential symbols to replace entities in all similar questions to obtain an updated second corpus. The server can receive user questions through the client, then use specific referential symbols to replace entities in the user questions, and perform question recall processing on the entity-replaced user questions in the second corpus to obtain at least one second candidate question. The user question is then matched with all second candidate questions using multi-dimensional feature similarity to select the corresponding target question and return the associated answer. In this application, for question-answering services in financial and medical scenarios, entity replacement technology is used to analyze the entity information in user questions, compare it with more data in the corpus, thereby expanding the retrieval scope to match relevant answers and effectively improving the recall rate. This enables the question-and-answer system to better address complex and ever-changing user needs, providing users with higher-quality and more efficient services, marking a significant step forward in the field of intelligent interaction in financial and medical scenarios. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this application.

[0030] Please refer to Figure 2, which is a flowchart of a question-and-answer method for application entity extraction provided in an embodiment of this application, including the following steps S201 and S202.

[0031] S201. Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus.

[0032] In this step, all similar questions associated with each standard question in the first corpus are also replaced using entity replacement technology, thereby updating the second corpus.

[0033] For example, the first corpus contains the standard question "[How to handle missing items in an order]", under which are similar questions such as "[Why are there only 3 boxes of medicine A, and no medicine B?]" and "[I bought two bottles of medicine C, why did I only receive one bottle]". Correspondingly, after entity replacement, the second corpus will contain the standard question "[How to handle missing items in an order]", under which will be similar questions such as "[Why are there only 3 boxes of medicine A?]" <product>,No <product>I bought two bottles. <product>Why did you only send one bottle?

[0034] For example, the first corpus contains the standard question "[Order not received for a long time]", under which there are similar questions such as "[Why hasn't medicine D arrived yet]". Correspondingly, after entity replacement, the second corpus will contain the standard question "[How to handle missing items in an order]", under which there will be similar questions such as "[Why...]". <product>It hasn't arrived yet.

[0035] S202. Replace the entities in the user question with specific referential symbols, and perform question recall processing on the user question after entity replacement in the second corpus to obtain at least one second candidate question. Then, perform multi-dimensional feature similarity matching processing on the user question and all second candidate questions to select the corresponding target question and return the associated answer.

[0036] The question-and-answer methods provided in this step can be applied to intelligent customer service or intelligent assistant question-and-answer engines in various application scenarios. Intelligent question-and-answer engines are usually implemented through a server-side component, which can receive user questions in real time through the client. For example, in the field of medical applications, users often ask questions via telephone or online chat, and it is often necessary to use an intelligent question-and-answer engine to answer some of the customer's medical questions in order to improve the efficiency of medical promotion and user experience.

[0037] For example, a user question received by the client might be, "I bought three bottles of Vitamin E, why did I only receive one?". After receiving this user question, the client needs to replace the entities in the user question with corresponding specific referential symbols to obtain the replaced user question, "I bought two bottles." <product>"Why did you only send one bottle?" Then, the replaced user question is used to perform question recall processing in the second corpus to obtain at least one second candidate question. The user question is then matched with all second candidate questions using multi-dimensional feature similarity to select the corresponding target question and return the associated answer.

[0038] In this embodiment, based on the techniques of steps S201-S202, a more in-depth analysis can be performed on complex user questions in financial and medical question-and-answer scenarios. When traditional question-and-answer systems fail, the system analyzes the entities within the user's question and compares them with a larger volume of similar questions. This expands the system's retrieval scope, enabling it to accurately match the most relevant similar questions and related answers from massive amounts of data, effectively improving the recall rate.

[0039] For example, in a financial Q&A scenario, a user asks, "Why hasn't the interest from my fixed deposit at XX Bank, which matured three years ago, been credited to my account?" In traditional Q&A systems, due to variations in question wording or limitations of the system's knowledge base, a direct answer might not be possible. However, in the technical solution of this embodiment, the system first identifies the entities "XX Bank" and "interest" in the question, and then replaces them with specific referential symbols, such as... <bank>"and" <interest>The question after the replacement was, "Why did I..." <bank>"My fixed deposit has matured, but my interest hasn't arrived yet?" Next, the system will perform question recall processing in the second corpus, finding standard or similar questions that are similar to the replaced question, such as... <bank>of <interest>Unsettled?". Finally, the system performs multi-dimensional feature similarity matching between these candidate questions and the original user question, selects the most matching target question, and returns its associated answer, thereby effectively solving the user's query.

[0040] It can be understood that the present application can provide technical support for intelligent question answering systems in different scenarios, especially in handling user questions with complex backgrounds and diverse expressions, which can significantly improve the accuracy of system answers and user satisfaction. By constructing and maintaining a corpus containing rich standard questions and similar questions, and applying entity replacement technology and multi-dimensional feature similarity matching algorithm, the present application can achieve accurate understanding and efficient answering of user questions. In addition, the present application also has good scalability and adaptability. With the development of business and changes in user needs, the content in the corpus can be continuously updated and improved to adapt to new question answering scenarios.

[0041] In an embodiment, after step S201 and before step S202, the method further comprises:

[0042] performing question recall processing on the received user question in the first corpus to obtain at least one first candidate question, and performing semantic similarity matching between the user question and all candidate questions to select a corresponding target question and return an associated answer; if no target question is selected, returning information prompting that the user question cannot be recognized; and performing step S202.

[0043] In this embodiment, after the user question is received, it is not searched in the second corpus, nor is the entity replaced in the user question, but the user question is directly matched with the corpus in the first corpus or other knowledge base without replacing entities.

[0044] If no content is matched, the entity is extracted from the user question using an entity extraction model, and the entity extraction result is replaced with a specific reference symbol, and then the replaced user question is searched in the second corpus.

[0045] In an embodiment, as shown in FIG. 3, step S201 includes steps S301-S303.

[0046] S301, confirming the similar questions associated with each standard question in the first corpus;

[0047] S302, using NER technology and regular matching to label entities in each similar question;

[0048] S303, replacing all similar entities of the same type with corresponding specific reference symbols to update the expression form of all similar questions, and obtaining an updated second corpus.

[0049] In this embodiment, the accuracy and efficiency of subsequent question-and-answer matching are improved by standardizing similar questions in the first corpus. Specifically, NER technology can accurately identify key entities in similar questions, while regular expression matching ensures the consistency and accuracy of entity annotation. In addition, uniformly replacing similar entities with corresponding specific referential symbols not only simplifies the expression of similar questions but also helps reduce matching errors caused by entity differences, thereby improving the performance of the entire question-and-answer system.

[0050] It should be noted that the key entities in similar questions depend on the specific application scenario. For example, in a medical Q&A scenario, disease names, drug names, and symptom descriptions can be considered entities. In a financial Q&A scenario, stock names, fund names, types of financial products, and financial indicators can be considered entities.

[0051] In one embodiment, as shown in FIG4, the step of performing question recall processing on the received user question in the first corpus to obtain at least one first candidate question, and performing semantic similarity matching between the user question and all candidate questions to select the corresponding target question and return the associated answer includes steps S401-S404.

[0052] S401. Obtain word segments from the user question and retrieve them from the first corpus based on the word segments to obtain at least one first candidate question;

[0053] S402. Input the user question and all first candidate questions into a pre-trained semantic matching model to calculate semantic similarity, and output the semantic similarity score between the user question and each first candidate question.

[0054] S403. If the highest value among all semantic similarity scores is greater than or equal to the preset similarity threshold, then the first candidate question corresponding to the highest value is taken as the target question and the associated answer is returned.

[0055] S404. If the highest value among all semantic similarity scores is less than the preset similarity threshold, a message indicating that the user's question cannot be identified will be returned.

[0056] In this embodiment, firstly, the system performs word segmentation on the user-input question, breaking the sentence into individual words or phrases for subsequent processing and analysis. After segmentation, the system searches for at least one first candidate question (the number of first candidate questions can be set) in a first corpus (which pre-stores a large number of similar questions and corresponding related answers, serving as the basis for system processing and matching). Next, the system inputs the user-input question and all recalled first candidate questions into a pre-trained semantic matching model. The semantic matching model can capture the semantic connections between the question and various similar questions; even if their expressions differ, as long as the meaning is similar, the model can provide a corresponding semantic similarity score. The purpose of this step is to find at least one first candidate question that is close to the user's question, so as to provide an accurate answer later. Then, the highest value among all semantic similarity scores is checked. If the highest value is greater than or equal to a preset similarity threshold, a candidate question that is sufficiently close to the user's question is considered to have been found and is used as the target question, and the associated answer is returned. However, if the highest value among all semantic similarity scores is less than the preset similarity threshold, a message indicating that the system cannot identify the user's question is returned, informing the user that the system cannot provide a relevant answer. Finally, the process jumps to step S202.

[0057] For example, in a medical question-and-answer scenario, a user asks, "I have a headache and a slight fever, what medicine should I take?" The system first segments the question into words, obtaining words such as "I," "headache," "and," "slight," and "fever." Then, the system searches the first corpus for candidate questions related to these segmented words, such as "What should I do if I have a headache and fever?" and "What medicine will help me recover quickly from a headache and fever?" Next, the system inputs the user's question and these candidate questions into a semantic matching model to calculate the semantic similarity score between them. Suppose that one of the candidate questions, "What medicine is best for a headache and fever?", has the highest semantic similarity score and exceeds a preset similarity threshold. The system will then select this question as the target question and return an associated answer, such as, "We suggest you measure your temperature first. If your fever is severe, please seek medical attention promptly. If it's just a mild fever, you can consider taking some antipyretic and analgesic medications." However, if the semantic similarity scores of all candidate questions are below the preset similarity threshold, the system will return a message indicating that it cannot recognize the user's question, such as, "We're sorry, I cannot understand your question. Please ask again or consult a professional doctor." Then, the system will proceed to step S202 to replace the user's question entity and perform an expanded search from the second corpus.

[0058] In one embodiment, as shown in FIG5, step S202 includes steps S501-S505.

[0059] S501. Replace the entities in the user question using specific symbolic references;

[0060] S502. Based on the entity information in the user question, recall processing is performed in the second corpus to obtain at least one second candidate question;

[0061] S503. Input the user question after entity replacement and all second candidate questions into the pre-trained multimodal matching model to calculate multidimensional feature similarity, and output the comprehensive similarity score between the user question after entity replacement and each second candidate question.

[0062] S504. If the highest value among all comprehensive similarity scores is greater than or equal to the preset similarity threshold, then the second candidate question corresponding to the highest value is taken as the target question and the associated answer is returned.

[0063] S505. If the highest value among all comprehensive similarity scores is less than the preset similarity threshold, the recall work in the second corpus will end and an unrecognizable information prompt will be returned.

[0064] In this embodiment, the system first identifies and understands the entities in the user question, and replaces them with specific referential symbols corresponding to those entities to reduce the complexity of the user question and improve processing efficiency. Then, based on this entity information, the system performs a recall operation in the first corpus to find at least one second candidate question that is close to the user question. Next, the system inputs the user question after entity replacement processing and all recalled second candidate questions into a pre-trained multimodal matching model. This model can perform multi-dimensional in-depth analysis using multi-dimensional feature similarity calculation methods and outputs a comprehensive similarity score between each second candidate question and the user question after entity replacement processing. Then, if the highest value reaches or exceeds a preset similarity threshold, the system considers the second candidate question corresponding to the highest value to be the best match for the user question, thus identifying it as the target question and returning an answer associated with the target question. However, if the highest value of all comprehensive similarity scores is lower than the preset similarity threshold, it is considered that no question sufficiently matching the user question can be found in the second corpus. In this case, the system terminates the recall operation in the second corpus and returns an unrecognizable information prompt to the user.

[0065] For example, in a financial Q&A scenario, a user asks, "What is the amount on my credit card bill?" The system identifies the entities "credit card" and "billing amount" in the user's question. Then, the system replaces "credit card" with specific symbols like "card" and "billing amount" with "¥," resulting in a modified user question like, "What is the ¥ on my 'card'?". Next, the system retrieves information from a second corpus based on the entity "credit card bill amount," potentially finding second candidate questions such as "How can I check the ¥ payment on my 'card'?" and "I want to check the ¥ balance on my 'card'?". Finally, the system inputs the modified user question and these second candidate questions into a pre-trained multimodal matching model to calculate multidimensional feature similarity. Suppose that the second candidate question, "How do I check the repayment amount of 'card' (¥)?", has the highest overall similarity score with the replaced user question, and this highest score is greater than or equal to a preset similarity threshold. Then, the system considers this second candidate question the best match for the user question, identifies it as the target question, and returns an answer associated with the target question, such as "Your credit card repayment amount is XX yuan." However, if the highest overall similarity score of all second candidate questions with the replaced user question is lower than the preset similarity threshold, then the system considers that no question can be found in the second corpus that sufficiently matches the user question, thus ending the recall process and returning a "question not recognized" message to the user.

[0066] In one embodiment, as shown in FIG6, step S503 includes steps S601-S604.

[0067] S601. Perform syntactic decomposition, semantic parsing and entity location marking on the user question after entity replacement processing to obtain the syntactic structure, semantic role and entity location information of the user question.

[0068] In this step, syntactic decomposition refers to breaking down the syntactic structure of the user question, such as subject, verb, and object. Semantic parsing refers to parsing the semantic roles of the user question, such as the action subject, object, and additional conditions. Entity position marking processing refers to marking the sequential position of specific referential symbols in the user question.

[0069] S602. Perform joint feature encoding on the syntactic structure, semantic roles, and entity location information of the user question to obtain a multi-dimensional hybrid vector;

[0070] In this step, the syntactic structure can be converted into a topological vector; the semantic roles can be converted into a relation matrix; the entity location information can be converted into a weight sequence; and then the three feature vectors are concatenated to obtain a multidimensional hybrid vector.

[0071] S603. Perform syntactic similarity calculation, semantic fit calculation, and entity position relevance calculation on the multidimensional mixed vector of the user question and the multidimensional mixed vector of each second candidate question to obtain syntactic similarity score, semantic fit score, and entity position relevance score.

[0072] S604. The syntactic similarity score, semantic fit score, and entity position relevance score are weighted and fused to output the final comprehensive similarity score.

[0073] In steps S603-S604, deep learning algorithms or machine learning models can be used to calculate syntactic similarity, semantic fit, and entity position relevance. These algorithms or models can capture subtle differences and deep connections between texts, thereby improving matching accuracy. Syntactic similarity calculation can be based on parse trees or dependency relations to assess the structural similarity between two sentences. Semantic fit calculation can utilize word embeddings, semantic role labeling, and other techniques to measure the semantic closeness between two sentences. Entity position relevance calculation focuses on the position and relative relationship of entities in the sentence to further improve matching accuracy. By weighted fusion of these scores, a comprehensive similarity score can be obtained to evaluate the overall matching degree between the user's question and the candidate questions.

[0074] In this embodiment, based on steps S601-S604, a refined matching analysis between the user's question and the second candidate question is achieved. This matching considers the similarity of syntactic structures and delves into the semantic level, even down to the processing of entity positions, thereby ensuring the comprehensiveness and accuracy of the matching. In this way, the system can more intelligently understand the user's intent, quickly locate the target question and related answers that best meet the user's needs, and improve user experience and satisfaction.

[0075] As can be seen, the above solution provides a more in-depth analysis of difficult user questions in the question-and-answer business of financial and medical scenarios. When traditional question-and-answer systems fail, the system will analyze the entities in the user's question and compare them with a larger volume of similar questions. In this way, the scope of the system's retrieval is expanded, and it can accurately match the most relevant similar questions and related answers from massive amounts of data, effectively improving the recall rate.

[0076] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0077] In one embodiment, a question-answering device 700 for application entity extraction is provided, which corresponds one-to-one with the question-answering method for application entity extraction in the above embodiments. As shown in FIG7, the question-answering device for application entity extraction includes a corpus update module 701 and a second query module 702. Detailed descriptions of each functional module are as follows:

[0078] The corpus update module 701 is used to identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions with specific referential symbols to obtain the updated second corpus.

[0079] The second query module 702 is used to replace entities in the user question with specific referential symbols, and to perform question recall processing on the user question after entity replacement in the second corpus to obtain at least one second candidate question. The user question is then matched with all second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

[0080] In one embodiment, the question-answering device applying entity extraction further includes:

[0081] The first query module is used to perform question retrieval processing on the received user questions in the first corpus to obtain at least one first candidate question, and to perform semantic similarity matching between the user questions and all candidate questions to select the corresponding target question and return the associated answer; if no target question is selected, the process jumps to the second query module 702.

[0082] In one embodiment, the first query module is specifically used for:

[0083] Obtain word segments from the user's question, and retrieve and process them from the first corpus based on the word segments to obtain at least one first candidate question;

[0084] The user question and all first candidate questions are input into a pre-trained semantic matching model to calculate semantic similarity, and the semantic similarity score between the user question and each first candidate question is output.

[0085] If the highest value among all semantic similarity scores is greater than or equal to the preset similarity threshold, then the first candidate question corresponding to the highest value is taken as the target question and the associated answer is returned;

[0086] If the highest value among all semantic similarity scores is less than the preset similarity threshold, a message indicating that the user's question cannot be identified will be returned.

[0087] In one embodiment, the corpus update module 701 is specifically used for:

[0088] Identify the similar questions associated with each standard question in the first corpus;

[0089] The NER technique and regular expression matching are used to annotate entities in each similarity problem;

[0090] The entity type replaces all similar entities with corresponding specific referential symbols to update the expression of all similar questions and obtain the updated second corpus.

[0091] In one embodiment, the second query module 702 is specifically used for:

[0092] The entities in the user's question are replaced using specific symbolic representations;

[0093] Based on the entity information in the user question, recall processing is performed in the second corpus to obtain at least one second candidate question;

[0094] The user question after entity replacement and all second candidate questions are input into a pre-trained multimodal matching model to calculate multidimensional feature similarity, and the comprehensive similarity score between the user question after entity replacement and each second candidate question is output.

[0095] If the highest value among all comprehensive similarity scores is greater than or equal to the preset similarity threshold, then the second candidate question corresponding to the highest value will be used as the target question and the associated answer will be returned.

[0096] If the highest value among all comprehensive similarity scores is less than the preset similarity threshold, the recall process in the second corpus will end and an unidentified message will be returned.

[0097] In one embodiment, the user question after entity replacement and all second candidate questions are input into a pre-trained multimodal matching model for multidimensional feature similarity calculation, and the comprehensive similarity score between the user question after entity replacement and each second candidate question is output, specifically for:

[0098] The user questions after entity replacement are syntactically decomposed, semantically parsed, and entity location marked to obtain the syntactic structure, semantic roles, and entity location information of the user questions.

[0099] The syntactic structure, semantic roles, and entity location information of the user question are jointly encoded to obtain a multi-dimensional hybrid vector.

[0100] The multidimensional hybrid vector of the user question and the multidimensional hybrid vector of each second candidate question are used to calculate syntactic similarity, semantic fit, and entity position relevance to obtain syntactic similarity score, semantic fit score, and entity position relevance score.

[0101] The syntactic similarity score, semantic fit score, and entity location relevance score are weighted and fused to output the final comprehensive similarity score.

[0102] This invention provides a question-answering device that uses entity extraction to conduct a more in-depth analysis of difficult user questions in financial and medical question-answering scenarios. When traditional question-answering systems fail, the system analyzes the entities in the user's question and compares them with a larger volume of similar questions. In this way, the scope of the system's retrieval is expanded, enabling it to accurately match the most relevant similar questions and related answers from massive amounts of data, effectively improving the recall rate.

[0103] Specific limitations regarding the question-answering device for application entity extraction can be found in the limitations of the question-answering method for application entity extraction described above, and will not be repeated here. Each module in the aforementioned question-answering device for application entity extraction can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0104] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram is shown in Figure 8. The computer device includes a processor, memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a question-and-answer method for application entity extraction on the server side.

[0105] In one embodiment, a computer device is provided, which can be a client, and its internal structure diagram is shown in Figure 9. The computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a question-and-answer method for application entity extraction on the client side.

[0106] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0107] Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus;

[0108] The entities in the user question are replaced using specific referential symbols. The user question with the replaced entities is then used for question recall in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

[0109] In one embodiment, a non-volatile computer-readable storage medium is provided, on which a computer program is stored, the computer program performing the following steps when executed by a processor:

[0110] Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus;

[0111] The entities in the user question are replaced using specific referential symbols. The user question with the replaced entities is then used for question recall in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

[0112] It should be noted that the functions or steps that can be implemented by the non-volatile computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0113] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0114] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0115] The embodiments described above are merely illustrative of the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application. Software tools or components not belonging to this company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.< / interest> < / bank> < / bank> < / interest> < / bank> < / product> < / product> < / product> < / product> < / product>

Claims

1. A question-answering method applying entity extraction, comprising: Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus; The user question is replaced with a specific symbol, and the user question with the replaced entity is retrieved in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

2. The question-answering method based on entity extraction as described in claim 1, wherein, Before the steps of replacing entities in the user question with specific referential symbols, retrieving the entity-replaced user question from the second corpus to obtain at least one second candidate question, and performing multi-dimensional feature similarity matching between the user question and all second candidate questions to select the corresponding target question and return the associated answer, the method further includes: The received user questions are processed in the first corpus to obtain at least one first candidate question. The user questions are then semantically similar to all candidate questions to select the corresponding target question and return the associated answer. If no target question is selected, a message indicating that the user question cannot be identified is returned; and the following steps are performed: replacing the entity in the user question with a specific symbol, retrieving the user question with the replaced entity in the second corpus to obtain at least one second candidate question, and performing multi-dimensional feature similarity matching between the user question and all second candidate questions to select the corresponding target question and return the associated answer.

3. The question-answering method based on entity extraction as described in claim 1, wherein, The process involves identifying all similar questions associated with each standard question in the first corpus, replacing entities in all similar questions using specific referential symbols, and obtaining an updated second corpus, including: Identify the similar questions associated with each standard question in the first corpus; The NER technique and regular expression matching are used to annotate entities in each similarity problem; The entity type replaces all similar entities with corresponding specific referential symbols to update the expression of all similar questions and obtain the updated second corpus.

4. The question-answering method based on entity extraction as described in claim 2, wherein, The step of performing question retrieval processing on the received user questions in the first corpus to obtain at least one first candidate question, and performing semantic similarity matching between the user questions and all candidate questions to select the corresponding target question and return the associated answer, includes: Obtain word segments from the user's question, and retrieve and process them from the first corpus based on the word segments to obtain at least one first candidate question; The user question and all first candidate questions are input into a pre-trained semantic matching model to calculate semantic similarity, and the semantic similarity score between the user question and each first candidate question is output. If the highest value among all semantic similarity scores is greater than or equal to the preset similarity threshold, then the first candidate question corresponding to the highest value is taken as the target question and the associated answer is returned; If the highest value among all semantic similarity scores is less than the preset similarity threshold, a message indicating that the user's question cannot be identified will be returned.

5. The question-answering method based on entity extraction as described in claim 1, wherein, The process of replacing entities in the user question with specific referential symbols, retrieving the entity-replaced user question from the second corpus to obtain at least one second candidate question, and performing multi-dimensional feature similarity matching between the user question and all second candidate questions to select the corresponding target question and return the associated answer includes: The entities in the user's question are replaced using specific symbolic representations; Based on the entity information in the user question, recall processing is performed in the second corpus to obtain at least one second candidate question; The user question after entity replacement and all second candidate questions are input into a pre-trained multimodal matching model to calculate multidimensional feature similarity, and the comprehensive similarity score between the user question after entity replacement and each second candidate question is output. If the highest value among all comprehensive similarity scores is greater than or equal to the preset similarity threshold, then the second candidate question corresponding to the highest value will be used as the target question and the associated answer will be returned. If the highest value among all comprehensive similarity scores is less than the preset similarity threshold, the recall process in the second corpus will end and an unidentified message will be returned.

6. The question-answering method based on entity extraction as described in claim 5, wherein, The process involves inputting the entity-replaced user question and all second candidate questions into a pre-trained multimodal matching model to calculate multidimensional feature similarity, and outputting a comprehensive similarity score between the entity-replaced user question and each second candidate question, including: The user questions after entity replacement are syntactically decomposed, semantically parsed, and entity location marked to obtain the syntactic structure, semantic roles, and entity location information of the user questions. The syntactic structure, semantic roles, and entity location information of the user question are jointly encoded to obtain a multi-dimensional hybrid vector. The multidimensional hybrid vector of the user question and the multidimensional hybrid vector of each second candidate question are used to calculate syntactic similarity, semantic fit, and entity position relevance to obtain syntactic similarity score, semantic fit score, and entity position relevance score. The syntactic similarity score, semantic fit score, and entity location relevance score are weighted and fused to output the final comprehensive similarity score.

7. The question-answering method based on entity extraction as described in claim 4, wherein, The step of obtaining word segments from the user's question and recalling them from the first corpus based on the word segments to obtain at least one first candidate question includes: The user's question sentence is broken down into individual words or phrases to obtain the word segmentation results; Based on the word segmentation results, at least one first candidate question related to it is searched in the first corpus, wherein the first corpus stores similar questions and corresponding related answers in advance.

8. The question-answering method based on entity extraction as described in claim 6, wherein, The process involves syntactic decomposition, semantic parsing, and entity location marking of the user question after entity replacement, resulting in the syntactic structure, semantic roles, and entity location information of the user question, including: The subject-verb-object syntactic structure of the user question is syntactically decomposed, the action subject, object and additional conditions in the user question are semantically parsed, and the sequential position of the specific referential symbols in the user question is marked with entity position marking; thus, the syntactic structure, semantic roles and entity position information of the user question are obtained.

9. A question-answering device applying entity extraction, wherein, include: The corpus update module is used to identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions with specific referential symbols to obtain the updated second corpus; The second query module is used to replace entities in the user question with specific referential symbols, and to perform question recall processing on the user question with the replaced entities in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, When the processor executes the computer program, it performs the following steps: Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus; The user question is replaced with a specific symbol, and the user question with the replaced entity is retrieved in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

11. The computer device according to claim 10, wherein, Before the steps of replacing entities in the user question with specific referential symbols, retrieving the entity-replaced user question from the second corpus to obtain at least one second candidate question, and performing multi-dimensional feature similarity matching between the user question and all second candidate questions to select the corresponding target question and return the associated answer, the method further includes: The received user questions are processed in the first corpus to obtain at least one first candidate question. The user questions are then semantically similar to all candidate questions to select the corresponding target question and return the associated answer. If no target question is selected, a message indicating that the user question cannot be identified is returned; and the following steps are performed: replacing entities in the user question with specific referential symbols, retrieving the user question with the replaced entities from the second corpus to obtain at least one second candidate question, and performing multi-dimensional feature similarity matching between the user question and all second candidate questions to select the corresponding target question and return the associated answer.

12. The computer device according to claim 10, wherein, The process involves identifying all similar questions associated with each standard question in the first corpus, replacing entities in all similar questions using specific referential symbols, and obtaining an updated second corpus, including: Identify the similar questions associated with each standard question in the first corpus; The NER technique and regular expression matching are used to annotate entities in each similarity problem; The entity type replaces all similar entities with corresponding specific referential symbols to update the expression of all similar questions and obtain the updated second corpus.

13. The computer device according to claim 11, wherein, The step of performing question retrieval processing on the received user questions in the first corpus to obtain at least one first candidate question, and performing semantic similarity matching between the user questions and all candidate questions to select the corresponding target question and return the associated answer, includes: Obtain word segments from the user's question, and retrieve and process them from the first corpus based on the word segments to obtain at least one first candidate question; The user question and all first candidate questions are input into a pre-trained semantic matching model to calculate semantic similarity, and the semantic similarity score between the user question and each first candidate question is output. If the highest value among all semantic similarity scores is greater than or equal to the preset similarity threshold, then the first candidate question corresponding to the highest value is taken as the target question and the associated answer is returned; If the highest value among all semantic similarity scores is less than the preset similarity threshold, a message indicating that the user's question cannot be identified will be returned.

14. The computer device according to claim 10, wherein, The process of replacing entities in the user question with specific referential symbols, retrieving the entity-replaced user question from the second corpus to obtain at least one second candidate question, and performing multi-dimensional feature similarity matching between the user question and all second candidate questions to select the corresponding target question and return the associated answer includes: The entities in the user's question are replaced using specific symbolic representations; Based on the entity information in the user question, recall processing is performed in the second corpus to obtain at least one second candidate question; The user question after entity replacement and all second candidate questions are input into a pre-trained multimodal matching model to calculate multidimensional feature similarity, and the comprehensive similarity score between the user question after entity replacement and each second candidate question is output. If the highest value among all comprehensive similarity scores is greater than or equal to the preset similarity threshold, then the second candidate question corresponding to the highest value will be used as the target question and the associated answer will be returned. If the highest value among all comprehensive similarity scores is less than the preset similarity threshold, the recall process in the second corpus will end and an unidentified message will be returned.

15. The computer device according to claim 14, wherein, The process involves inputting the entity-replaced user question and all second candidate questions into a pre-trained multimodal matching model to calculate multidimensional feature similarity, and outputting a comprehensive similarity score between the entity-replaced user question and each second candidate question, including: The user questions after entity replacement are syntactically decomposed, semantically parsed, and entity location marked to obtain the syntactic structure, semantic roles, and entity location information of the user questions. The syntactic structure, semantic roles, and entity location information of the user question are jointly encoded to obtain a multi-dimensional hybrid vector. The multidimensional hybrid vector of the user question and the multidimensional hybrid vector of each second candidate question are used to calculate syntactic similarity, semantic fit, and entity position relevance to obtain syntactic similarity score, semantic fit score, and entity position relevance score. The syntactic similarity score, semantic fit score, and entity location relevance score are weighted and fused to output the final comprehensive similarity score.

16. The computer device according to claim 13, wherein, The step of obtaining word segments from the user's question and recalling them from the first corpus based on the word segments to obtain at least one first candidate question includes: The user's question sentence is broken down into individual words or phrases to obtain the word segmentation results; Based on the word segmentation results, at least one first candidate question related to it is searched in the first corpus, wherein the first corpus stores similar questions and corresponding related answers in advance.

17. The computer device according to claim 15, wherein, The process involves syntactic decomposition, semantic parsing, and entity location marking of the user question after entity replacement, resulting in the syntactic structure, semantic roles, and entity location information of the user question, including: The subject-verb-object syntactic structure of the user question is syntactically decomposed, the action subject, object and additional conditions in the user question are semantically parsed, and the sequential position of the specific referential symbols in the user question is marked with entity position marking; thus, the syntactic structure, semantic roles and entity position information of the user question are obtained.

18. A non-volatile computer-readable storage medium storing a computer program, wherein, When the computer program is executed by the processor, it causes the processor to perform the following operations: Identify all similar questions associated with each standard question in the first corpus, and replace the entities in all similar questions using specific referential symbols to obtain the updated second corpus; The user question is replaced with a specific symbol, and the user question with the replaced entity is retrieved in the second corpus to obtain at least one second candidate question. The user question is then matched with all the second candidate questions using multidimensional feature similarity to select the corresponding target question and return the associated answer.

19. The non-volatile computer-readable storage medium according to claim 18, wherein, Before the steps of replacing entities in the user question with specific referential symbols, retrieving the entity-replaced user question from the second corpus to obtain at least one second candidate question, and performing multi-dimensional feature similarity matching between the user question and all second candidate questions to select the corresponding target question and return the associated answer, the method further includes: The received user questions are processed in the first corpus to obtain at least one first candidate question. The user questions are then semantically similar to all candidate questions to select the corresponding target question and return the associated answer. If no target question is selected, a message indicating that the user question cannot be identified is returned; and the following steps are performed: replacing the entity in the user question with a specific symbol, retrieving the user question with the replaced entity in the second corpus to obtain at least one second candidate question, and performing multi-dimensional feature similarity matching between the user question and all second candidate questions to select the corresponding target question and return the associated answer.

20. The non-volatile computer-readable storage medium according to claim 18, wherein, The process involves identifying all similar questions associated with each standard question in the first corpus, replacing entities in all similar questions using specific referential symbols, and obtaining an updated second corpus, including: Identify the similar questions associated with each standard question in the first corpus; The NER technique and regular expression matching are used to annotate entities in each similarity problem; The entity type replaces all similar entities with corresponding specific referential symbols to update the expression of all similar questions and obtain the updated second corpus.