Voice real-time question and answer processing method based on large model and domain knowledge base

Abstract

The invention discloses a voice real-time question and answer processing method based on a large model and a domain knowledge base, which belongs to the technical field of computer data processing, and comprises the following steps: acquiring a voice input stream of a user, carrying out intelligent sound wave deconstruction on the voice input stream, generating a text stream, and carrying out dependency syntactic analysis and semantic role labeling; generating intention information and key information, dynamically accessing a domain knowledge base, executing predictive loading, generating a preloaded data subset, performing fusion processing by combining the intention information, the key information and the preloaded data subset, generating a fusion result, performing accuracy verification and correction, generating a natural language answer, and converting the natural language answer into voice output. The technical scheme of combining knowledge predictive loading driven by intention analysis, context fusion and answer traceability correction is adopted, and low-delay response, high-precision intention understanding and high-factuality answer of voice questions and answers can be achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer data processing, and in particular to a voice real-time question and answer processing method based on a large model and a domain knowledge base. BACKGROUND

[0002] As an important form of human-computer interaction, the core of the voice real-time question and answer system is to use computers for data processing to understand the user's natural language instructions and provide accurate information feedback. This kind of system usually integrates speech recognition, natural language processing, information retrieval and large language models, etc. It aims to provide users with a convenient and efficient way of knowledge acquisition and service execution through simulating human dialogue, and has a wide range of applications in intelligent customer service, personal assistants, vehicle-mounted systems, etc.

[0003] In the prior art, the implementation of voice question and answer processing usually adopts a serialized processing flow. First, the user's voice input is converted into text by a speech recognition module. Then, the natural language understanding module analyzes the text and extracts keywords or simple intentions. Next, according to the extracted information, a query request is sent to the backend knowledge base or search engine. After obtaining the relevant data, these data are input as context information to the large language model, and the model organizes the language and generates the final natural language answer. Finally, the text answer is converted into speech by a speech synthesis module and output to the user.

[0004] However, the above-mentioned prior art solution has several technical defects in actual application. First, the entire processing flow is strictly serial, and the next step must wait for the previous step to complete. When dealing with complex problems, multiple database queries and network requests will accumulate into a long response delay, affecting the real-time and smoothness of user interaction. Secondly, the large language model does not fully utilize the input knowledge and lacks effective supervision when generating answers, which can easily lead to deviation from the facts and fabrication of information, resulting in the accuracy of the final output answer content being unable to be guaranteed. In addition, the intention understanding method based on keywords or simple templates is difficult to cope with users' colloquial, diverse and novel expressions, resulting in limited understanding ability and weak service robustness of the system. SUMMARY

[0005] To solve the above problems, the present application provides a voice real-time question and answer processing method based on a large model and a domain knowledge base, which adopts a technical solution combining intention analysis driven knowledge predictive loading, context fusion and answer traceability correction, and can realize low-latency response, high-precision intention understanding and high-fact answer of voice question and answer.

[0006] The above-mentioned object can be achieved by the following solution: A voice real-time question and answer processing method based on a large model and a domain knowledge base, comprising: acquiring a voice input stream of a user, and performing acoustic intelligent deconstruction on the voice input stream to generate a text stream; performing dependency syntax analysis and semantic role labeling on the text stream, analyzing the core predicate and argument structure, generating intent information and key information; based on the intent information and the key information, dynamically accessing the domain knowledge base to locate an initial query node, and performing intent-driven predictive loading to obtain a cluster of adjacent knowledge nodes and analyze the priority of knowledge gaps, and generate a preloading data subset; constructing a context graph structure, fusing and processing the intent information, the key information and the preloading data subset through a context perception algorithm, and performing weighted pruning processing based on the context graph structure to generate a fusion result; based on the fusion result, constructing an answer draft and trace information, combining the preloading data subset to evaluate, obtaining defect content and performing constraint coding iteration correction to generate a natural language answer, and converting the natural language answer into a voice output.

[0007] Optionally, the generating a text stream comprises: collecting a continuous voice input stream, and performing sound source separation on the voice input stream to extract target speaker voice features; performing parallel fast decoding and fine decoding on the target speaker voice features, the fast decoding generating a preliminary text hypothesis with lower time delay, and the fine decoding performing deep recognition under the guidance of the preliminary text hypothesis; performing stream confidence evaluation on the output of the fine decoding to identify and correct low-confidence recognition fragments to obtain a corrected recognition result; at a pause point or sentence semantic termination point of the user voice, integrating and confirming the corrected recognition result to generate a text stream.

[0008] Optionally, the generating intent information and key information comprises: performing dependency syntax analysis and semantic role labeling on the text stream to deconstruct the core predicate and argument structure of the sentence; mapping the core predicate and argument structure to a potential interactive action space defined by the ontology structure of the domain knowledge base to infer the deep intention of user interaction; identifying entities, attributes, constraint conditions and relationship triples related to the deep intention in the text stream to construct a structured query fragment; associating and integrating the deep intention and the structured query fragment to generate intent information and key information.

[0009] Optionally, the method further comprises: when the core predicate and argument structure cannot be directly mapped to the potential interactive action space, decomposing the predicate in the core predicate and argument structure into a basic semantic primitive; in the potential interactive action space, searching for an operation chain composed of the basic semantic primitive that can be logically combined to realize the core predicate and argument structure; dynamically instantiating the operation chain as a transactional execution plan to generate an adaptive analysis result for unseen intentions.

[0010] Optionally, the generating the preloading data subset comprises: locating an initial query node in a domain knowledge base based on the intent information and the key information; performing multi-hop path traversal and semantic expansion from the initial query node to predict a cluster of adjacent knowledge nodes that the user is likely to focus on; obtaining user historical interaction preferences and a current dialogue state, and combining the relevance to the cluster of adjacent knowledge nodes to calculate a dynamic loading priority of each node; and extracting data entities and associated attributes related to high-priority nodes from the domain knowledge base asynchronously and differentially according to the dynamic loading priority to generate the preloading data subset.

[0011] Optionally, the calculating the dynamic loading priority of each node comprises: constructing a vectorized representation of a user cognitive state based on the user historical interaction preferences and the current dialogue state; obtaining a current user intent and inferring a set of knowledge components from an ontology level of the domain knowledge base; comparing the vectorized representation and the set of knowledge components to calculate a knowledge gap vector; and assigning a loading priority to a node in the cluster of adjacent knowledge nodes that meets a preset condition for filling a cognitive blind spot according to the knowledge gap vector.

[0012] Optionally, the generating the fusion result comprises: constructing a dynamic context graph structure, wherein the intent information and the key information are used as a query root node, and the data entities in the preloading data subset are used as knowledge leaf nodes; calculating semantic correlation strengths between the query root node and each of the knowledge leaf nodes, and dynamically weighting edges in the context graph structure according to the semantic correlation strengths to obtain weights of the edges; pruning the context graph structure according to the weights of the edges to form a focused reasoning path; and serializing the focused reasoning path to generate the fusion result.

[0013] Optionally, the generating the natural language answer and converting the natural language answer into voice output comprises: driving a language generation process based on the fusion result to produce a preliminary natural language answer draft and synchronously generate provenance information of key arguments in the answer draft; cross-referencing the provenance information with the preloading data subset to evaluate the answer draft and identify defective content; iteratively modifying and verifying the provenance of the defective content to obtain a final answer text that passes the verification; and adapting the final answer text that passes the verification to an expression style and emotional color and performing voice synthesis driving to generate the natural language answer and convert the natural language answer into voice output.

[0014] Optionally, the iterative correction and traceability verification based on the defect content comprises: converting the defect content into a set of veto constraints and a set of guiding constraints; encoding the veto constraints and the guiding constraints to generate a logical bias for the language generation process; under the guidance of the logical bias, performing a new round of text generation to generate a final natural language answer that is strengthened in factual consistency.

[0015] Based on the same inventive concept, the application also provides a voice real-time question and answer processing system based on a large model and a domain knowledge base, which comprises: a voice recognition module for obtaining a voice input stream of a user and intelligently deconstructing the voice input stream to generate a text stream; an intent analysis module for performing dependency syntax analysis and semantic role labeling on the text stream, analyzing core predicates and argument structures, and generating intent information and key information; a knowledge preloading module for dynamically accessing a domain knowledge base based on the intent information and the key information, positioning an initial query node, performing intent-driven predictive loading, obtaining a cluster of adjacent knowledge nodes and analyzing knowledge gap priorities, and generating a preloaded data subset; a context fusion module for constructing a context graph structure, fusing the intent information, the key information and the preloaded data subset through a context perception algorithm, and performing weighting pruning processing based on the context graph structure to generate a fusion result; and an answer generation output module for constructing an answer draft and traceability information based on the fusion result, combining the preloaded data subset for evaluation, obtaining defect content and performing constraint coding iterative correction, generating a natural language answer, and converting the natural language answer into a voice output.

[0016] Compared with the prior art, the application has the following advantages: 1. The application converts a serial information processing process into a parallel prediction and preparation process by deeply analyzing the user's intent and driving the predictive loading of knowledge, shortens the overall time consumption from receiving the user's voice to generating the answer, improves the real-time response speed of the question and answer system, and optimizes the interactive experience of the user in continuous and multi-round conversations.

[0017] 2. The application establishes a closed-loop answer quality control mechanism from generation to verification, performs forced cross-checking and iterative correction on the content generated by the large language model, makes the final output natural language answer strictly consistent with the domain knowledge base in terms of facts, effectively suppresses the generation of factual errors and information illusion, and thus improves the accuracy and credibility of the question and answer results.

[0018] 3. By introducing adaptive parsing capabilities for unseen intentions, this invention decomposes complex, undefined instructions into basic semantic primitives and dynamically plans their execution. This enables the system to handle novel and complex user requests that exceed its preset action space, significantly enhancing the system's generalization ability and intelligence level, making human-computer interaction more flexible and natural, and expanding its application scenarios.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a real-time voice question-answering processing method based on a large model and a domain knowledge base, according to an embodiment of the present invention.

[0022] Figure 2 This is the core predicate and action template mapping score heatmap of the embodiments of the present invention.

[0023] Figure 3 This is a radar chart showing the dynamic loading priority of knowledge nodes in an embodiment of the present invention.

[0024] Figure 4 The following diagram illustrates the construction and pruning of the atlas in an embodiment of the present invention.

[0025] Figure 5 This is a schematic diagram of the structure of a real-time voice question answering system based on a large model and a domain knowledge base, according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Reference Figure 1One embodiment of the present invention proposes a real-time voice question answering processing method based on a large model and a domain knowledge base. It adopts a technical solution that combines intent parsing-driven knowledge predictive loading, context fusion, and answer source correction, which can achieve low-latency response, high-precision intent understanding, and high-factuality answers in voice question answering.

[0028] The method described in this embodiment specifically includes: Acquire the user's voice input stream and perform intelligent sound wave deconstruction on the voice input stream to generate a text stream; Optionally, the generation of the text stream includes: A continuous voice input stream is acquired, and the voice input stream is separated into sound sources to extract the voice features of the target speaker. Parallel fast decoding and fine decoding are performed on the speech features of the target speaker. The fast decoding generates a preliminary text hypothesis with low latency, and the fine decoding performs deep recognition under the guidance of the preliminary text hypothesis. The output of the fine decoding is subjected to streaming confidence evaluation to identify and correct low-confidence recognition segments, thereby obtaining the corrected recognition result; At the pauses or semantic ends of the user's speech, the corrected recognition results are integrated and confirmed to generate a text stream.

[0029] Specifically, the user's voice input stream is continuously acquired via a microphone device, containing the user's voice as well as any ambient noise or other interfering sound sources. To ensure recognition accuracy, source separation processing is performed on this voice input stream. Source separation techniques, such as beamforming or independent component analysis, can isolate the target speaker's voice signal from the mixed audio signal, thereby suppressing noise and interference. The target speaker's voice features are then extracted. The time-domain voice waveform is converted into a series of short-time feature vectors that characterize the voice content, such as Mel-frequency cepstral coefficients. Parallel fast decoding and fine decoding are then performed on the target speaker's voice features. Fast decoding uses a computationally inexpensive acoustic and language model to perform a rapid initial recognition of the voice features, generating an initial text hypothesis with very low latency. Simultaneously, or with a slight delay, fine decoding begins. Fine decoding utilizes a larger, more powerful deep neural network model to perform deep recognition within the smaller search space defined by the initial text hypothesis. For example, fine decoding can re-score and sort the word candidates generated by fast decoding, thereby improving recognition accuracy without sacrificing too much speed. Simultaneously, streaming confidence assessment is performed while finely decoding the output text. This assessment process calculates a confidence score for each word or phrase generated by fine decoding to quantify the degree of certainty regarding the recognition result. For calculating the recognized segment... confidence score ,have: ; in, Representative and Identification Fragments The corresponding target speaker's speech feature sequence; This means that given speech features Under the given conditions, the recognition result is The posterior probability can be calculated by the fine decoder during the decoding process by normalizing the path probabilities in the word candidate lattice. When the confidence score of a certain recognition segment is lower than the preset confidence threshold, the recognition segment is marked as a low-confidence recognition segment. These low-confidence recognition segments are then corrected, such as by selecting a substitute result with the second-highest score but higher confidence from the decoder's candidate list, or by using a contextual language model for error correction, thereby generating a corrected recognition result. Finally, the user's pronunciation state is continuously monitored by a speech activity detection algorithm. When a natural pause in the user's speech is detected, such as the silence period after a sentence is finished, or when a relatively complete semantic endpoint of a sentence is determined by shallow syntactic analysis of the recognized part of the text, the previously accumulated and corrected recognition results are integrated. This confirmed and complete sentence or phrase constitutes the final output text stream.

[0030] Exemplarily, in a noisy power service hall environment, a customer consults a power customer service staff member wearing a portable assistant name tag, saying, "Hello, I would like to ask, which is more cost-effective for residential peak-valley electricity and ladder electricity prices? Also, what materials are required to apply for the installation of a charging pile now? Is it very troublesome?" A voice input stream containing the customer's voice, the conversations of people around, and background broadcasts is collected through a microphone array. First, the beamforming technology based on a deep neural network is applied for sound source separation to extract the clear speech features of the target speaker, which are a series of Mel-frequency cepstral coefficient vectors. Then, a lightweight recurrent neural network model is used to quickly generate a preliminary text hypothesis within 100 milliseconds. At the same time, the fine decoding module uses an acoustic model trained based on the DeepSeek large model and containing power industry terms for deep recognition. It is noted that "saving electricity price" and "installation registration" are words with similar pronunciations but lower probabilities in this business scenario, so they are corrected to "ladder electricity price" and "installation application". When the "installation registration" segment is output by the fine decoding, a streaming confidence assessment is performed simultaneously. By calculating its posterior probability, the confidence score of this segment is obtained as 0.65. Since 0.65 is lower than the preset confidence threshold of 0.8, this segment is identified as a low-confidence recognition segment. Immediately, the second-highest confidence result in the decoder candidate list, which is 0.92, the substitute result "installation application", is found and used for correction to obtain the corrected recognition result. When the voice activity detection algorithm monitors that there is more than 500 milliseconds of silence after the customer finishes speaking the last sentence, it is judged that this is a sentence semantic end point, and all the previously corrected and accumulated recognition results are integrated and confirmed, finally generating a stable and accurate text stream, "Hello, I would like to ask, which is more cost-effective for residential peak-valley electricity and ladder electricity prices? Also, what materials are required to apply for the installation of a charging pile now? Is it very troublesome?" By combining sound source separation, parallel decoding, and streaming confidence correction based on power domain knowledge, an accurate text stream can be generated in real time for the name tag assistant in a noisy power service hall environment with low latency.

[0031] Perform dependency syntactic analysis and semantic role annotation on the text stream, parse the core predicate and argument structure, and generate intent information and key information; Optionally, the generation of intent information and key information includes: Perform dependency syntactic analysis and semantic role annotation on the text stream, and deconstruct the core predicate and argument structure of the sentence; Map the core predicate and argument structure to the potential interaction action space defined by the domain knowledge base ontology structure to infer the deep intent of user interaction; Identify entities, attributes, constraint conditions, and relationship triples related to the deep intent in the text stream, and construct a structured query fragment; The deep intent and the structured query fragments are associated and integrated to generate intent information and key information.

[0032] Specifically, firstly, a deep language understanding model is applied to perform dependency parsing and semantic role labeling. Dependency parsing aims to reveal the grammatical dependencies between words in the text stream, constructing a dependency tree rooted at the core verb to clarify the principal-subordinate relationships among sentence components. Semantic role labeling further identifies the semantic roles played by each argument surrounding the core predicate, such as agent, patient, time, and place, based on the grammatical structure, thus deconstructing a natural language sentence into a standardized core predicate and argument structure. Next, the user's deep intent will be inferred. The domain knowledge base ontology structure is a systematic description of knowledge in a specific domain, defining the concept categories, attributes, and relationships between them within that domain. Based on this ontology, the potential interaction action space is constructed as a set containing all supported atomic or compound operations, each action defined by a standard predicate and a set of desired argument roles. For computational action templates... With input structure Mapping score ,have: ; in, It is the core predicate of the input text; It is an action template The predicate; It is the first of the input text Individual role; It is an action template The Individual role; The representative will write the text An embedding function that converts it into its semantic vector representation; It is a function for calculating the similarity between two vectors, and cosine similarity can be used; It is a weighting coefficient used to balance the importance of predicate similarity and argument structure similarity, and its value is obtained by optimization on the validation set; It is the total number of arguments contained in the input structure; The expected total number of arguments in the action template; For each input argument In all arguments of the action template In the given context, find the maximum semantic similarity. For example... Figure 2As shown, the mapping scores between the core predicate and the action template are visualized; the darker the cell color, the lower the corresponding mapping score. The action template with the highest score is selected as the inferred deep user intent. After determining the user's deep intent, entities, attributes, constraints, and relation triples strongly related to that intent are precisely identified in the original text stream. This process is a task-oriented named entity recognition and relation extraction task, focusing on the specific information corresponding to the argument roles required by the intent action template. For example, if the deep intent is to query the weather, location entities, time entities, and specific weather attributes will be specifically searched in the text stream. These identified information elements are organized into a structured query fragment, and all parameters required to execute the user command are encapsulated in a machine-readable format. Finally, the deep intent and the structured query fragment are associated and integrated. This integration process uses the deep intent as the operation command and the structured query fragment as the specific parameters of the command. The final output consists of intent information and key information. The intent information clarifies the type of operation the user wishes to perform, while the key information provides all the context parameters required to perform the operation. Together, they form a complete, unambiguous, structured user instruction that drives subsequent knowledge base access and answer generation.

[0033] For example, upon receiving the aforementioned text stream, dependency parsing and semantic role labeling are used to identify the core predicate "(more) cost-effective" and the argument structure as (comparison object 1 "peak-valley electricity", comparison object 2 "tiered electricity pricing") for the first sentence "Which is more cost-effective". Similarly, for the second sentence "What materials are needed to apply for charging pile installation", the core predicate "(query) materials" and the argument structure as (query topic "charging pile installation"). Then, the parsed core predicate and argument structure are mapped to the potential interaction action space defined by the power knowledge base ontology structure. A predefined action template is "compare (object 1, object 2, attribute)", where the attribute defaults to "economic efficiency". When calculating the mapping score, the core predicate of the input structure is "(more) cost-effective", and the predicate of the action template is "compare". Since the semantic vectors of "cost-effective" and "compare economic efficiency" have high similarity and the arguments match well, the calculated mapping score is also high. Therefore, the deep intent of the user interaction is inferred to be "electricity price comparison". Similarly, another deep intent is inferred as "business processing inquiry". Next, guided by this deep intent, structured query fragments are identified and constructed in the text stream, specifically: business type "electricity pricing scheme", entity "peak-valley electricity", entity "tiered electricity pricing", business type "electricity connection application", and entity "charging pile". Finally, the inferred deep intent and structured query fragments are integrated to generate intent information, namely "electricity price comparison" and "business processing inquiry", as well as key information, namely {comparison object "[peak-valley electricity, tiered electricity pricing]"} and {query topic "[charging pile connection application]", query attribute "[required materials]"}. Through deep language understanding and mapping with the domain action space, multiple business intents and all key information can be accurately parsed from the customer's colloquial questions.

[0034] Optionally, the method further includes: When the core predicate and argument structure cannot be directly mapped to the potential interaction action space, the predicates in the core predicate and argument structure are decomposed into basic semantic primitives. In the potential interaction space, search for operation chains composed of the basic semantic primitives that can be logically combined to realize the core predicate and argument structure; The operation chain is dynamically instantiated into a transactional execution plan, generating adaptive parsing results for unseen intentions.

[0035] Specifically, the core predicate, which cannot be directly mapped, is first decomposed into a set of basic semantic primitives. These basic semantic primitives are predefined and represent the lowest-level, indivisible atomic operations in the domain, such as "query" and "comparison." The decomposition process relies on a semantic decomposition dictionary or a trained semantic parsing model, capable of mapping a complex verb concept to a sequence of basic semantic primitives. Then, using this set of basic semantic primitives as the target, a search is performed within a set of atomic actions, also composed of basic semantic primitives. This set of atomic actions is a subset of the potential interaction action space, aiming to construct an operation chain logically equivalent to the original core predicate function. This search process can be viewed as a planning problem, finding an optimal path in a graph composed of basic atomic actions. A candidate operation chain consists of a series of atomic actions. For calculating candidate operation chains... Validity score ,have: ; in, These are the original, core predicates that cannot be directly mapped; It is the first in the operation chain Atom actions; The function represents an abstract representation of the logical combination of the effects of each action in the chain, and the semantic vector of the result is calculated by methods such as ordered weighted combination of individual action vectors; It is to perform a single atomic action. The estimated cost, which can be pre-set based on its computational complexity and resource consumption; `<parameter>` is a hyperparameter used as a penalty factor to balance semantic matching degree and execution cost, and is obtained through experimental tuning. The candidate operation chain with the highest validity score can be selected using a heuristic search algorithm, such as A* search. Once the optimal operation chain is found, it is dynamically instantiated into a transactional execution plan. The instantiation process fills the parameter slots of each atomic action in the operation chain with the entities, attributes, and constraints carried in the original core predicate and argument structure—that is, the key information—forming an executable sequence of specific steps. For example, an abstract operation chain of "find and compare," when faced with the input "compare the prices of product A and product B," will be instantiated into two specific steps: "query (the price of product A)" and "query (the price of product B)," with an additional "compare" operation. "Transactional" ensures the atomicity of the execution plan; that is, all steps in the entire operation chain either execute successfully or roll back entirely if any step fails. The final instantiated and transactional execution plan is the adaptive parsing result for the unseen intent.

[0036] For example, when addressing the customer question "Is it very troublesome?", assume that the complex subjective action "querying the degree of trouble" is not directly defined in the potential interaction action space. In this case, the adaptive parsing process is activated. First, the core predicate "(assess) trouble" is decomposed into a set of predefined basic semantic primitives, such as "query," "statistics," and "assessment." Then, an optimal operation chain that can achieve this intent is automatically searched and constructed. Through a search algorithm, a candidate operation chain with the highest validity score is found: "[query (subject, required materials), query (subject, processing procedure), query (subject, average time), comprehensive assessment (result set)]". When calculating the validity score, "(assess) trouble" is the original core predicate, and "query" or "comprehensive assessment" is the atomic action in the operation chain. Because the combined semantics of this operation chain highly match "assessing the degree of trouble" and its execution cost is reasonable, it scores the highest. Finally, this operation chain is dynamically instantiated into a transactional execution plan. Key information from the text stream, such as the subject "charging pile installation," is populated into this operation chain, dynamically generating a specific transactional execution plan. This plan queries the materials, procedures, and average time required for charging pile installation and obtains adaptive parsing results based on this information. This approach flexibly addresses customers' ambiguous subjective questions by breaking them down into a series of specific, queryable basic operations and dynamically combining them into an executable plan, thus enhancing robustness and adaptability.

[0037] Based on the intent information and key information, the domain knowledge base is dynamically accessed to locate the initial query node, and intent-driven predictive loading is executed to obtain a cluster of neighboring knowledge nodes and analyze the priority of knowledge gaps to generate a preloaded data subset. Optionally, generating the preloaded data subset includes: Based on the intent information and key information, the initial query node is located in the domain knowledge base; Starting from the initial query node, perform multi-hop path traversal and semantic expansion to predict the cluster of neighboring knowledge nodes that the user may be interested in; Obtain the user's historical interaction preferences and current dialogue status, and combine them with the correlation with the neighboring knowledge node cluster to calculate the dynamic loading priority of each node; Based on the dynamic loading priority, data entities and their associated attributes related to high-priority nodes are extracted asynchronously and differentially from the domain knowledge base to generate a preloaded data subset.

[0038] Specifically, the process begins with a precise search within the domain knowledge base, based on intent and key information, to locate the initial query node. This domain knowledge base is typically a graph-structured knowledge graph, where nodes represent entities and edges represent relationships between entities. Entities extracted from the key information, such as product names, personal names, or place names, are directly used as query entry points. These are matched against and their corresponding specific nodes within the knowledge graph, forming the initial query node set. Starting from these initial query nodes, multi-hop path traversal and semantic expansion are performed. The goal is to proactively predict and define the relevant knowledge range that the user might be interested in, forming a cluster of neighboring knowledge nodes. Multi-hop path traversal means not only focusing on the one-hop neighbors directly connected to the initial node but also exploring second-hop, third-hop, and even further nodes along relational paths in the knowledge graph. Semantic expansion provides directional guidance for this traversal process, prioritizing expansion along relational types that are semantically more relevant to the current user's intent. For example, if the intent is "comparison," edges related to "with specifications" or "price" will be traversed first, rather than edges related to "produced in." This guided exploration allows us to construct a cluster of neighboring knowledge nodes highly relevant to the current query. Next, to make data loading more personalized and forward-looking, we need to calculate the dynamic loading priority of each node in this cluster. To do this, we first obtain the user's historical interaction preferences and the current dialogue state. User historical interaction preferences can be obtained by vectorizing and aggregating the entities and intents of the user's past queries to obtain a preference vector. The current dialogue state includes the entities and topics already mentioned in this round of dialogue. For calculating a neighboring node... Dynamic loading priority ,have: ; in, It is a node The relevance score to the current query can be obtained by looking at the nodes. The distance to the graph path or the semantic similarity of the path to the initial query node is calculated. It is a node The correlation score with the user's historical interaction preferences is calculated by nodes. The similarity between the semantic vector and the user preference vector is obtained; It is a node The relevance score to the current dialogue state is calculated by the node. The similarity to the dialogue context vector is obtained; , , These are preset weighting coefficients, summing to one, used to adjust the influence ratio of different correlation components. They are obtained by optimizing metrics such as response latency on offline datasets. Finally, asynchronous and differentiated data extraction is performed from the domain knowledge base. Asynchronous extraction means that the data loading process is executed in an independent background thread, without blocking the main dialogue flow, thus ensuring the smoothness of front-end interaction. Differentiated extraction refers to setting a priority threshold, initiating data extraction requests only for nodes with a priority higher than this threshold (i.e., high-priority nodes), and obtaining these data entities and their key correlation attribute information from the domain knowledge base. This extracted structured data set constitutes the final pre-loaded data subset, like a small and refined data cache tailored to this query.

[0039] For example, the system first locates two entity nodes representing "peak-valley electricity" and "tiered pricing" in the electricity knowledge base as initial query nodes. Starting from these two nodes, a multi-hop path traversal and semantic expansion are performed. Since the user's intent involves "cost-effectiveness," expansion is prioritized along relational edges such as "applicable conditions" and "billing rules." It is predicted that the user might also be interested in "monthly electricity consumption tiers" and "prices for different voltage levels," thereby exploring related clusters of neighboring knowledge nodes. Next, the dynamic loading priority of each node in this cluster is calculated. , , The priority can be set to 0.5, 0.3, or 0.2. After comprehensive calculation, nodes such as "low-voltage residential users" and "monthly electricity consumption exceeding 400 kWh" received higher loading priority. Finally, in a separate background thread, data related to all high-priority nodes with a priority higher than the preset threshold are asynchronously extracted. This includes specific data entities and their associated attributes such as "peak, flat, and valley electricity periods and corresponding electricity prices," "electricity consumption and price standards for each tier of tiered electricity pricing," and "identity certificates and parking space ownership certificates required for charging pile installation," forming a pre-loaded data subset. By predictively loading data related to the user's current and potential points of interest, a foundation is laid for the subsequent rapid and accurate generation of answers, improving the smoothness of the interaction.

[0040] Optionally, calculating the dynamic loading priority of each node includes: Based on the user's historical interaction preferences and the current dialogue state, a vectorized representation of the user's cognitive state is constructed; Obtain the current user intent and infer the set of knowledge components from the ontology level of the domain knowledge base; Based on the comparison between the vectorized representation and the knowledge component set, the knowledge gap vector is calculated; Based on the knowledge gap vector, nodes in the neighboring knowledge node cluster that can meet the preset conditions for filling cognitive blind spots are assigned loading priority.

[0041] Specifically, a vectorized representation of the user's current cognitive state needs to be constructed. The user's historical interaction preferences can be obtained by aggregating the semantic embedding vectors of entities and intents involved in all past interactions. The current dialogue state is obtained by encoding the text content up to this point in the dialogue. These two vectorized representations are weighted and summed or fused using other mechanisms to ultimately form a comprehensive user cognitive state vector. Simultaneously, the user's current intent information is obtained, and inference is performed at the ontology level of the domain knowledge base based on this intent information. The ontology level of the domain knowledge base defines the abstract categories and conceptual frameworks of knowledge. For example, in the automotive domain, the ontology level might include high-level concepts such as "powertrain" and "chassis system." When the user's intent points to a specific entity, upward tracing is performed to infer a set of knowledge components most relevant to that entity and the user's intent. For example, when a user queries the acceleration performance of a car, the inferred set of relevant knowledge components might include ontology concepts such as "engine" and "transmission." Next, a knowledge gap vector is calculated based on the constructed user cognitive state vector and the inferred set of knowledge components. The core of this calculation is to assess the user's understanding of each necessary knowledge component. For each component in the knowledge component set, the similarity between the user's cognitive state vector and the semantic vector of that component is calculated. This similarity reflects the relevance of the user's historical and current conversational content to that knowledge component; each dimension of the knowledge gap vector corresponds to a knowledge component. This similarity is used to calculate the user's cognitive state vector within each knowledge component. Cognitive gap value ,have: ; in, It is a knowledge component The pre-trained semantic embedding vectors; The user's cognitive state vector is generated by weighted summation or other fusion mechanisms from the user's historical interaction preference vector and the current dialogue state vector. This is a preset threshold for cognitive familiarity, representing the degree to which a user is considered sufficiently familiar with a knowledge component. When the similarity between a user's cognitive state and that of a knowledge component is lower than a certain threshold... hour, A positive value indicates the degree of cognitive gap; conversely, a negative value indicates a lower degree of similarity. If so, it is believed that there is no cognitive gap. The value is zero. This knowledge gap vector intuitively depicts the dimensions of information a user still needs to understand in order to fully comprehend the current problem. Finally, it iterates through each node in the neighboring knowledge node cluster and determines whether the information contained in that node can effectively fill the identified cognitive blind spots. Specifically, if the information content of a node is semantically highly relevant to a knowledge component with a large gap value, the node is considered to meet the preset conditions for filling the cognitive blind spot. Based on this degree of matching, the node is assigned a higher loading priority. Figure 3 As shown in the radar chart, the loading priority of three different types of knowledge nodes in six key knowledge component domains is displayed. Node A scores the highest in acceleration performance dimension with 0.8, and the lowest in user evaluation dimension with 0.4. Node B stands out in the smart cockpit dimension with 0.9, and is relatively balanced in other dimensions. Node C scores the highest in the price range dimension with 0.9, and is relatively low in the technology dimension.

[0042] For example, firstly, based on the customer's historical interaction preferences, such as previous inquiries about "electricity bill inquiry" and "account number binding," and the current dialogue state (discussing "electricity pricing scheme" and "charging pile installation"), a comprehensive user cognitive state vector is constructed by weighted summing of the embedding vectors of these pieces of information. Simultaneously, based on the current user intent "what materials are needed for charging pile installation," a set of relevant knowledge components is inferred from the ontology level of the power knowledge base, including high-level concepts such as "installation qualifications," "grid capacity," and "construction specifications." Next, a knowledge gap vector is calculated. For a knowledge component in the knowledge component set, such as "community grid capacity assessment," the user's cognitive gap value for that component is calculated, where "community grid capacity assessment" is used as a pre-trained semantic embedding vector, and the cognitive familiarity threshold is set to 0.8. Assuming the customer has never inquired about information related to power grid engineering in the past, the user cognitive state vector and... The cosine similarity is only 0.2. Therefore, the user's cognitive gap value is 0.6, indicating a significant cognitive blind spot in this area. Finally, the priority of neighboring knowledge nodes is adjusted based on this knowledge gap vector. In the neighboring knowledge node cluster, the node "Application for Electricity Capacity Expansion Process" is given a high loading priority because its content is semantically highly correlated with the knowledge component "Community Power Grid Capacity Assessment" with a high gap value, thus meeting the preset condition for filling the cognitive blind spot. This method upgrades the data preloading strategy from simple "relevance"-driven to "cognitive gap"-driven, providing customers with more forward-looking and inspiring information services.

[0043] A context graph structure is constructed, and the intent information, key information, and preloaded data subset are fused using a context-aware algorithm. Weighting and pruning are then performed based on the context graph structure to generate a fusion result. Optionally, the generation of the fusion result includes: Construct a dynamic context graph structure, wherein the intent information and key information serve as the query root node, and the data entities in the preloaded data subset serve as knowledge leaf nodes; Calculate the semantic association strength between the query root node and each of the knowledge leaf nodes, and dynamically assign weights to the edges in the context graph structure based on the semantic association strength to obtain the edge weights; Based on the weights of the edges, the context graph structure is pruned to form focused reasoning paths; The focused inference path is serialized to generate a fusion result.

[0044] Specifically, a dynamic context graph structure is first constructed. In this temporarily constructed context graph structure, intent information and key information are uniformly represented as one or more query root nodes. These nodes are the starting point for reasoning and encapsulate the user's core needs. Simultaneously, each data entity in the preloaded data subset, such as specific knowledge points or facts, is instantiated as a knowledge leaf node. Query root nodes and knowledge leaf nodes are connected through potential logical relationships, collectively forming a local knowledge network around the current query. Next, the degree of association between different nodes in this context graph structure needs to be quantified, i.e., the semantic association strength between the query root nodes and each knowledge leaf node needs to be calculated. This strength is not a simple binary connection, but a continuous value calculated through a deep semantic matching model. For each query root node and knowledge leaf node, the semantic association strength between them can be calculated using a learnable fusion function, such as a small neural network, mapping the three dimensions of similarity between the query root node and knowledge leaf node in the semantic vector space, the shortest path features between them in the original knowledge base, and the fit of the knowledge leaf node with the current dialogue context, to a unified association strength score. Based on the calculated semantic association strength, each edge in the context graph structure pointing from the query root node to a knowledge leaf node is dynamically weighted to obtain the edge weight. Then, pruning is performed to eliminate noise and retain only the most core knowledge paths, thus forming a focused reasoning path. Pruning can be done by setting a weight threshold, i.e., removing all edges with weights below a preset threshold and their connected leaf nodes. A more refined method is to use algorithms such as maximum spanning trees, which, while ensuring graph connectivity, retain the edges with the highest total weight, thereby constructing a backbone reasoning network containing the highest-value information. This pruned network structure clearly demonstrates which key knowledge points, starting from the user's question, logically lead to a complete and relevant answer. Figure 4As shown, subgraph (a) represents the complete context graph structure before pruning, with the root being the query root node, first-level nodes being the entity nodes, second-level nodes being the attribute nodes, and third-level nodes being the extended knowledge nodes. Subgraph (b) represents the focused inference path after the final pruning, where edges with weights below 0.5 are removed, and leaf nodes isolated due to edge removal are removed. This focused inference path then needs to be serialized to generate a fusion result. The serialization process converts the nodes and weighted edges in the pruned graph into a structured text description according to a specific traversal order, such as a depth-first or breadth-first traversal starting from the query root node. This text not only contains the content of the selected knowledge entities but also explicitly indicates the logical relationships between them and their respective relevance to the user query. If a formatted text similar to "Regarding [query entity], the following points are worth noting: First, in [aspect A], its characteristic is [knowledge point 1], which is highly relevant to your question; second, in [aspect B], there is [knowledge point 2]..." can be generated, then this is the final fusion result.

[0045] For example, a dynamic context graph is first constructed by combining customer intent information ("comparing electricity price affordability", "inquiring about installation materials") with a pre-loaded subset of data. The user intent serves as the root node, while pre-loaded data such as "tiered electricity price first tier: 0.52 yuan / kWh" and "installation requires identification" serve as knowledge leaf nodes. The edges are dynamically weighted based on the semantic strength of the association between the two. For instance, the association strength weight between "comparing affordability" and "tiered electricity price first tier: 0.52 yuan / kWh" is 0.9, while the association weight with "installation requires identification" is only 0.2. Subsequently, pruning is performed to remove edges with lower weights, such as removing edges pointing from "comparing affordability" to "installation materials" information, thus forming two focused reasoning paths: one around electricity price comparison, and the other around installation materials. Finally, these paths are serialized into a structured fusion result, clearly connecting each customer's question with the most directly relevant answer data. By constructing and pruning context graphs, customer service personnel can accurately filter out the core logical chain related to the customer's current problem from a massive amount of pre-loaded power knowledge.

[0046] Based on the fusion results, a draft answer and source information are constructed. The draft answer is evaluated in conjunction with the preloaded data subset to obtain defective content and perform constraint coding iterative correction to generate a natural language answer. The natural language answer is then converted into speech output.

[0047] Optionally, generating a natural language response and converting the natural language response into speech output includes: Based on the fusion results, the language generation process is driven to produce a preliminary natural language response draft, and source information of key arguments in the response draft is generated simultaneously. The source information is cross-referenced and verified with the preloaded data subset to evaluate the draft answer and identify defective content. Based on the aforementioned defects, iterative corrections and source tracing verification are performed to obtain the final verified response text. The final verified answer text is adapted in terms of expression style and emotional tone and driven by speech synthesis to generate a natural language answer and convert it into speech output.

[0048] Specifically, the process begins by using this fusion result as the core for contextual and content guidance, driving a large language generation model such as DeepSeek or a similar architecture to produce an initial natural language draft response. Simultaneously, source information for key arguments in the draft response is generated. That is, for each factual statement or data sourced from the domain knowledge base in the draft response, a pointer or identifier pointing to its original source in the preloaded data subset is recorded. Next, the content in the draft response is cross-referenced and verified against its source information. Each key argument in the draft response with source information is automatically traced back to its corresponding data entity in the preloaded data subset for comparison. This verification process aims to evaluate the fidelity and completeness of the draft response. It checks for factual errors (statements inconsistent with the knowledge source), missing information (high-priority knowledge points from the fusion result not reflected in the draft response), and over-reasoning or "illusions" (statements in the draft response that cannot be supported by any evidence in the preloaded data subset). Through this verification, problematic parts of the draft response, i.e., defective content, can be automatically identified. For each identified defect, a corresponding correction strategy is implemented. If it's a factual error, the incorrect part is directly replaced with the correct information from the knowledge source; if it's missing information, the language generation model is instructed to include the missing key knowledge points in the next generation; if it's hallucination content, it is deleted or rewritten into a more rigorous speculative statement. After each correction, source verification is performed again to ensure the new content is consistent with the knowledge source. This iterative process is repeated until there are no more identifiable defects in the entire response text, resulting in the final verified response text. Finally, the response text is adapted in terms of expression style and emotional tone. Based on user profiles, dialogue history, or explicit instructions, the tone of the response can be adjusted to carry specific emotional connotations, such as friendliness and patience. After polishing, a speech synthesis engine is driven to convert this final verified response text into a high-quality speech waveform, which serves as the final natural language response and is played to the user through a speaker device, completing the closed loop of the entire real-time voice question-and-answer process.

[0049] For example, based on the fusion results, a language model trained on a DeepSeek large model is driven to generate a preliminary draft answer: "Regarding electricity prices, if your monthly electricity consumption is high, peak-valley electricity is more cost-effective. As for charging pile installation, you need to prepare your property ownership certificate and ID card, and we guarantee installation within a week." Source information is generated simultaneously for each key argument. Subsequently, cross-reference verification reveals that the draft answer omits a comparison of "applicable scenarios for tiered electricity pricing," which is a high-priority element in the fusion results, thus identifying it as a defect. Simultaneously, the compliance monitoring engine identifies "guarantee within a week" as an "absolute promise" among 20 sensitive word categories, also marking it as defective. An iterative correction and source verification process is then initiated, forcing the model to supplement electricity price comparison information based on this defective content and correcting the non-compliant statement. After revision and re-verification, the final verified response text was obtained: "Regarding electricity pricing, if your monthly electricity consumption frequently exceeds 400 kWh and you tend to use electricity during off-peak hours at night, then peak-valley pricing is usually more cost-effective; otherwise, tiered pricing is more suitable. Regarding charging station installation, you need to prepare identification documents and parking space ownership certificates, etc. After submitting the application, we will generally conduct an on-site inspection within 5 business days." Finally, the response text was style-adapted, and a clear, professional voice response was generated by a speech synthesis engine and played back to the customer. By introducing a cross-reference verification mechanism based on traceability information and an iterative correction mechanism using a compliance monitoring engine, factual errors and non-compliant statements were effectively reduced, providing customers with natural, accurate, and reliable business answers.

[0050] Optionally, the iterative correction and source tracing verification based on the defective content includes: The aforementioned defects are transformed into a set of veto constraints and a set of guiding constraints; The veto constraint and the guiding constraint are encoded to generate logical biases for the language generation process; Guided by the aforementioned logical bias, a new round of text generation is performed, producing a final natural language response reinforced with factual consistency.

[0051] Specifically, these deficient contents are first transformed into rejection and guidance constraints. For identified factual errors or illusory content, rejection constraints are generated to prevent the language generation model from producing statements that contradict the domain knowledge base in subsequent rounds of generation. For example, if the draft mistakenly states "Product A's price is 100 yuan," while the domain knowledge base states "150 yuan," a rejection constraint is generated, formally represented as "Prohibit the generation of sentences about 'Product A's price' that contain '100 yuan.'" For identified information omissions, guidance constraints are generated, mandating that the model must cover the omitted key knowledge points in the generated content. If an important difference in product comparison is not mentioned, a guidance constraint is generated, represented as "Must generate content describing the difference between 'Product A' and 'Product B' in terms of 'battery life.'" Subsequently, these textual rejection and guidance constraints need to be encoded. This encoding process transforms high-level semantic instructions into low-level adjustments to the model's output probability distribution. For veto constraints, the encoded result is to apply a large negative bias to the output probability of target words that might lead to the generation of prohibited content at each step of the model's word generation. For example, when generating a sentence about "product A price," the output probability value of the word "100" is subtracted by a near-infinite number, making its sampling probability approach zero. For guiding constraints, the encoded result is to introduce preferences into the attention mechanism or decoding strategy. For example, by modifying the context vector or increasing the score of candidate sequences containing key knowledge points during bundle search decoding, the likelihood of relevant content being generated can be increased. These encoded constraints together constitute a comprehensive logical bias signal. Guided by the logical bias, a new round of controlled text generation is performed. In this round of generation, the large language model is no longer completely free to continue writing text according to its internal probability distribution, but its decisions at each step are subject to real-time intervention and calibration by the aforementioned logical bias. This process is repeated, and source verification is performed again after each round of generation until the generated text fully meets the constraints and no longer generates new defective content, thus obtaining the final natural language answer.

[0052] For example, the identified defects are "omission of the applicable instructions for tiered electricity pricing" and "use of non-compliant commitment words such as 'guarantee'", which are then transformed into a set of guiding and disqualifying constraints. The guiding constraint is stated as "mandatory requirement that the generated content must include an explanation of the relationship between 'tiered electricity pricing' and 'monthly electricity consumption'". The disqualifying constraint is stated as "prohibition of using absolute terms such as 'guarantee' and 'certainty' when describing the processing time limit". Subsequently, encoding is performed to generate logical biases. For the guiding constraints, during the bundle search process in decoding, candidate sentences containing keywords such as "tiered electricity pricing" and "electricity consumption tiers" are given additional scores. For the disqualifying constraints, when generating words describing the processing time limit, a very large negative value is applied to the output probability of the word "guarantee", making its sampling probability close to zero. Finally, guided by these logical biases, a new round of text generation is performed, directly generating the statement, "...If your monthly electricity consumption is low, tiered electricity pricing is more suitable...We usually conduct an on-site inspection within 5 business days.", which serves as part of the final natural language response. By transforming abstract correction needs and compliance requirements into precise and quantifiable logical biases for the language model generation process, the efficiency and effectiveness of the correction are improved, enhancing the accuracy and compliance of the output content.

[0053] Based on the same inventive concept, such as Figure 5 As shown, the present invention also provides a real-time voice question-answering processing system based on a large model and a domain knowledge base, the system comprising: The speech recognition module is used to acquire the user's speech input stream and perform intelligent sound wave deconstruction on the speech input stream to generate a text stream; The intent parsing module is used to perform dependency parsing and semantic role labeling on the text stream, parse the core predicate and argument structure, and generate intent information and key information. The knowledge preloading module is used to dynamically access the domain knowledge base to locate the initial query node based on the intent information and key information, and perform intent-driven predictive loading to obtain a cluster of neighboring knowledge nodes and analyze the priority of knowledge gaps to generate a preloaded data subset. The context fusion module is used to construct a context graph structure, fuse the intent information, the key information, and the preloaded data subset through a context-aware algorithm, and perform weighting and pruning based on the context graph structure to generate a fusion result. The answer generation and output module is used to construct an answer draft and traceability information based on the fusion result, evaluate it in combination with the preloaded data subset, obtain defective content and perform constraint coding iterative correction, generate a natural language answer, and convert the natural language answer into speech output.

[0054] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0055] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A real-time speech question-answering processing method based on a large model and a domain knowledge base, characterized in that, The method includes: Acquire the user's voice input stream and perform intelligent sound wave deconstruction on the voice input stream to generate a text stream; Dependency parsing and semantic role labeling are performed on the text stream to parse the core predicate and argument structure and generate intent information and key information; Based on the intent information and key information, the domain knowledge base is dynamically accessed to locate the initial query node, and intent-driven predictive loading is executed to obtain a cluster of neighboring knowledge nodes and analyze the priority of knowledge gaps to generate a preloaded data subset. A context graph structure is constructed, and the intent information, key information, and preloaded data subset are fused using a context-aware algorithm. Weighting and pruning are then performed based on the context graph structure to generate a fusion result. Based on the fusion results, a draft answer and source information are constructed. The draft answer is evaluated in conjunction with the preloaded data subset to obtain defective content and perform constraint coding iterative correction to generate a natural language answer. The natural language answer is then converted into speech output.

2. The real-time speech question answering method based on a large model and domain knowledge base according to claim 1, characterized in that, The generated text stream includes: A continuous voice input stream is acquired, and the voice input stream is separated into sound sources to extract the voice features of the target speaker. Parallel fast decoding and fine decoding are performed on the speech features of the target speaker. The fast decoding generates a preliminary text hypothesis with low latency, and the fine decoding performs deep recognition under the guidance of the preliminary text hypothesis. The output of the fine decoding is subjected to streaming confidence evaluation to identify and correct low-confidence recognition segments, thereby obtaining the corrected recognition result; At the pauses or semantic ends of the user's speech, the corrected recognition results are integrated and confirmed to generate a text stream.

3. The real-time speech question answering method based on a large model and domain knowledge base according to claim 1, characterized in that, The generated intent information and key information include: Dependency parsing and semantic role labeling are performed on the text stream to deconstruct the core predicate and argument structure of the sentences; By mapping the core predicate and argument structure to the potential interaction action space defined by the domain knowledge base ontology structure, the deep intent of user interaction can be inferred. Identify entities, attributes, constraints, and relation triples related to the deep intent in the text stream to construct structured query fragments; The deep intent and the structured query fragments are associated and integrated to generate intent information and key information.

4. The real-time speech question answering method based on a large model and domain knowledge base according to claim 3, characterized in that, The method further includes: When the core predicate and argument structure cannot be directly mapped to the potential interaction action space, the predicates in the core predicate and argument structure are decomposed into basic semantic primitives. In the potential interaction space, search for operation chains composed of the basic semantic primitives that can be logically combined to realize the core predicate and argument structure; The operation chain is dynamically instantiated into a transactional execution plan, generating adaptive parsing results for unseen intentions.

5. The real-time speech question answering method based on a large model and domain knowledge base according to claim 1, characterized in that, The generated preloaded data subset includes: Based on the intent information and key information, the initial query node is located in the domain knowledge base; Starting from the initial query node, perform multi-hop path traversal and semantic expansion to predict the cluster of neighboring knowledge nodes that the user may be interested in; Obtain the user's historical interaction preferences and current dialogue status, and combine them with the correlation with the neighboring knowledge node cluster to calculate the dynamic loading priority of each node; Based on the dynamic loading priority, data entities and their associated attributes related to high-priority nodes are extracted asynchronously and differentially from the domain knowledge base to generate a preloaded data subset.

6. The real-time speech question answering method based on a large model and domain knowledge base according to claim 5, characterized in that, The calculation of the dynamic loading priority of each node includes: Based on the user's historical interaction preferences and the current dialogue state, a vectorized representation of the user's cognitive state is constructed; Obtain the current user intent and infer the set of knowledge components from the ontology level of the domain knowledge base; Based on the comparison between the vectorized representation and the knowledge component set, the knowledge gap vector is calculated; Based on the knowledge gap vector, nodes in the neighboring knowledge node cluster that can meet the preset conditions for filling cognitive blind spots are assigned loading priority.

7. The real-time speech question answering method based on a large model and domain knowledge base according to claim 1, characterized in that, The generated fusion results include: Construct a dynamic context graph structure, wherein the intent information and key information serve as the query root node, and the data entities in the preloaded data subset serve as knowledge leaf nodes; Calculate the semantic association strength between the query root node and each of the knowledge leaf nodes, and dynamically assign weights to the edges in the context graph structure based on the semantic association strength to obtain the edge weights; Based on the weights of the edges, the context graph structure is pruned to form focused reasoning paths; The focused inference path is serialized to generate a fusion result.

8. The real-time speech question answering method based on a large model and domain knowledge base according to claim 1, characterized in that, The step of generating a natural language response and converting the natural language response into speech output includes: Based on the fusion results, the language generation process is driven to produce a preliminary natural language response draft, and source information of key arguments in the response draft is generated simultaneously. The source information is cross-referenced and verified with the preloaded data subset to evaluate the draft answer and identify defective content. Based on the aforementioned defects, iterative corrections and source tracing verification are performed to obtain the final verified response text. The final verified answer text is adapted in terms of expression style and emotional tone and driven by speech synthesis to generate a natural language answer and convert it into speech output.

9. The real-time speech question answering method based on a large model and domain knowledge base according to claim 1, characterized in that, The iterative correction and source tracing verification based on the defective content includes: The aforementioned defects are transformed into a set of veto constraints and a set of guiding constraints; The veto constraint and the guiding constraint are encoded to generate logical biases for the language generation process; Guided by the aforementioned logical bias, a new round of text generation is performed, producing a final natural language response reinforced with factual consistency.

10. A real-time speech question-answering processing system based on a large model and a domain knowledge base, applied to the real-time speech question-answering processing method based on a large model and a domain knowledge base as described in any one of claims 1-9, characterized in that, The system includes: The speech recognition module is used to acquire the user's speech input stream and perform intelligent sound wave deconstruction on the speech input stream to generate a text stream; The intent parsing module is used to perform dependency parsing and semantic role labeling on the text stream, parse the core predicate and argument structure, and generate intent information and key information. The knowledge preloading module is used to dynamically access the domain knowledge base to locate the initial query node based on the intent information and key information, and perform intent-driven predictive loading to obtain a cluster of neighboring knowledge nodes and analyze the priority of knowledge gaps to generate a preloaded data subset. The context fusion module is used to construct a context graph structure, fuse the intent information, the key information, and the preloaded data subset through a context-aware algorithm, and perform weighting and pruning based on the context graph structure to generate a fusion result. The answer generation and output module is used to construct an answer draft and traceability information based on the fusion result, evaluate it in combination with the preloaded data subset, obtain defective content and perform constraint coding iterative correction, generate a natural language answer, and convert the natural language answer into speech output.

Images