Digital human intelligent question-answering system and method for water affair scene

Abstract

The invention provides a digital human intelligent question-answering system and method for a water affair scene, and relates to the technical field of intelligent water affair, and the method comprises the steps: an inquiry receiving module obtains a user question through a digital human; the data updating module obtains real-time water affair data and updates the extended database; the question retrieval module is connected with the intelligent question and answer model and the water affair knowledge base, semantic matching retrieval is performed in the water affair knowledge base based on the user question through the intelligent question and answer model, and a matched question is obtained; an answer matching module queries a matching answer in an answer data set of the water affair knowledge base based on the matching question; the real-time data query module queries the associated real-time water affair data in the extended database based on the matching problem, and generates a real-time data query tag; and the interaction feedback module displays the matching answer and the real-time data query tag through a digital person. According to the invention, the technical problem of low water affair knowledge propagation efficiency caused by tedious water affair knowledge acquisition modes in the prior art is solved, and the water affair knowledge propagation efficiency and propagation effect are improved.

Description

Technical Field

[0001] The present application relates to the field of smart water technology, and specifically to a digital human intelligent question-answering system and method for water scenarios. Background Art

[0002] As a key area for safeguarding people's livelihood and promoting economic and social development, the importance of the water industry is self-evident. However, due to the limitations of concepts and technology, the traditional water industry often adopts extensive management concepts, resulting in many deficiencies in corporate culture promotion, business knowledge acquisition, knowledge inheritance and reuse.

[0003] First, water companies have long relied on paper documents, oral teaching, and unstructured information exchange. When users query water-related information, they often need to go through complex processes or channels, resulting in inefficient knowledge acquisition and slow updates. In addition, there is a lack of innovation in corporate culture promotion and product publicity, making it difficult to attract the attention of a new generation of users. Secondly, the ability to inherit and reuse knowledge is insufficient. The accumulation of knowledge in the water industry is often closely related to personal experience. There is a lack of effective mechanisms to transform these experiences into reusable digital assets, resulting in knowledge loss and repeated learning, as well as slow knowledge updates within the water industry, making it difficult to meet the rapidly developing market needs. Summary of the invention

[0004] The present application provides a digital human intelligent question-answering system and method for water affairs scenarios, which solves the technical problem that the traditional way of acquiring water affairs knowledge is cumbersome and leads to low efficiency in the dissemination of water affairs knowledge, and achieves the technical effect of improving the efficiency and effect of water affairs knowledge dissemination, while enhancing the user experience.

[0005] In view of the above problems, on the one hand, the present application provides a digital human intelligent question and answer system for water affairs scenarios, the system comprising an inquiry receiving module, the inquiry receiving module being used to interact with users through digital humans to obtain user questions; a data updating module, the data updating module being used to obtain real-time water affairs data and update an extended database, the extended database being associated with a water affairs knowledge base, the water affairs knowledge base comprising a question data set and an answer data set, the question data set and the answer data set being in a corresponding relationship; a question retrieval module, the question retrieval module connecting an intelligent question and answer model and a water affairs knowledge base, performing semantic matching retrieval in the water affairs knowledge base based on the user questions through the intelligent question and answer model to obtain matching questions, wherein the intelligent question and answer model comprises a pre-screening unit, a preliminary sorting unit and a precise sorting unit, The pre-screening unit performs preliminary matching on the question data set in the water affairs knowledge base based on the user question to obtain a candidate question set, the preliminary sorting unit performs preliminary matching sorting on the candidate question set based on the user question to determine a coarse-ranked question set, and the precise sorting unit performs deep matching screening on the coarse-ranked question set based on the user question to obtain the matching question; an answer matching module, the answer matching module is used to query the answer data set in the water affairs knowledge base based on the matching question for a matching answer; a real-time data query module, the real-time data query module is used to query the associated real-time water affairs data in the extended database based on the matching question, and generate a real-time data query tag for the matching answer; an interactive feedback module, the interactive feedback module is used to display the matching answer and the real-time data query tag through the digital human.

[0006] On the other hand, the present application also provides: a digital human intelligent question-answering method for water affairs scenarios, the method comprising: obtaining user questions through information interaction between digital humans and users; obtaining real-time water affairs data and updating an extended database, wherein the extended database is associated with a water affairs knowledge base, and the water affairs knowledge base includes a question data set and an answer data set, and the question data set and the answer data set have a corresponding relationship; connecting an intelligent question-answering model and a water affairs knowledge base, and performing semantic matching retrieval in the water affairs knowledge base based on the user questions through the intelligent question-answering model to obtain matching questions, wherein the intelligent question-answering model includes a pre-screening unit, a preliminary sorting unit and a precise sorting unit, The pre-screening unit performs preliminary matching on the question data set in the water affairs knowledge base based on the user question to obtain a candidate question set, the preliminary sorting unit performs preliminary matching sorting on the candidate question set based on the user question to determine a rough-ranked question set, and the precise sorting unit performs deep matching screening on the rough-ranked question set based on the user question to obtain the matching question; based on the matching question, a matching answer is queried in the answer data set of the water affairs knowledge base; based on the matching question, the associated real-time water affairs data is queried in the extended database to generate a real-time data query tag for the matching answer; and the matching answer and the real-time data query tag are displayed through the digital human. One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] The inquiry receiving module exchanges information with the user through the digital human to obtain the user's questions, thereby improving the convenience of the user's questions and the system's response speed. The data updating module obtains real-time water affairs data, updates the extended database, and provides data preparation for the subsequent retrieval process. The question retrieval module connects the intelligent question-answering model and the water affairs knowledge base, and performs semantic matching retrieval in the water affairs knowledge base based on the user's questions through the intelligent question-answering model to obtain matching questions, wherein the intelligent question-answering model includes a pre-screening unit, a preliminary sorting unit, and a precise sorting unit. The pre-screening unit performs preliminary matching on the question data set in the water affairs knowledge base based on the user's questions to obtain a candidate question set, quickly filters a large amount of irrelevant information, improves retrieval efficiency, and reduces the computational burden of subsequent processing. The preliminary sorting unit performs preliminary matching sorting on the candidate question set based on the user's questions, determines a rough sorting question set, further improves the accuracy of matching, and reduces the amount of data processed by the subsequent precise sorting unit. The precise sorting unit performs deep matching screening on the rough sorting question set based on the user's questions to obtain the matching questions, thereby ensuring the accuracy of matching questions and improving the reliability of answers. The answer matching module searches for the matching answer in the answer data set of the water affairs knowledge base based on the matching question, ensuring the relevance and accuracy of the answer. The real-time data query module searches for the associated real-time water affairs data in the extended database based on the matching question, and generates a real-time data query tag for the matching answer, so that the user can directly obtain the latest water affairs data. The interactive feedback module displays the matching answer and the real-time data query tag through the digital human, providing a user-friendly interactive experience, so that the user can obtain the answer to the question and real-time water affairs data according to the needs.

[0008] In summary, this application realizes the automation of the whole process from user question acquisition to answer output through the collaborative work of multiple modules. Through multi-level matching and screening, the search scope is gradually narrowed, the retrieval efficiency is improved, and the high-quality output of the answer is ensured. In addition, the system can adapt to the usage habits of different users by interacting with the user through digital people, improve the user experience, enhance the usability and popularity of the system, and efficiently and accurately answer users' questions about water affairs, thereby improving the efficiency and effect of the dissemination of water affairs knowledge.

[0009] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1A schematic diagram of the structure of a digital human intelligent question-answering system for water affairs scenarios provided in an embodiment of the present application;

[0011] Figure 2 A schematic diagram of a flow chart of determining a rough-ranked question set in a digital human intelligent question-answering system for a water affairs scenario provided in an embodiment of the present application;

[0012] Figure 3 A schematic diagram of a process for determining matching problems in a digital human intelligent question-answering system for water affairs scenarios provided in an embodiment of the present application;

[0013] Figure 4 A flowchart of a digital human intelligent question-answering method for water affairs scenarios provided in an embodiment of the present application.

[0014] Explanation of the accompanying drawings: inquiry receiving module 10, data updating module 20, question retrieval module 30, answer matching module 40, real-time data query module 50, interactive feedback module 60. DETAILED DESCRIPTION

[0015] The embodiment of the present application provides a digital human intelligent question-answering system and method for water affairs scenarios, which solves the technical problem that the traditional way of acquiring water affairs knowledge is cumbersome and leads to low efficiency in the dissemination of water affairs knowledge. Through the interaction between the digital human and the user, intelligent and automated question answering is achieved, thereby achieving the technical effect of improving the efficiency and effect of water affairs knowledge dissemination and enhancing the user experience.

[0016] Embodiment 1, as Figure 1 As shown, the embodiment of the present application provides a digital human intelligent question-answering system for water affairs scenarios, and the system includes:

[0017] The inquiry receiving module 10 is used to exchange information with the user through the digital human and obtain the user's questions.

[0018] Specifically, this module collects and understands the questions raised by users through effective information interaction between the constructed digital human and users. In this process, users ask questions to the digital human through voice or text. The digital human converts the user's voice input into text through voice recognition technology, and then uses natural language processing technology, including semantic analysis and intent recognition, to accurately understand user needs, store or mark the questions, and prepare for subsequent retrieval and matching.

[0019] The digital human described in the embodiment of the present application can be constructed based on a 3D high-precision digital human model library. According to specific business needs, based on the 3D high-precision digital human model library, select a suitable digital human model, customize the digital human, including the customization of clothing, hairstyle, and badge, and perform digital human clothing solution, bone binding, expression binding, and action animation. Exemplarily, the digital human can be customized based on the dress requirements of a water company's water service customer service personnel, such as hairstyle, clothing, etc., including the customization of 1 set of clothing, 1 hairstyle, and 1 badge. In order to achieve the animation effect of the digital human model, the skeleton needs to be bound to the model. According to actual needs, select a suitable skeleton system and binding method to ensure that the movements and expressions of the digital human model are more natural and smooth. Then add appropriate animation effects to the digital human model. This includes facial expression animation, body movement animation, etc., and realize the dynamic display of the virtual image through manual animation. For example, drive lip shape and action according to semantics. Finally, use the rendering engine Unity for real-time rendering, including detailed rendering and simulation of light, background, special effects, etc. in the scene to enhance the overall visual effect.

[0020] Through 3D digital human technology, the inquiry receiving module 10 can effectively communicate with the user, accurately obtain and record the user's questions, and at the same time improve the convenience of the user's questions and enhance the user experience.

[0021] The data updating module 20 is used to obtain real-time water affairs data and update the extended database. The extended database is associated with the water affairs knowledge base. The water affairs knowledge base includes a question data set and an answer data set. The question data set and the answer data set have a corresponding relationship.

[0022] Specifically, real-time water affairs data refers to water affairs-related data collected and monitored in real time, including water quality monitoring data, equipment operation data, pipe network data, water affairs policy data, etc. These data usually come from sensors, monitoring equipment or real-time data sources. The extended database is a database used to store real-time water affairs data and is associated with the water affairs knowledge base. The water affairs knowledge base is a database that stores a large number of water affairs-related questions and answers, including a question data set consisting of multiple questions and an answer data set consisting of answers to the questions.

[0023] By setting up a data acquisition interface or connecting sensor devices, the latest data on water affairs can be obtained in real time. For example, a smart water meter can be used to collect hourly water usage data, or a water quality monitoring sensor can be used to obtain water pollution indicators. The database management system is used to write the acquired real-time water affairs data into an extended database for subsequent query and analysis. The data update module 20 updates and displays the latest water affairs data in real time by updating the extended database, while maintaining the relevance to the existing knowledge base, thereby improving the accuracy and practicality of the intelligent question-answering system.

[0024] A question retrieval module 30 is provided, wherein the question retrieval module 30 is connected to an intelligent question-answering model and a water affairs knowledge base, and a semantic matching search is performed in the water affairs knowledge base based on the user question through the intelligent question-answering model to obtain matching questions, wherein the intelligent question-answering model includes a pre-screening unit, a preliminary sorting unit and a precise sorting unit, wherein the pre-screening unit performs a preliminary matching of the question data set in the water affairs knowledge base based on the user question to obtain a candidate question set, the preliminary sorting unit performs a preliminary matching sorting of the candidate question set based on the user question to determine a coarse-sorted question set, and the precise sorting unit performs a deep matching screening of the coarse-sorted question set based on the user question to obtain the matching question.

[0025] Specifically, the intelligent question-answering model is a model trained through deep learning, natural language processing and other technologies. It can understand user questions and find answers that match the questions in the knowledge base. The intelligent question-answering model includes a pre-screening unit, a preliminary sorting unit and a precise sorting unit, which perform semantic matching on user input questions in turn and gradually obtain accurate question matching results. As the basic data source of the intelligent question-answering system, the water affairs knowledge base is a database that stores a large number of water affairs-related questions and answers, including a question data set consisting of multiple questions and an answer data set consisting of answers to the questions.

[0026] The candidate question set is a set of questions that are preliminarily matched with the user's questions and are screened out by the pre-screening unit from the water affairs knowledge base. The intelligent question-answering model receives the acquired user questions, extracts the keywords in the user questions through the pre-screening unit, matches them in the water affairs knowledge base, searches for similar questions stored in the water affairs knowledge base, and obtains the candidate question set. In this way, relevant questions can be quickly screened out from a large amount of water affairs question-answering data, improving the efficiency of question matching and reducing the amount of data for subsequent semantic matching.

[0027] The coarse-ranked question set is a set of candidate questions that are smaller in number but more similar to the user's questions after being screened and sorted by the preliminary sorting unit. The preliminary sorting unit conducts a deeper semantic analysis and sorting of the candidate question set obtained by the preliminary matching to determine a more accurate coarse-ranked question set. This unit uses more complex natural language processing technologies, such as deep learning models and semantic analysis algorithms, to evaluate the semantic similarity between each candidate question and the user's original question, and then screen out questions with higher similarity and sort these questions according to similarity to determine the coarse-ranked question set, providing a high-quality candidate set for the subsequent precise sorting unit.

[0028] The matching questions are the questions stored in the water knowledge base that best match the user's question intent after being processed by the precise sorting unit. The precise sorting unit performs final deep semantic matching and screening on the roughly sorted question set to determine the best matching question. The unit uses advanced semantic analysis algorithms and deep learning models, such as BERT, to deeply understand and analyze each question, and screen out the questions that best match the user's intent from the roughly sorted question set, thereby ensuring the accuracy and reliability of the answers. For example, for the question "How to deal with scale?", the system can accurately find "detailed steps for scale removal" from the candidate set as the final matching question through deep semantic matching, ensuring the accuracy and practicality of the answer.

[0029] The answer matching module 40 is used to query the matching answer in the answer data set of the water affairs knowledge base based on the matching question.

[0030] Specifically, the matching answer refers to the answer corresponding to the matching question in the water affairs knowledge base. The answer matching module 40 obtains the matching question that best meets the user's intention determined by the question retrieval module 30, and then finds the most appropriate answer in the answer data set based on the correspondence between the question and the answer in the water affairs knowledge base to meet the user's needs.

[0031] A real-time data query module 50 is used to query the associated real-time water affairs data in the extended database based on the matching question and generate a real-time data query tag for the matching answer.

[0032] Specifically, the associated real-time water data refers to the latest water data associated with the matching question in the extended database through the data association field. The real-time data query label is used to indicate to the user that the answer contains real-time data and provide a link or button that can be clicked or viewed so that the user can further view the latest data.

[0033] The real-time data query module 50 obtains the identified matching questions, and then queries the extended database for real-time water data related to the matching questions. For example, if the matching question is "What is the current water quality?", the module will search the extended database for the latest water quality monitoring data, such as pH value, turbidity, oxygen content, etc. Through the pre-established data association fields, the module can quickly locate these real-time data. After querying the relevant data, the module will generate a real-time data query tag for the matching answer found. This tag is usually an indicator attached to the answer, such as "View the latest water quality data", so that users can directly view the latest water data.

[0034] The interactive feedback module 60 is used to display the matching answer and the real-time data query tag through the digital human.

[0035] Specifically, the interactive feedback module 60 converts the matching answers from text to voice, and plays them through the constructed virtual image of the digital human, so that the user can receive the answers in the most natural way. The digital human model can be displayed through a variety of terminals such as all-in-one machines, large screens, small screens, PCs and smart phones to facilitate users to obtain the required water affairs knowledge. At the same time, a link or button is displayed on the terminal to display the real-time data query label. The digital human will instruct the user to obtain the latest water quality data by clicking on this label. When the user clicks on this label, the latest water affairs data can be directly viewed. This method not only allows users to obtain instant and accurate information, but also improves the user experience through interaction, ensuring that users can easily obtain all the information they need.

[0036] Furthermore, the embodiment of the present application constructs a water affairs knowledge base, and the execution steps include:

[0037] In combination with business needs, water affairs knowledge question and answer data are collected; the water affairs knowledge question and answer data are organized into a question data set and an answer data set, wherein the question data set includes a standard question set and an extended question set, and both the standard question set and the extended question set are marked with keyword tags; a mapping relationship among the standard question set, the extended question set and the answer data set is established, wherein each standard question data in the standard question set corresponds to one answer data in the answer data set, and each standard question data in the standard question set corresponds to multiple extended question data in the extended question set; the water affairs knowledge base is constructed based on the standard question set, the extended question set, the answer data set and the mapping relationship; the water affairs knowledge question and answer data are interactively acquired and updated, and the water affairs knowledge base is updated.

[0038] Specifically, standard question data refers to questions with standardized wording determined during the data collection process. Each standard question data has a corresponding answer data in the answer data set. At the same time, each standard question data is also associated with multiple extended question data, which refer to question data with different expressions from the standard question data but similar core content. For example, the standard question data may be a direct question like "How to detect contaminants in household water?", while the extended question data may be a variant question like "How to check contaminants in household water?"

[0039] The process of building a water affairs knowledge base begins with the collection of water affairs knowledge question and answer data. These data usually come from professional knowledge in the water affairs field, common user questions and related literature, including real-time water supply operation, the latest water affairs information, standards and specifications, systems, business operation guidelines, management experience, etc. In addition, high-frequency chat knowledge question and answer data can be collected to meet users' question and answer needs for weather, festivals, constellations, news, etc. The collected data is then sorted into a question data set and an answer data set. In this process, the question data set is further divided into a standard question set and an extended question set, both of which are marked with keyword tags for easy classification and retrieval. This process can be completed by natural language processing tools. According to the correspondence between the collected questions and data, a mapping relationship between the standard question set, the extended question set and the answer data set is established. Each standard question corresponds to a specific answer, and at the same time, a standard question may correspond to multiple extended questions, which share the same answer. This mapping relationship ensures that each question can find the corresponding answer. The water affairs knowledge base is constructed by combining the standard question set, the extended question set, the answer data set and the mapping relationship through a database management system.

[0040] After the water affairs knowledge base is built, water affairs knowledge question and answer data is regularly acquired and updated interactively to maintain the timeliness and accuracy of the knowledge base. In addition, the water affairs knowledge base can also be configured with a search interface to facilitate staff to quickly retrieve and query knowledge content directly through keywords.

[0041] Furthermore, the execution steps of the data updating module 20 in the embodiment of the present application include:

[0042] Acquire a water affairs scenario, traverse multiple water affairs data sources based on the water affairs scenario to collect water affairs data, and build an extended database; acquire real-time water affairs data, and extract data type characteristics of the real-time water affairs data, wherein the data type characteristics include water quality monitoring data, equipment operation data, pipe network data, and water affairs policy data; establish associated fields for the real-time water affairs data according to the data type characteristics, and update the association relationship between the real-time water affairs data and the standard questions based on the associated fields and the keyword tags to obtain an updated extended database.

[0043] Specifically, water service scenarios refer to the actual operations and management backgrounds involved in specific water service environments, such as urban water supply systems and sewage treatment plants. Water service data sources are data collection points or systems that can provide water service-related information, such as sensor networks, monitoring systems, historical databases, etc. Real-time water service data is data collected in real time from the water service system, which can reflect the current system status, such as water quality, equipment operation, pipe network pressure, etc. Data type features refer to attributes in real-time water service data that have classification significance and can help classify and process data.

[0044] The data update module 20 first understands and obtains the current water environment, such as urban water supply or sewage treatment, and then collects data related to the scene by traversing multiple water data sources in the scene. For example, in the urban water supply scene, water quality monitoring data can be collected from the sensor network, or the equipment operation status can be obtained from the equipment management system. These data are used to build an extended database.

[0045] Next, the module will continuously obtain real-time water data from these data sources. The type characteristics of real-time water data can include multiple aspects, such as water quality monitoring data (such as pH value, turbidity, etc.), equipment operation data (such as pump station status, energy consumption, etc.), pipeline network data (such as pipeline network pressure, flow, etc.), and data related to water policies issued by the government. These type characteristics help to classify and process data of different categories. Based on these data type characteristics, associated fields are created for each type of real-time water data. These fields are used to associate real-time data with standard questions in the water knowledge base. For example, water quality monitoring data can be associated with questions about "abnormal water quality" in the knowledge base, while equipment operation data may be associated with questions related to "equipment failure".

[0046] Finally, the module updates these associated fields based on the keyword tags, thereby establishing an association relationship between real-time data and standard questions in the extended database. For example, suppose there is a question in the knowledge base, "How to deal with water quality exceeding the standard?" When the water quality monitoring data in the real-time data finds an abnormal pH value, the associated field will match this abnormal data with the question, thereby updating the extended database. This process ensures that when users ask relevant questions, the system can provide accurate answers based on the latest real-time data.

[0047] Furthermore, the embodiment of the present application constructs an intelligent question-answering model, and the execution steps include:

[0048] Acquire a search engine, configure the Okapi BM25 information retrieval algorithm, and acquire the pre-screening unit; construct an initial recurrent neural network, train the initial recurrent neural network with the data in the water affairs knowledge base, and acquire the preliminary sorting unit, wherein the initial recurrent neural network includes a multi-level semantic mapping layer, a multi-level fusion coding layer, and an inference decision layer; acquire a pre-trained BERT language model, migrate the BERT language model based on the water affairs knowledge base, train the TinyBERT model, and acquire the precise sorting unit, wherein the TinyBERT model includes an input layer, a deep semantic encoder, an interactive parsing layer, and an output layer; integrate the pre-screening unit, the preliminary sorting unit, and the precise sorting unit in series to obtain an integrated model; evaluate the integrated model by accuracy, recall rate, and F1 score, and acquire the intelligent question-answering model.

[0049] Specifically, the intelligent question-answering model includes a pre-screening unit, an initial sorting unit, and a precise sorting unit. The construction process includes: using the Elasticsearch search engine and configuring the Okapi BM25 algorithm to obtain a pre-screening unit, which can quickly retrieve possibly related questions based on user questions.

[0050] Build an initial recurrent neural network including a multi-level semantic mapping layer, a multi-level fusion encoding layer, and an inference decision layer, such as a long short-term memory network (LSTM). The multi-level semantic mapping layer is used to convert the input data into a vector representation. The multi-level fusion encoding layer is used to further encode the vector output by the embedding layer to capture deeper information. The inference decision layer is used to make inferences and predictions based on the information in the encoding layer. Obtain data from the water affairs knowledge base to train the initial recurrent neural network to obtain a primary sorting unit, improve the intelligent question-answering model's ability to understand water affairs issues, and screen out more relevant documents.

[0051] The BERT language model is a pre-trained language model based on the Transformer architecture that can understand the complex relationships and contextual meanings of words in a sentence. The TinyBERT model is a miniaturized BERT model designed to handle specific tasks more efficiently. Get a pre-trained large BERT model as a teacher model. This model already contains a lot of language knowledge and patterns and can provide excellent performance on a variety of natural language processing tasks. Design the structure of the TinyBERT model, including the input layer, deep semantic encoder, interactive parsing layer, and output layer. The input layer is used to receive input data. The deep semantic encoder is used to convert the input data into a vector representation. The interactive parsing layer is used to process the relationship between different input data. The output layer is used to generate the final output. Transfer learning is performed on this teacher model based on the water affairs knowledge base to ensure that the TinyBERT model can adapt to specific water affairs scenarios.

[0052] Specifically, it includes: using the teacher model to process the text data in the water affairs knowledge base and generate soft labels. These soft labels contain the predicted probability of each category by the teacher model. Train the TinyBERT model to imitate the output of the teacher model. Make the output of the TinyBERT model, that is, the hard label, close to the soft label of the Teacher model, and make the intermediate layer representation of the TinyBERT model close to the intermediate layer representation of the teacher model. Use the optimized loss function to evaluate the TinyBERT model and adjust its parameters until it achieves a satisfactory inference speed and model size while maintaining a high accuracy, and obtain a TinyBERT model suitable for water affairs scenarios, that is, a precise sorting unit, which can be executed quickly on resource-constrained devices while maintaining a high question-answering accuracy. Starting from the pre-trained BERT model, the TinyBERT model is trained through transfer learning and fine-tuning using the water affairs knowledge base.

[0053] The trained pre-screening unit, preliminary sorting unit and precise sorting unit are connected in series, that is, the output of the previous unit is used as the input of the next unit, to ensure that the entire model can smoothly complete the entire process from preliminary screening to deep matching, and obtain an integrated model. The integrated model is evaluated by accuracy, recall rate and F1 score. The calculation formulas of each indicator are as follows:

[0054]

[0055] Among them, TP represents the number of samples that the model predicts to be similar and are actually similar; FP represents the number of samples that the model predicts to be similar but are actually dissimilar; FN represents the number of samples that the model predicts to be dissimilar but are actually similar. The F1 score combines the precision and recall rates to prevent bias caused by sample imbalance and better evaluate the model effect.

[0056] Through the above steps, the system can obtain a trained intelligent question-answering model, perform multi-level matching in the water affairs knowledge base according to user questions, obtain questions and answers that best meet user needs, and improve users' efficiency and experience in acquiring water affairs knowledge.

[0057] Furthermore, the question retrieval module 30 of the embodiment of the present application is also used to perform the following steps:

[0058] Obtain the user question, perform text analysis on the user question, and obtain a keyword list; based on the keyword list, traverse the question data set to perform keyword matching retrieval to obtain a keyword matching question set; calculate the matching score between the user question and the keyword matching question set through the OkapiBM25 information retrieval algorithm to obtain multiple question matching scores; determine whether the multiple question matching scores are greater than the preset matching scores respectively, and if so, add the question data with a question matching score greater than the preset matching score to the candidate question set.

[0059] Specifically, the keyword list refers to the set of keywords extracted from the user's questions. The keyword matching question set refers to the set of questions in the question dataset whose keyword labels match the keyword list. The matching score is the similarity score between the user's question and the questions in the keyword matching question set calculated based on the algorithm. The preset matching score is a preset score limit used to filter out relevant candidate questions.

[0060] First, the user's question is obtained, and then the natural language processing library, such as NLTK or spaCy, is used to perform text parsing on the user's question to obtain a list of keywords. These keywords are the core vocabulary in the question. Next, the module traverses the question data set based on these keywords and performs keyword matching retrieval to obtain the keyword matching question set. This process is similar to executing a search query containing multiple keywords in a database. Then, the Okapi BM25 information retrieval algorithm is used to calculate the matching score between the user's question and the keyword matching question set. This algorithm considers the frequency of keywords in the question, the length of the document, and the frequency of documents in the entire data set to generate a matching score for each question. A preset matching score is pre-set according to the retrieval accuracy requirements, and the question data with a question matching score greater than this preset value is filtered out and added to the candidate question set.

[0061] Through the above steps, the question retrieval module 30 can efficiently filter out the candidate questions most relevant to the user's questions from the question data set, laying the foundation for subsequent accurate matching and answer retrieval.

[0062] Further, such as Figure 2 As shown, the question retrieval module 30 in the embodiment of the present application is also used to perform the following steps:

[0063] The user question and the candidate question set are respectively mapped into character-level text feature vectors through the multi-level semantic mapping layer to generate a user question character vector and a candidate question character vector; the user question and the candidate question set are respectively mapped into word-level text feature vectors through the multi-level semantic mapping layer to generate a user question word vector and a candidate question word vector; the character vector and the word vector corresponding to the same question are semantically fused through the multi-level fusion coding layer to generate a user question feature vector and multiple candidate question feature vectors; the user question feature vector is respectively paired with multiple candidate question feature vectors to obtain multiple feature vector pairs; the semantic similarities of the multiple feature vector pairs are respectively calculated through the inference decision layer to obtain multiple first semantic similarities; the candidate question set is sorted based on the multiple first semantic similarities to obtain a coarse-sorted question sequence; the first K questions in the coarse-sorted question sequence whose first semantic similarities are greater than a similarity threshold are screened to generate a coarse-sorted question set, wherein K is a positive integer representing the question capacity of the coarse-sorted question sequence; when the number of questions in the coarse-sorted question sequence is less than K, the coarse-sorted question sequence is determined as the coarse-sorted question set.

[0064] Specifically, the first semantic similarity is a quantitative indicator that measures the semantic similarity between the user's question and the questions in the candidate question set. The coarse sorting question set is the most relevant question set that is screened out after being processed by the coarse sorting module. The similarity threshold is a preset limit used to judge the relevance of questions.

[0065] First, a single character or word from the user question and candidate question set is extracted through a multi-level semantic mapping layer, and mapped to a text feature representation vector from the two granularities of character and word, to obtain the user question character vector, candidate question character vector, user question vocabulary vector and candidate question vocabulary vector, and use these vectors as the input of the multi-level fusion encoding layer. Among them, the character vector refers to the vector obtained by mapping the question text sequence to a character-level text feature vector; the vocabulary vector refers to the vector obtained by mapping the question text sequence to a vocabulary-level text feature vector.

[0066] The multi-level fusion coding layer extracts the character vector and vocabulary vector corresponding to the same question, and fuses the semantic features of these two types of vectors to obtain the user question feature vector and multiple candidate question feature vectors. These feature vectors are then input into the reasoning decision layer for semantic similarity calculation. Among them, the user question feature vector and the candidate question feature vector refer to the comprehensive feature vectors formed by combining the character vector and vocabulary vector of the same question through the coding layer.

[0067] The reasoning decision layer pairs the user question feature vector with multiple candidate question feature vectors to obtain multiple feature vector pairs, each of which contains a user question feature vector and a candidate question feature vector. The reasoning decision layer calculates the semantic similarity between the two texts by comparing the semantic feature vectors of the two texts. The similarity calculation can be based on the distance metric of the vectors, such as cosine similarity. Finally, multiple first semantic similarities are output. These similarity values ​​are between 0 and 1. The larger the value, the higher the match. 1 means exactly the same, and 0 means completely different. Based on these first semantic similarities, the module sorts the candidate question set to obtain a coarse-sequenced question sequence.

[0068] Finally, the module selects the first K questions in the rough-sorted question sequence whose first semantic similarity is greater than the similarity threshold to obtain a rough-sorted question set. Wherein, K is a positive integer, representing the problem capacity of the rough-sorted question sequence. Exemplarily, K=10, the module compares the first semantic similarity corresponding to each question in the rough-sorted question sequence with the similarity threshold, and removes those questions with low similarity. For the remaining question data in the rough-sorted question sequence, the top 10 question data are selected and added to the rough-sorted question set. When the number of remaining questions is less than 10, all the remaining questions are retained as a rough-sorted question set. Through this process, candidate questions with higher similarity can be further determined, while reducing the amount of data input to the precise sorting unit, ensuring the accuracy of model calculations while improving the calculation speed.

[0069] Further, such as Figure 3 As shown, the question retrieval module 30 in the embodiment of the present application is also used to perform the following steps:

[0070] The user question and the rough-ranked question set are received through the input layer, and the character vectors and vocabulary vectors corresponding to the user question and the multiple candidate questions in the rough-ranked question set are queried to obtain the user question vector set and the multiple candidate question vector sets; the user question vector set and the multiple candidate question vector sets are respectively encoded through the deep semantic encoder to generate a user question sentence vector and a multiple candidate question sentence vectors; the semantic similarities of the user question sentence vector and the multiple candidate question sentence vectors are respectively calculated through the interactive parsing layer to obtain multiple second semantic similarities; the matching probability of the multiple candidate questions and the user question is output through the output layer based on the multiple second semantic similarities, and the rough-ranked question set is screened and sorted according to the matching probability to determine the matching question.

[0071] Specifically, the user question vector set and the candidate question vector set refer to the vector sets composed of the character vectors and vocabulary vectors corresponding to the user questions and candidate questions, respectively. The input layer of the precise sorting unit queries the character vectors and vocabulary vectors corresponding to multiple candidate questions in the user questions and coarse sorting question sets, summarizes the user question vector set and multiple candidate question vector sets, and inputs these vector sets into the deep semantic encoder.

[0072] The deep semantic encoder encodes all vectors in each vector set through a Transformer or recurrent neural network model to generate a user question sentence vector and multiple candidate question sentence vectors, which are then input into the interactive parsing layer. The interactive parsing layer uses an attention mechanism to measure the relevance between user questions and candidate questions, and obtains multiple second semantic similarities. This similarity measures the semantic proximity of two sentence vectors. The attention mechanism is a method that can dynamically focus on different parts of the input sequence, which can capture the complex relationship between two vectors.

[0073] Finally, the output layer outputs the matching probability of each candidate question and the user question based on these second semantic similarities. The matching probability reflects the matching degree between each candidate question and the user question. This is a binary classification task, usually calculated by a classification model, such as logistic regression or softmax function, outputting 1 or 0. 1 indicates a match and 0 indicates a mismatch.

[0074] Through the above steps, the question retrieval module 30 can effectively filter out the most relevant questions from the roughly sorted question set, thereby ensuring that the answer matching module 40 can provide the most accurate answers to users.

[0075] In summary, the digital human intelligent question-answering system for water affairs provided in the embodiment of the present application has the following technical effects:

[0076] The inquiry receiving module 10 exchanges information with the user through the digital human to obtain the user's questions, thereby improving the convenience of the user's questions and the response speed of the system. The data updating module 20 obtains real-time water affairs data, updates the extended database, and provides data preparation for the subsequent retrieval process. The question retrieval module 30 connects the intelligent question-answering model and the water affairs knowledge base, and performs semantic matching retrieval in the water affairs knowledge base based on the user's questions through the intelligent question-answering model to obtain matching questions, wherein the intelligent question-answering model includes a pre-screening unit, a preliminary sorting unit and a precise sorting unit, and the pre-screening unit performs preliminary matching on the question data set in the water affairs knowledge base based on the user's questions to obtain a candidate question set, quickly filters a large amount of irrelevant information, improves retrieval efficiency, and reduces the computational burden of subsequent processing. The preliminary sorting unit performs preliminary matching sorting on the candidate question set based on the user's questions, determines a rough sorting question set, further improves the accuracy of matching, and reduces the amount of data processed by the subsequent precise sorting unit. The precise sorting unit performs deep matching screening on the rough sorting question set based on the user's questions to obtain the matching questions, thereby ensuring the accuracy of matching questions and improving the reliability of answers. The answer matching module 40 searches for the matching answer in the answer data set of the water affairs knowledge base based on the matching question, ensuring the relevance and accuracy of the answer. The real-time data query module 50 searches for the associated real-time water affairs data in the extended database based on the matching question, and generates a real-time data query tag for the matching answer, so that the user can directly obtain the latest water affairs data. The interactive feedback module 60 displays the matching answer and the real-time data query tag through the digital human, providing a user-friendly interactive experience, so that the user can obtain the answer to the question and real-time water affairs data according to the needs.

[0077] Overall, the embodiment of the present application realizes the automation of the whole process from user question acquisition to answer output through the collaborative work of multiple modules. Through multi-level matching and screening, the search scope is gradually narrowed, the retrieval efficiency is improved, and the high-quality output of the answer is ensured. In addition, the system also interacts with users through digital humans, which can adapt to the usage habits of different users, improve the user experience, enhance the usability and popularity of the system, and efficiently and accurately answer users' questions about water affairs, thereby improving the efficiency and effect of the dissemination of water affairs knowledge.

[0078] Embodiment 2, as Figure 4 As shown, the embodiment of the present application provides a digital human intelligent question-answering method for water affairs scenarios, the method comprising:

[0079] Through digital humans, information is exchanged with users to obtain user questions.

[0080] Real-time water affairs data is acquired, and an extended database is updated. The extended database is associated with a water affairs knowledge base. The water affairs knowledge base includes a question data set and an answer data set. The question data set and the answer data set have a corresponding relationship.

[0081] Connect the intelligent question-answering model and the water affairs knowledge base, and perform semantic matching retrieval in the water affairs knowledge base based on the user questions through the intelligent question-answering model to obtain matching questions, wherein the intelligent question-answering model includes a pre-screening unit, a preliminary sorting unit and a precise sorting unit, the pre-screening unit performs preliminary matching on the question data set in the water affairs knowledge base based on the user questions to obtain a candidate question set, the preliminary sorting unit performs preliminary matching sorting on the candidate question set based on the user questions to determine a coarse-ranked question set, and the precise sorting unit performs deep matching screening on the coarse-ranked question set based on the user questions to obtain the matching questions.

[0082] A matching answer is queried in the answer data set of the water affairs knowledge base based on the matching question.

[0083] The extended database is queried for associated real-time water affairs data based on the matching question, and a real-time data query tag is generated for the matching answer.

[0084] The matching answer and the real-time data query tag are displayed through the digital human.

[0085] Furthermore, the embodiment of the present application constructs a water affairs knowledge base, including:

[0086] In combination with business needs, water affairs knowledge question and answer data are collected; the water affairs knowledge question and answer data are organized into a question data set and an answer data set, wherein the question data set includes a standard question set and an extended question set, and both the standard question set and the extended question set are marked with keyword tags; a mapping relationship among the standard question set, the extended question set and the answer data set is established, wherein each standard question data in the standard question set corresponds to one answer data in the answer data set, and each standard question data in the standard question set corresponds to multiple extended question data in the extended question set; the water affairs knowledge base is constructed based on the standard question set, the extended question set, the answer data set and the mapping relationship; the water affairs knowledge question and answer data are interactively acquired and updated, and the water affairs knowledge base is updated.

[0087] Furthermore, the embodiment of the present application obtains real-time water service data and updates the extended database, including:

[0088] Acquire a water affairs scenario, traverse multiple water affairs data sources based on the water affairs scenario to collect water affairs data, and build an extended database; acquire real-time water affairs data, and extract data type characteristics of the real-time water affairs data, wherein the data type characteristics include water quality monitoring data, equipment operation data, pipe network data, and water affairs policy data; establish associated fields for the real-time water affairs data according to the data type characteristics, and update the association relationship between the real-time water affairs data and the standard questions based on the associated fields and the keyword tags to obtain an updated extended database.

[0089] Preferably, the embodiment of the present application constructs an intelligent question-answering model, including:

[0090] Acquire a search engine, configure the Okapi BM25 information retrieval algorithm, and acquire the pre-screening unit; construct an initial recurrent neural network, train the initial recurrent neural network with the data in the water affairs knowledge base, and acquire the preliminary sorting unit, wherein the initial recurrent neural network includes a multi-level semantic mapping layer, a multi-level fusion coding layer, and an inference decision layer; acquire a pre-trained BERT language model, migrate the BERT language model based on the water affairs knowledge base, train the TinyBERT model, and acquire the precise sorting unit, wherein the TinyBERT model includes an input layer, a deep semantic encoder, an interactive parsing layer, and an output layer; integrate the pre-screening unit, the preliminary sorting unit, and the precise sorting unit in series to obtain an integrated model; evaluate the integrated model by accuracy, recall rate, and F1 score, and acquire the intelligent question-answering model.

[0091] Furthermore, in the embodiment of the present application, the pre-screening unit preliminarily matches the question data set in the water affairs knowledge base based on the user question to obtain a candidate question set, including:

[0092] Obtain the user question, perform text analysis on the user question, and obtain a keyword list; based on the keyword list, traverse the question data set to perform keyword matching retrieval to obtain a keyword matching question set; calculate the matching score between the user question and the keyword matching question set through the OkapiBM25 information retrieval algorithm to obtain multiple question matching scores; determine whether the multiple question matching scores are greater than the preset matching scores respectively, and if so, add the question data with a question matching score greater than the preset matching score to the candidate question set.

[0093] Preferably, the preliminary ranking unit in the embodiment of the present application performs preliminary matching ranking on the candidate question set based on the user question to determine a rough ranking question set, including:

[0094] The user question and the candidate question set are respectively mapped into character-level text feature vectors through the multi-level semantic mapping layer to generate a user question character vector and a candidate question character vector; the user question and the candidate question set are respectively mapped into word-level text feature vectors through the multi-level semantic mapping layer to generate a user question word vector and a candidate question word vector; the character vector and the word vector corresponding to the same question are semantically fused through the multi-level fusion coding layer to generate a user question feature vector and multiple candidate question feature vectors; the user question feature vector is respectively paired with multiple candidate question feature vectors to obtain multiple feature vector pairs; the semantic similarities of the multiple feature vector pairs are respectively calculated through the inference decision layer to obtain multiple first semantic similarities; the candidate question set is sorted based on the multiple first semantic similarities to obtain a coarse-sorted question sequence; the first K questions in the coarse-sorted question sequence whose first semantic similarities are greater than a similarity threshold are screened to generate a coarse-sorted question set, wherein K is a positive integer representing the question capacity of the coarse-sorted question sequence; when the number of questions in the coarse-sorted question sequence is less than K, the coarse-sorted question sequence is determined as the coarse-sorted question set.

[0095] Furthermore, the precise sorting unit in the embodiment of the present application performs deep matching screening on the rough sorted question set based on the user question to obtain the matching question, including:

[0096] The user question and the rough-ranked question set are received through the input layer, and the character vectors and vocabulary vectors corresponding to the user question and the multiple candidate questions in the rough-ranked question set are queried to obtain the user question vector set and the multiple candidate question vector sets; the user question vector set and the multiple candidate question vector sets are respectively encoded through the deep semantic encoder to generate a user question sentence vector and a multiple candidate question sentence vectors; the semantic similarities of the user question sentence vector and the multiple candidate question sentence vectors are respectively calculated through the interactive parsing layer to obtain multiple second semantic similarities; the matching probability of the multiple candidate questions and the user question is output through the output layer based on the multiple second semantic similarities, and the rough-ranked question set is screened and sorted according to the matching probability to determine the matching question.

[0097] Through the above detailed description of a digital human intelligent question and answer system for water conservancy scenarios, those skilled in the art can clearly understand a digital human intelligent question and answer method for water conservancy scenarios in this embodiment. For the method disclosed in Embodiment 2, since it corresponds to the system disclosed in Embodiment 1 and has corresponding steps and beneficial effects, please refer to the system part description for relevant matters.

[0098] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A digital human intelligent question-answering system for water affairs scenarios, characterized in that: The system comprises: A query receiving module, which is used to interact with the user through the digital human to obtain the user's questions; A data updating module, the data updating module is used to obtain real-time water affairs data and update an extended database, the extended database is associated with a water affairs knowledge base, the water affairs knowledge base includes a question data set and an answer data set, the question data set and the answer data set are in a corresponding relationship; A question retrieval module, wherein the question retrieval module is connected to the intelligent question-answering model and the water affairs knowledge base, and performs semantic matching retrieval in the water affairs knowledge base based on the user question through the intelligent question-answering model to obtain matching questions, wherein the intelligent question-answering model includes a pre-screening unit, a preliminary sorting unit and a precise sorting unit, wherein the pre-screening unit performs preliminary matching on the question data set in the water affairs knowledge base based on the user question to obtain a candidate question set, the preliminary sorting unit performs preliminary matching sorting on the candidate question set based on the user question to determine a rough-ranked question set, and the precise sorting unit performs deep matching screening on the rough-ranked question set based on the user question to obtain the matching question; An answer matching module, the answer matching module is used to query a matching answer in an answer data set of the water affairs knowledge base based on the matching question; A real-time data query module, the real-time data query module is used to query the extended database for associated real-time water affairs data based on the matching question, and generate a real-time data query tag for the matching answer; An interactive feedback module, wherein the interactive feedback module is used to display the matching answer and the real-time data query tag through the digital human.

2. The digital human intelligent question-answering system for water affairs scenarios according to claim 1, characterized in that: To build a water affairs knowledge base, the execution steps include: Collect water affairs knowledge question and answer data based on business needs; The water affairs knowledge question and answer data is sorted into a question data set and an answer data set, wherein the question data set includes a standard question set and an extended question set, and both the standard question set and the extended question set are marked with keyword tags; Establishing a mapping relationship among a standard question set, an extended question set, and an answer data set, wherein each standard question data in the standard question set corresponds to one answer data in the answer data set, and each standard question data in the standard question set corresponds to multiple extended question data in the extended question set; Constructing the water affairs knowledge base based on the standard question set, the extended question set, the answer data set and the mapping relationship; Interactively obtain and update water affairs knowledge question and answer data, and update the water affairs knowledge base.

3. The digital human intelligent question-answering system for water affairs scenarios according to claim 2, characterized in that: The execution steps of the data updating module include: Acquire a water affairs scenario, traverse multiple water affairs data sources based on the water affairs scenario to collect water affairs data, and build an extended database; Acquire real-time water affairs data, and extract data type characteristics of the real-time water affairs data, wherein the data type characteristics include water quality monitoring data, equipment operation data, pipe network data, and water affairs policy data; Establish an associated field for the real-time water affairs data according to the data type characteristics, update the association relationship between the real-time water affairs data and the standard question based on the associated field and the keyword tag, and obtain an updated extended database.

4. The digital human intelligent question-answering system for water affairs scenarios according to claim 2, characterized in that: To build an intelligent question-answering model, the execution steps include: Obtain a search engine, configure an Okapi BM25 information retrieval algorithm, and obtain the pre-screening unit; Building an initial recurrent neural network, training the initial recurrent neural network with the data in the water affairs knowledge base, and obtaining the primary sorting unit, wherein the initial recurrent neural network includes a multi-level semantic mapping layer, a multi-level fusion coding layer, and an inference decision layer; Obtain a pre-trained BERT language model, migrate the BERT language model based on the water affairs knowledge base, train a TinyBERT model, and obtain the precise sorting unit, wherein the TinyBERT model includes an input layer, a deep semantic encoder, an interactive parsing layer, and an output layer; The pre-screening unit, the preliminary sorting unit, and the precise sorting unit are integrated in series to obtain an integrated model; The integrated model is evaluated by accuracy, recall and F1 score to obtain the intelligent question answering model.

5. The digital human intelligent question-answering system for water affairs scenarios according to claim 4, characterized in that: The question retrieval module performs the following steps: Obtain the user question, perform text analysis on the user question, and obtain a keyword list; Based on the keyword list, traverse the question data set to perform keyword matching retrieval to obtain a keyword matching question set; The Okapi BM25 information retrieval algorithm is used to calculate the matching scores between the user question and the keyword matching question set, and multiple question matching scores are obtained; It is determined whether the matching scores of the multiple questions are greater than the preset matching scores respectively. If so, the question data with the matching scores greater than the preset matching scores are added to the candidate question set.

6. The digital human intelligent question-answering system for water affairs scenarios according to claim 5, characterized in that: The question retrieval module performs the following steps: Mapping the user question and the candidate question set into character-level text feature vectors respectively through the multi-level semantic mapping layer to generate a user question character vector and a candidate question character vector; Mapping the user question and the candidate question set into vocabulary-level text feature vector mappings respectively through the multi-level semantic mapping layer to generate a user question vocabulary vector and a candidate question vocabulary vector; Through the multi-level fusion coding layer, the semantic features of the character vector and vocabulary vector corresponding to the same question are fused to generate a user question feature vector and multiple candidate question feature vectors; Pairing the user question feature vector with the plurality of candidate question feature vectors respectively to obtain a plurality of feature vector pairs; Calculating the semantic similarities of the plurality of feature vector pairs respectively through the inference decision layer to obtain a plurality of first semantic similarities; Sorting the candidate question set based on the multiple first semantic similarities to obtain a roughly sorted question sequence; Filter the first K questions in the coarse-sorted question sequence whose first semantic similarity is greater than a similarity threshold to generate a coarse-sorted question set, where K is a positive integer representing the question capacity of the coarse-sorted question sequence; When the number of questions in the coarsely sorted question sequence is less than K, the coarsely sorted question sequence is determined as the coarsely sorted question set.

7. The digital human intelligent question-answering system for water affairs scenarios according to claim 6, characterized in that: The question retrieval module performs the following steps: Receiving user questions and a rough-ranked question set through an input layer, querying character vectors and vocabulary vectors corresponding to the user questions and a plurality of candidate questions in the rough-ranked question set, and obtaining a user question vector set and a plurality of candidate question vector sets; Encode the user question vector set and the plurality of candidate question vector sets respectively through a deep semantic encoder to generate a user question sentence vector and a plurality of candidate question sentence vectors; Calculating the semantic similarities of the user question sentence vector and the plurality of candidate question sentence vectors respectively through the interactive parsing layer to obtain a plurality of second semantic similarities; The output layer outputs the matching probabilities between the multiple candidate questions and the user questions based on the multiple second semantic similarities, and the coarse-sorted question set is screened and sorted according to the matching probabilities to determine the matching questions.

8. A digital human intelligent question-answering method for water affairs scenarios, characterized in that: The method is performed by a digital human intelligent question-answering system for water affairs scenarios according to any one of claims 1 to 7, and the method comprises: Interact with users through digital humans to obtain user questions; Acquire real-time water affairs data and update an extended database, wherein the extended database is associated with a water affairs knowledge base, wherein the water affairs knowledge base includes a question data set and an answer data set, and the question data set and the answer data set are in a corresponding relationship; Connecting an intelligent question-answering model and a water affairs knowledge base, performing semantic matching retrieval in the water affairs knowledge base based on the user questions through the intelligent question-answering model to obtain matching questions, wherein the intelligent question-answering model includes a pre-screening unit, a preliminary sorting unit and a precise sorting unit, the pre-screening unit performs preliminary matching on the question data set in the water affairs knowledge base based on the user questions to obtain a candidate question set, the preliminary sorting unit performs preliminary matching sorting on the candidate question set based on the user questions to determine a rough-ranked question set, and the precise sorting unit performs deep matching screening on the rough-ranked question set based on the user questions to obtain the matching questions; Based on the matching question, searching for a matching answer in the answer data set of the water affairs knowledge base; Based on the matching question, the extended database is searched for associated real-time water affairs data, and a real-time data query tag is generated for the matching answer; The matching answer and the real-time data query tag are displayed through the digital human.

Images