Government affair information intelligent processing system and method based on voice recognition

Abstract

The invention relates to the technical field of government affair services, in particular to a government affair information intelligent processing system and method based on voice recognition. Comprising a voice acquisition preprocessing module, a voice recognition dynamic dictionary updating module, a natural language understanding policy element association module, a policy knowledge graph logical reasoning network module, a semantic mapping adaptive standardization module, a policy matching multi-round iterative optimization module and an interactive guidance system state feedback module. All the modules are sequentially connected through data streams, and voice signal noise reduction, text recognition optimization, intention classification and entity extraction, policy logical reasoning, mapping from spoken language to standard policies, multi-round policy matching sorting and interactive guidance are achieved. Through dynamic dictionary updating, multi-strategy adaptive mapping, reinforcement learning driving optimization and a closed-loop feedback mechanism, a semantic gap between policy text normalization and entrepreneurs spoken language cognition is effectively bridged, and policy matching accuracy and resource utilization efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of government affairs service, and in particular to a government affair information intelligent processing system and method based on voice recognition. BACKGROUND

[0002] The existing policy recommendation technology based on keyword matching or simple classification is usually established on the basis of a natural language processing framework, and through text analysis and feature extraction on user queries and policy documents, the content is converted into a computable form. Then, the system uses keyword matching algorithms or pre-trained classification models to compare the queries with the entries in the policy library, and generates a recommendation result according to semantic similarity or category labels. This process relies on the synergy of statistical learning and rule engines to achieve the practical function of quickly filtering relevant policies from massive data.

[0003] The existing policy recommendation technology based on keyword matching or simple classification has the following technical pain points. Specifically, there is a fundamental contradiction between the standardization requirements of policy texts and the randomness of natural language descriptions. Policy provisions usually use precise legal terminology and structured expressions to define the applicable objects and conditions, such as "the total amount of R&D expenses accounts for a certain percentage of the total sales revenue" or "a certain number of college graduates are recruited and pay social security as required". However, entrepreneurs often describe their actual business scenarios in colloquial language, such as "investing a considerable amount of money in experimental equipment" or "recruiting a few college graduates this year". The difference in expression methods makes it difficult for existing information retrieval-based matching systems to establish accurate mappings and understand that "investing a considerable amount of money" may correspond to the "R&D expense ratio" policy indicator, or determine whether "a few college graduates" meet the specific requirements of the number and social security in the policy. Ultimately, a large number of entrepreneurial subjects who meet the policy conditions but do not express them properly are excluded from the matching results, resulting in a mismatch between policy resources and beneficiaries. SUMMARY

[0004] To address the deficiencies in the prior art, the present application provides a government affair information intelligent processing system and method based on voice recognition, which solves the technical problem that due to the huge "semantic gap" between the professional and standardized expression of entrepreneurial policy texts and the daily colloquial and scenario-based cognition of entrepreneurs, entrepreneurs cannot accurately locate and successfully apply for support policies that are highly matched to their actual situation, resulting in policy resource mismatch and waste.

[0005] To solve the above technical problems, the specific content of the present application is as follows: In a first aspect, the present application provides a government affair information intelligent processing system based on voice recognition, comprising: a voice acquisition and preprocessing module that acquires the voice signal of an entrepreneur and performs noise reduction and endpoint detection, and outputs a pure voice frame sequence; The speech recognition dynamic dictionary update module receives the clean speech frame sequence from the speech acquisition and preprocessing module, generates the recognition text through the speech recognition model, and optimizes the recognition process using the dynamic dictionary update mechanism. The Natural Language Understanding Policy Element Association Module receives the recognized text from the Speech Recognition Dynamic Dictionary Update Module, performs intent classification and entity extraction, queries the Policy Knowledge Graph Logical Reasoning Network Module to add policy association weights to entities, and outputs a weighted semantic unit sequence. The policy knowledge graph logical reasoning network module stores the hypergraph structure of policy clauses, receives the semantic unit sequence from the natural language understanding policy element association module, performs logical reasoning, and derives the policy condition set; The semantic mapping adaptive standardization module receives the weighted semantic unit sequence from the natural language understanding policy element association module, converts spoken entities into standard policy indicators through multi-strategy mapping, receives feedback confidence from the policy matching multi-round iterative optimization module to adaptively adjust the mapping strategy, and outputs a structured query vector. The policy matching multi-round iterative optimization module receives the structured query vector from the semantic mapping adaptive standardization module, calculates the matching degree on the policy knowledge graph, and drives multi-round query reconstruction through reinforcement learning to output a ranked policy list; The interactive guidance system status feedback module receives the sorted policy list from the policy matching multi-round iterative optimization module, generates guidance questions based on the uncertainty of the list status, collects user feedback data, and distributes the user feedback data to the speech recognition dynamic dictionary update module and the semantic mapping adaptive standardization module.

[0006] Furthermore, in the speech recognition-based intelligent government information processing system of the present invention, the speech acquisition and preprocessing module uses a microphone array to acquire speech signals, uses beamforming technology to suppress environmental noise, uses an endpoint detection algorithm to extract effective speech segments, and outputs the clean speech frame sequence to the speech recognition dynamic dictionary update module.

[0007] Furthermore, in the speech recognition-based intelligent processing system for government information described in this invention, the speech recognition dynamic dictionary update module adopts an end-to-end speech recognition model and introduces an adversarial domain adaptive method to reduce the impact of dialect variations; the dynamic dictionary update mechanism receives high-frequency policy terms output by the natural language understanding policy element association module, adjusts the decoding weights of the acoustic model and the language model, and forms a reverse optimization path from semantic understanding to speech recognition.

[0008] Further, the voice recognition-based government information intelligent processing system provided by the application, the natural language understanding policy element association module uses a Transformer architecture model for intent recognition and entity recognition, and the entity recognition combines a policy domain dictionary and a conditional random field model; a policy element pre-association mechanism queries a policy knowledge graph logical reasoning network module when performing entity extraction, assigns a policy association weight to the entity, and generates a weighted semantic unit sequence that is transmitted to a semantic mapping adaptive standardization module.

[0009] Further, the voice recognition-based government information intelligent processing system provided by the application, the policy knowledge graph logical reasoning network module adopts a hypergraph to represent the logical relationship between policy indicators and declaration conditions, integrates a rule reasoning engine to deduce implicit policy requirements, receives an entity query request from the natural language understanding policy element association module, outputs a policy condition set to a policy matching multi-round iteration optimization module, and feeds back an entity policy association weight to the natural language understanding policy element association module.

[0010] Further, the voice recognition-based government information intelligent processing system provided by the application, the semantic mapping adaptive standardization module performs rule mapping, fuzzy set membership degree calculation and collaborative filtering prediction in parallel, and maps a spoken description into a standard policy field; the mapping confidence score of the policy matching multi-round iteration optimization module is received, the strategy weight is adjusted through the gradient descent method, and the structured query vector is output to the policy matching multi-round iteration optimization module.

[0011] Further, the voice recognition-based government information intelligent processing system provided by the application, the policy matching multi-round iteration optimization module embeds the structured query vector into the vector space of the policy knowledge graph, and uses a graph neural network to calculate the node similarity; a reinforcement learning framework takes the declaration success rate as a reward signal, optimizes the query reconstruction strategy, outputs the sorted policy list to an interactive guidance system state feedback module, and feeds back the matching confidence to the semantic mapping adaptive standardization module.

[0012] Further, the voice recognition-based government information intelligent processing system provided by the application, the interactive guidance system state feedback module generates a guidance question sequence based on a partially observable Markov decision process model; user feedback data synchronously updates the acoustic model of the voice recognition dynamic dictionary update module, the entity recognition strategy of the natural language understanding policy element association module and the mapping weight of the semantic mapping adaptive standardization module, forming a cross-module collaborative learning cycle.

[0013] Further, the voice recognition-based government information intelligent processing system provided by the application realizes multi-level self-adaptation through bidirectional data flow among modules: the recognition confidence of the voice recognition dynamic dictionary updating module triggers the recalibration of the natural language understanding policy element association module, the matching result of the policy matching multi-round iteration optimization module feeds back to optimize the strategy of the semantic mapping self-adaptation standardization module, and the state feedback module of the interactive guidance system feeds back the user behavior data to drive dynamic dictionary updating and mapping adjustment, so that the system has the continuous evolution ability of policy language model and knowledge model.

[0014] In a second aspect, the application provides a voice recognition-based government information intelligent processing method, which is applied to the voice recognition-based government information intelligent processing system and includes the following steps: Step 1: collecting the voice signal of an entrepreneur, performing noise reduction and endpoint detection, and outputting a pure voice frame sequence; Step 2: receiving the pure voice frame sequence, generating a recognition text through a voice recognition model, and optimizing the recognition process by using a dynamic dictionary updating mechanism; Step 3: receiving the recognition text, performing intent classification and entity extraction, querying a policy knowledge graph to add a policy association weight to the entity, and outputting a weighted semantic unit sequence; Step 4: storing a hypergraph structure of a policy clause, receiving the semantic unit sequence to perform logical reasoning, and deriving a policy condition set; Step 5: receiving the weighted semantic unit sequence, converting the spoken entity into a standard policy index through multi-strategy mapping, receiving feedback confidence to adaptively adjust the mapping strategy, and outputting a structured query vector; Step 6: receiving the structured query vector, calculating a matching degree on the policy knowledge graph, and outputting a ranked policy list through reinforcement learning-driven multi-round query reconstruction; Step 7: receiving the ranked policy list, generating a guidance question according to the list state uncertainty, collecting user feedback data, and using the user feedback data to update the dynamic dictionary and semantic mapping strategy of the voice recognition process.

[0015] The application has the following advantages: The application obtains the speech signal of an entrepreneur through a speech acquisition preprocessing module and outputs a pure speech frame sequence, a speech recognition dynamic dictionary updating module optimizes the recognition process by using a dynamic dictionary updating mechanism to improve the speech recognition accuracy, a natural language understanding policy element association module performs intent classification and entity extraction, and queries a policy knowledge graph to add a policy association weight to the entity to output a weighted semantic unit sequence, a policy knowledge graph logical reasoning network module derives a policy condition set through logical reasoning, a semantic mapping adaptive standardization module converts the spoken entity into a standard policy index through multi-strategy mapping and adaptively adjusts the mapping strategy based on the feedback confidence, and a policy matching multi-round iteration optimization module calculates the matching degree using a graph neural network and reconstructs through multi-round query driven by reinforcement learning, and an interactive guidance system state feedback module generates a guidance question according to the state uncertainty and collects user feedback data to form a closed-loop optimization path; the system realizes multi-level adaptation through bidirectional data flow between modules, so that the speech recognition accuracy is optimized as the usage frequency of policy terms increases, the semantic mapping strategy is dynamically adjusted according to the matching effect, and the policy matching process gradually approaches the optimal solution through multi-round iteration, thereby effectively bridging the semantic gap between the standardized expression of the policy text and the spoken cognitive understanding of the entrepreneur, and realizing the accurate matching and efficient use of policy resources. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on the drawings.

[0017] Figure 1 The flowchart of the intelligent processing method of government information based on speech recognition of the present application. DETAILED DESCRIPTION

[0018] In order to make the technical solutions of the present application clearer, the following will combine the specific embodiments of the present application and the corresponding drawings to clearly and completely describe the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. The following will combine the drawings to specifically describe the present application provided by the embodiments of the present application. In order to better understand the purpose of the present application, the following will further describe the present application in detail.

[0019] In the first aspect, the intelligent processing system of government information based on speech recognition of the present application comprises: A speech acquisition preprocessing module obtains the speech signal of an entrepreneur and performs noise reduction and endpoint detection to output a pure speech frame sequence. The voice recognition dynamic dictionary updating module receives the pure voice frame sequence from the voice collection preprocessing module, generates recognition text through a voice recognition model, and optimizes the recognition process using a dynamic dictionary updating mechanism. The natural language understanding policy element association module receives the recognition text from the voice recognition dynamic dictionary updating module, performs intent classification and entity extraction, queries the policy knowledge graph logical reasoning network module from the policy knowledge graph to add policy association weights to entities, and outputs a weighted semantic unit sequence. The policy knowledge graph logical reasoning network module stores the hypergraph structure of policy provisions, receives the semantic unit sequence from the natural language understanding policy element association module for logical reasoning, and derives a policy condition set. The semantic mapping adaptive standardization module receives the weighted semantic unit sequence from the natural language understanding policy element association module, converts spoken language entities into standard policy indicators through multi-strategy mapping, receives feedback confidence from the policy matching multi-round iteration optimization module to adaptively adjust the mapping strategy, and outputs a structured query vector. The policy matching multi-round iteration optimization module receives the structured query vector from the semantic mapping adaptive standardization module, calculates the matching degree on the policy knowledge graph, and drives multi-round query reconstruction through reinforcement learning, and outputs a ranked policy list. The interactive guidance system state feedback module receives the ranked policy list from the policy matching multi-round iteration optimization module, generates guidance questions according to the list state uncertainty, collects user feedback data, and distributes the user feedback data to the voice recognition dynamic dictionary updating module and the semantic mapping adaptive standardization module.

[0020] The voice collection preprocessing module collects the voice signal of the entrepreneur through a microphone array, uses beamforming technology to suppress environmental noise, and extracts the effective voice segment through an endpoint detection algorithm, and outputs a pure voice frame sequence. This process provides high-quality input data for subsequent voice recognition.

[0021] The voice recognition dynamic dictionary updating module receives the pure voice frame sequence, uses an end-to-end voice recognition model to convert the voice into recognition text, and optimizes the recognition process using a dynamic dictionary updating mechanism. The dynamic dictionary dynamically adjusts the decoding weights of the acoustic model and the language model according to the usage frequency of policy terms, thereby reducing the recognition errors caused by dialects and spoken language variants and improving the recognition accuracy.

[0022] The Natural Language Understanding (NLE) policy element association module receives and recognizes text, performing intent classification and entity extraction. Intent classification employs a Transformer architecture model for fine-grained classification, while entity extraction combines a policy domain dictionary and a conditional random field model to improve accuracy. The module also queries a policy knowledge graph logic reasoning network module to add policy-related weights to the extracted entities, outputting a weighted sequence of semantic units to provide structured information for subsequent policy matching.

[0023] The policy knowledge graph logic reasoning network module stores a hypergraph structure of policy provisions, where nodes represent policy indicators and application conditions, and hyperedges characterize logical relationships. The module receives a weighted sequence of semantic units and, through a logic reasoning engine based on Datalog rules, derives a complete set of policy conditions, revealing the implicit requirements between policies.

[0024] The semantic mapping adaptive standardization module receives a weighted sequence of semantic units and transforms spoken entities into standard policy indicators through multi-strategy mapping. Strategies include rule mapping, fuzzy set membership calculation, and collaborative filtering prediction. The module also receives feedback confidence from the policy matching multi-round iterative optimization module, dynamically adjusts the mapping strategy weights using gradient descent, and outputs a structured query vector.

[0025] The policy matching multi-round iterative optimization module receives structured query vectors, calculates the matching degree using a graph neural network on a policy knowledge graph, and drives multi-round query reconstruction through a reinforcement learning framework. Reinforcement learning uses the application success rate as a reward signal to optimize the query strategy, gradually narrowing down the policy candidate set and outputting a ranked policy list.

[0026] The interactive guidance system's state feedback module receives a list of ranking policies and generates guidance questions based on a partially observable Markov decision process model to address state uncertainty. This module collects user feedback data and distributes it to the speech recognition dynamic dictionary update module and the semantic mapping adaptive standardization module. These updates update the acoustic model, entity recognition strategy, and mapping weights, forming a closed-loop optimization path and improving system adaptability.

[0027] Specifically, in the speech recognition-based intelligent government information processing system of the present invention, the speech acquisition and preprocessing module uses a microphone array to acquire speech signals, uses beamforming technology to suppress environmental noise, uses an endpoint detection algorithm to extract effective speech segments, and outputs the clean speech frame sequence to the speech recognition dynamic dictionary update module.

[0028] The voice collection preprocessing module adopts a microphone array to collect voice signals. The microphone array is composed of multiple microphone sensors in a specific geometric layout, and multi-channel synchronous sampling is used to obtain multi-channel voice signals, thereby capturing the spatial and temporal characteristics of sound. The collected analog voice signals are converted into digital signals by an analog-to-digital converter, providing a basis for subsequent digital processing.

[0029] The module uses a beamforming technique to suppress environmental noise. The beamforming technique calculates the relative delay and weighting coefficient between each microphone channel through an adaptive algorithm, forming a directional beam main lobe pointing to the user's sound source direction. The beamforming process uses spatial filtering principles to enhance the target voice signal while attenuating interference noise from other directions, thereby improving signal quality.

[0030] The module uses an endpoint detection algorithm to extract valid voice segments. The endpoint detection algorithm dynamically sets double-threshold thresholds based on short-time energy and short-time zero-crossing rate characteristics, and identifies the starting point and ending point of the voice segment by analyzing the amplitude change and frequency characteristics of the voice signal. The endpoint detection algorithm distinguishes between voice activity regions and silent segments, eliminates invalid noise segments, and retains pure voice data.

[0031] The extracted valid voice segments are subjected to frame processing, which divides continuous voice signals into a sequence of short-time frames of fixed length. Overlapping is used between frames to maintain temporal continuity. Each frame of voice signal is windowed using a window function such as the Hamming window to reduce spectral leakage. Finally, the module outputs the processed pure voice frame sequence to the voice recognition dynamic dictionary update module for use in subsequent voice recognition processes.

[0032] Specifically, the voice recognition-based government information intelligent processing system described in the present application uses an end-to-end speech recognition model in the voice recognition dynamic dictionary update module, and introduces an adversarial domain adaptation method to reduce the influence of dialect variation. The dynamic dictionary update mechanism receives high-frequency policy terms output by the natural language understanding policy element association module, adjusts the decoding weights of the acoustic model and the language model, and forms a reverse optimization path from semantic understanding to speech recognition.

[0033] The voice recognition dynamic dictionary update module uses an end-to-end speech recognition model to process the pure voice frame sequence. The end-to-end speech recognition model converts the voice frame sequence into a high-dimensional feature representation through an encoder, and then uses a decoder to directly map the feature sequence to a text sequence, realizing end-to-end conversion from voice to text. The encoder usually uses a convolutional neural network combined with a recurrent neural network structure to extract acoustic features of the voice frame by frame; the decoder dynamically aligns the acoustic features and text labels based on the attention mechanism to generate recognized text.

[0034] The module introduces an adversarial domain adaptation method to reduce the influence of dialect variation, and the adversarial domain adaptation method adds a domain discriminator in the training process of the speech recognition model, the domain discriminator attempts to distinguish whether the input speech features come from a standard speech domain or a dialect domain, and the encoder learns to generate a domain-invariant feature representation through adversarial training. The adversarial domain adaptation method minimizes the feature distribution difference between the dialect domain and the standard domain, and improves the robustness of the model to dialect variations.

[0035] The dynamic dictionary updating mechanism receives high-frequency policy terms output by the natural language understanding policy element association module, obtains the high-frequency policy terms by counting the frequency of the appearance of entities in the natural language understanding policy element association module, and forms a dynamic dictionary word table. The dynamic dictionary updating mechanism integrates the set of high-frequency policy terms into the decoding dictionary of the speech recognition model, and preferentially matches policy-related words.

[0036] The dynamic dictionary updating mechanism adjusts the decoding weights of the acoustic model and the language model, the acoustic model decoding weight adjustment is based on the acoustic feature likelihood of the high-frequency policy term, and the language model decoding weight adjustment is based on the probability distribution of the policy term in the language model. In the decoding process, the acoustic model and the language model give higher weights to the high-frequency policy term, and improve the recognition priority of the policy term.

[0037] The speech recognition dynamic dictionary updating module forms a reverse optimization path from semantic understanding to speech recognition through weight adjustment, the high-frequency policy term in the semantic understanding stage is fed back to the speech recognition stage, and the decoding strategy is dynamically optimized. The reverse optimization path enables the speech recognition model to adapt to the policy dialogue scene, gradually improves the recognition accuracy of the policy term, and reduces the semantic deviation between the spoken language expression and the policy text.

[0038] Specifically, the speech recognition-based government information intelligent processing system provided by the application uses a Transformer architecture model for intent recognition and entity recognition in the natural language understanding policy element association module, and combines a policy domain dictionary and a conditional random field model for entity recognition; the policy element pre-association mechanism queries the policy knowledge graph logical reasoning network module when extracting entities, assigns policy association weights to the entities, and generates the weighted semantic unit sequence to be transmitted to the semantic mapping adaptive standardization module.

[0039] The natural language understanding policy element association module receives the recognized text output by the speech recognition dynamic dictionary updating module, first performs word segmentation and vectorization processing on the text, and converts the text into a word vector sequence. The module uses a Transformer architecture model to process the word vector sequence, calculates the dependency relationship between the word vectors through a self-attention mechanism, and generates context-dependent encoding representations. The encoder output of the Transformer architecture model is used for intent recognition and entity recognition two sub-tasks.

[0040] The intent recognition subtask maps the output of the Transformer encoder to the intent category space through a fully connected layer, calculates the probability distribution of each intent category using the Softmax function, and selects the category with the highest probability as the user intent classification result. Intent recognition covers multiple government scenario intents such as policy query, qualification assessment, and declaration guidance.

[0041] The entity recognition subtask combines a policy domain dictionary and a conditional random field model for sequence labeling. The policy domain dictionary contains policy-related terms and entity type mapping rules, providing domain prior knowledge for entity recognition. The conditional random field model receives the output features of the Transformer encoder, learns the transition probabilities between labels, and decodes the optimal entity label sequence through the Viterbi algorithm. The entity recognition process labels policy element entities such as personnel size, capital investment, and industry type.

[0042] The policy element pre-association mechanism sends the identified entity list to the policy knowledge graph logical reasoning network module for querying after entity extraction is completed. The policy knowledge graph logical reasoning network module returns the association strength indicator of each entity with the policy clause, and calculates the policy association weight based on the node degree and path weight of the entity in the knowledge graph. The policy element pre-association mechanism assigns a policy association weight to each entity based on the returned association strength, and the weight value reflects the importance of the entity in the policy matching process.

[0043] The module combines the weighted entities and intent classification results into a semantic unit sequence, each semantic unit containing entity content, entity type, and policy association weight. The semantic unit sequence is arranged in the original order in the text, maintaining semantic coherence. Finally, the natural language understanding policy element association module transmits the weighted semantic unit sequence to the semantic mapping adaptive standardization module to provide structured input for subsequent policy matching.

[0044] Specifically, the government information intelligent processing system based on speech recognition provided by the present application adopts a hypergraph to represent the logical relationship between policy indicators and declaration conditions, integrates a rule-based reasoning engine to deduce implicit policy requirements; receives entity query requests from the natural language understanding policy element association module, outputs policy condition sets to the policy matching multi-round iteration optimization module, and feeds back entity policy association weights to the natural language understanding policy element association module.

[0045] The policy knowledge graph logical reasoning network module stores policy clauses in a hypergraph structure, which is composed of nodes and hyperedges. The nodes represent policy indicators and declaration condition entities, and the hyperedges represent complex logical relationships between multiple nodes. During the construction of the hypergraph, policy clauses are parsed into atomic condition units, each of which is mapped to a hypergraph node, and logical connection relationships such as AND, OR, and NOT are mapped to hyperedges, forming the topology of the policy knowledge graph.

[0046] The module integrates a rule reasoning engine, which is based on a Datalog logical rule base containing derivation rules defined by policy domain experts. The rule reasoning engine receives hypergraph structure data and derives implicit policy requirements from explicit policy conditions through pattern matching and rule triggering mechanisms. The reasoning process uses a forward chain reasoning algorithm to traverse hypergraph nodes and hyperedges, generating a complete policy condition dependency relationship network.

[0047] The policy knowledge graph logical reasoning network module receives entity query requests from the natural language understanding policy element association module, which contains a list of entities in the semantic unit sequence. The module performs similarity matching between the entity list and the hypergraph nodes, calculates the cosine similarity between entity vectors and node vectors using graph embedding algorithms, and identifies relevant policy nodes.

[0048] The module queries the policy conditions associated with the matching nodes through a hypergraph traversal algorithm. The traversal algorithm starts from the matching node and searches for connected node sets along the hyperedge direction, collecting all directly and indirectly associated policy conditions. The search process uses a breadth-first strategy to ensure coverage of all possible condition paths, forming a policy condition set.

[0049] The policy condition set is processed through logical normalization, which converts the condition expressions into standard conjunctive normal form, eliminates redundant conditions, and merges duplicate clauses. The normalized policy condition set is output to the policy matching multi-round iteration optimization module as the baseline condition set for policy matching.

[0050] The module calculates entity policy association weights based on the centrality index of entities in the hypergraph and the condition association strength. The centrality index uses the node betweenness centrality algorithm, and the association strength is weighted by the path weight of the entity and the policy condition and the rule confidence. The calculated entity policy association weights are fed back to the natural language understanding policy element association module for optimizing the weight distribution of semantic units.

[0051] The policy knowledge graph logical reasoning network module achieves deep mining and association analysis of policy conditions through the synergistic effect of hypergraph querying and rule reasoning, providing a logically complete condition basis for policy matching. The feedback mechanism of the module forms a closed-loop optimization with the natural language understanding module, improving the accuracy of policy element identification.

[0052] Specifically, the speech recognition-based government information intelligent processing system of the present application, the semantic mapping adaptive standardization module performs rule mapping, fuzzy set membership calculation and collaborative filtering prediction in parallel, maps the spoken description into a standard policy field; receives the mapping confidence score of the policy matching multi-round iteration optimization module, adjusts the policy weight by gradient descent method, and outputs the structured query vector to the policy matching multi-round iteration optimization module.

[0053] The semantic mapping adaptive standardization module receives the weighted semantic unit sequence transmitted by the natural language understanding policy element association module, which contains entity content, entity type and policy association weight. The module preprocesses the semantic unit sequence, parses the spoken description part of each semantic unit, and extracts the key feature vector as the mapping input.

[0054] The module performs rule mapping strategy in parallel, and the rule mapping strategy is based on the pre-defined policy term mapping rule library, which contains the corresponding relationship between spoken expression and standard policy field. The rule mapping process matches the spoken entity with the pattern in the rule library through the pattern matching algorithm, and converts it directly into the standard policy index if the matching is successful. Rule mapping adopts a multi-level rule priority mechanism, which preferentially matches accurate rules and then matches fuzzy rules.

[0055] The module performs fuzzy set membership calculation strategy in parallel, and the fuzzy set membership calculation strategy is for numerical or degree type spoken description, such as "more" "higher" and other fuzzy quantifiers. The strategy constructs the membership function of fuzzy set, and the membership function defines the mapping relationship between spoken quantifier and standard policy numerical range. The calculation process evaluates the membership degree of spoken description belonging to each standard policy numerical interval through the membership function, and selects the interval with the highest membership degree as the mapping result.

[0056] The module performs collaborative filtering prediction strategy in parallel, and the collaborative filtering prediction strategy uses similar cases in historical mapping data for prediction. The strategy constructs a mapping collaborative filtering matrix based on user behavior data, which records the mapping relationship between historical spoken description and standard policy field. The prediction process calculates the similarity between the current spoken description and the historical cases, aggregates the mapping results of similar cases with weights, and generates the predicted standard policy index.

[0057] The module performs weighted fusion on the output of the three mapping strategies, and the weighted fusion calculates the comprehensive mapping result according to the initial weight of each strategy. The weight distribution is dynamically initialized based on the historical accuracy of the strategy, and the fusion process adopts a weighted voting mechanism to generate the preliminary standard policy field mapping.

[0058] The module receives the mapping confidence score fed back by the policy matching multi-round iteration optimization module, and the mapping confidence score reflects the accuracy of the last mapping result in policy matching. The score data is used as a supervision signal to adjust the weight of each mapping strategy. The adjustment process uses the gradient descent method, which takes the mapping confidence score as the loss function to calculate the weight adjustment gradient and iteratively updates the strategy weight value.

[0059] The module re-weights and fuses the mapping results according to the updated strategy weight to generate a final set of standard policy indicators. The indicator set is structured and converted into a fixed-dimensional structured query vector. The structured query vector contains information such as policy indicator type, indicator value, and confidence, and encodes the policy query condition in vector form. Finally, the module outputs the structured query vector to the policy matching multi-round iteration optimization module for policy matching calculation.

[0060] Specifically, the policy matching multi-round iteration optimization module embeds the structured query vector into the vector space of the policy knowledge graph and uses a graph neural network to calculate node similarity. The reinforcement learning framework uses the success rate of the application as a reward signal to optimize the query reconstruction strategy, outputs the sorted policy list to the interactive guidance system state feedback module, and feeds back the policy matching confidence to the semantic mapping adaptive standardization module.

[0061] The policy matching multi-round iteration optimization module receives the structured query vector transmitted by the semantic mapping adaptive standardization module, which contains information such as standard policy indicator type, indicator value, and confidence. The module first embeds the structured query vector into the vector space of the policy knowledge graph. The embedding process uses graph embedding algorithms such as Node2Vec or TransE to map the nodes and relationships in the policy knowledge graph into low-dimensional vector representations, forming points in the vector space. The embedding algorithm learns the semantic association between nodes based on random walk or translation mechanism, so that similar policy nodes are close in distance in the vector space, laying a foundation for subsequent similarity calculation.

[0062] The module uses a graph neural network to calculate node similarity, which uses a graph convolutional network or a graph attention network architecture. The number of network layers is dynamically configured according to the size of the policy knowledge graph. The graph neural network receives the embedded vector representation and aggregates neighbor node information through a message passing mechanism to update the feature representation of each node. The similarity calculation process uses cosine similarity or Euclidean distance measurement to compare the vector representation of the structured query vector with each policy node in the policy knowledge graph to generate a preliminary matching score. The multi-layer convolution operation of the graph neural network can capture high-order relationships between nodes, improving the accuracy of similarity calculation.

[0063] The module integrates a reinforcement learning framework that uses the success rate of policy declaration as a reward signal derived from historical declaration data or user feedback. The framework uses Q-learning or policy gradient algorithms, and the agent selects query reconstruction actions based on the current matching state, such as adding filter conditions or adjusting weights. The optimization process updates the policy network parameters based on the reward signal, allowing the query reconstruction policy to gradually converge to the optimal solution. Through multiple iterations, the reinforcement learning framework explores different query combinations and dynamically adjusts the matching strategy to maximize the success rate of declaration.

[0064] The module optimizes the matching process through a query reconstruction policy that dynamically adjusts the composition of the structured query vector based on the output of the reinforcement learning. The policy includes condition expansion, weight redistribution, or logical operator modification. The reconstructed query vector is re-embedded in the policy knowledge graph vector space, and the node similarity is recalculated. The optimization process is executed in a loop, and each iteration refines the query policy based on the matching score and reward signal, gradually narrowing down the policy candidate set and improving matching accuracy.

[0065] The module outputs a ranked policy list to the interactive guidance system state feedback module, with the list arranged in descending order of matching score and containing policy names, matching scores, and recommended reasons. The output process uses a Top-K retrieval algorithm to retain the top matching policies, ensuring the practicality and operability of the list. At the same time, the module calculates the matching confidence, which is based on the stability of the matching score and the historical accuracy rate, reflecting the reliability of the current matching.

[0066] The module feeds back the matching confidence to the semantic mapping adaptive standardization module, using the feedback data as a supervision signal to adjust the mapping strategy weights. The matching confidence affects the weight update of the semantic mapping module through gradient descent, forming a closed-loop optimization path. The feedback mechanism allows the semantic mapping process to adapt to the policy matching requirements, reducing the deviation of spoken language to standard policy mapping, and improving the overall performance of the system.

[0067] Specifically, the speech recognition-based government information intelligent processing system described in the present application generates a sequence of guidance questions based on a partially observable Markov decision process model through the interactive guidance system state feedback module; the user feedback data synchronously updates the acoustic model of the speech recognition dynamic dictionary update module, the entity recognition strategy of the natural language understanding policy element association module, and the mapping weight of the semantic mapping adaptive standardization module, forming a cross-module collaborative learning cycle.

[0068] The interactive guidance system state feedback module receives a ranked policy list from the policy matching multi-round iterative optimization module, which contains information such as policy name, matching score, and recommendation reason. The module assesses the uncertainty of the list state based on a partially observable Markov decision process model, which measures uncertainty through information entropy or confidence interval, reflecting the reliability of the policy matching result. The partially observable Markov decision process model models the system state as a hidden Markov chain, estimates the current state probability distribution through Bayesian filtering, and thus quantifies the uncertainty.

[0069] The module generates a guidance question sequence using the strategy function of the partially observable Markov decision process model, which selects the optimal action based on state uncertainty, i.e., generates a specific guidance question. The guidance question sequence is designed to clarify ambiguous points in policy matching or supplement missing information, such as asking for details of enterprise size or specific values of capital investment. The question generation process uses template matching or natural language generation technology to ensure that the question is relevant to the policy context.

[0070] The module presents the guidance question sequence through a user interaction interface and collects user voice or text feedback data. User feedback data contains answers to guidance questions, and the module preprocesses feedback data, such as converting voice feedback to text through speech recognition and tokenizing and vectorizing text feedback. The collection process records user response time and content integrity, providing a basis for subsequent analysis.

[0071] User feedback data is synchronously distributed to the speech recognition dynamic dictionary update module, the natural language understanding policy element association module, and the semantic mapping adaptive standardization module. The distribution mechanism is based on message queues or event-driven architecture to ensure real-time data transmission. The distribution strategy dynamically routes feedback data based on its type, such as sending voice data to the speech recognition module first and sending text data to the natural language understanding module.

[0072] After receiving user feedback data, the speech recognition dynamic dictionary update module updates the acoustic model parameters. The update process uses online learning algorithms such as stochastic gradient descent to adjust the weights of the acoustic feature extraction network using feedback data. Acoustic model updates reduce recognition errors caused by dialect or accent variations, improving speech recognition accuracy.

[0073] The natural language understanding policy element association module uses user feedback data to optimize entity recognition strategies. The optimization process adjusts the feature function weights of the conditional random field model through incremental learning, and the entity recognition strategy update is based on entity annotation information in the feedback data. Strategy optimization improves the accuracy of intent classification and entity extraction, making the semantic unit sequence more accurately reflect user needs.

[0074] The semantic mapping adaptive standardization module adjusts the mapping weight according to the user feedback data. The mapping weight adjustment adopts the gradient descent method, taking the mapping accuracy in the feedback data as the loss function, and dynamically updating the weight values of the rule mapping, fuzzy set membership degree calculation and collaborative filtering prediction strategy. The weight optimization makes the mapping from spoken language to standard policy more adaptive to the actual use scene.

[0075] The feedback process forms a cross-module collaborative learning cycle, and the user feedback data drives the continuous optimization of the parameters of each module. The collaborative learning cycle is realized through bidirectional data flow between modules, so that the system has the evolution ability of policy language model and knowledge model. The cycle process reduces the uncertainty of policy matching, improves the system interaction efficiency and resource matching accuracy.

[0076] Specifically, the government information intelligent processing system based on speech recognition provided by the application realizes multi-level adaptation through bidirectional data flow between modules: the recognition confidence of the speech recognition dynamic dictionary updating module triggers the recalibration of the natural language understanding policy element association module, the matching result of the policy matching multi-round iteration optimization module feeds back to optimize the strategy of the semantic mapping adaptive standardization module, and the user behavior data of the interactive guidance system state feedback module drives the dynamic dictionary updating and mapping adjustment, so that the system has the continuous evolution ability of policy language model and knowledge model.

[0077] The system realizes multi-level adaptation through bidirectional data flow between modules. The bidirectional data flow adopts message queue or event bus architecture to realize real-time transmission and synchronization of data between modules. The data flow includes forward data flow and reverse data flow. The forward data flow transmits voice signals, text sequences and query vectors, and the reverse data flow transmits confidence scores, matching results and user feedback data. This kind of bidirectional flow mechanism enables the system to dynamically adjust the parameters of each module and realize continuous optimization.

[0078] The speech recognition dynamic dictionary updating module calculates the recognition confidence, which is calculated based on the probability distribution output by the speech recognition model and the acoustic feature likelihood. The recognition confidence reflects the reliability of the speech recognition result. The module compares the confidence level with a threshold value. When the recognition confidence is lower than the preset threshold value, the module generates a recalibration signal, which is transmitted to the natural language understanding policy element association module through the reverse data flow.

[0079] After receiving the recalibration signal, the natural language understanding policy element association module starts the recalibration process. The recalibration process includes re-parsing the recognized text, adjusting the intent classification model parameters and updating the entity extraction strategy. The module uses incremental learning algorithms such as online stochastic gradient descent to adjust the attention weight of the Transformer architecture model and the feature function of the conditional random field model using the latest user feedback data. Recalibration improves the accuracy of intent recognition and entity extraction, and reduces the error of semantic unit sequence.

[0080] The policy matching multi-round iteration optimization module calculates the matching result, which is based on the similarity scores output by the graph neural network and the policy declaration success rate. The matching result reflects the effect of policy matching, and the module evaluates the matching quality through the reinforcement learning framework. The matching result is fed back to the semantic mapping adaptive standardization module through the reverse data flow as a supervision signal for optimizing the mapping strategy.

[0081] After receiving the matching result, the semantic mapping adaptive standardization module optimizes the mapping strategy. The optimization process uses the gradient descent method, taking the matching result as the loss function, to adjust the weights of the rule mapping, fuzzy set membership degree calculation, and collaborative filtering prediction strategy. The strategy optimization makes the mapping from spoken language to standard policy more in line with the actual policy requirements, improving the quality of the structured query vector.

[0082] The interactive guidance system state feedback module collects user behavior data, including the user's answers to guidance questions, interaction time, and selection preferences. The module captures behavior data through the user interface and performs preprocessing such as data cleaning and feature extraction. User behavior data is distributed to the speech recognition dynamic dictionary update module and the semantic mapping adaptive standardization module through the reverse data flow.

[0083] The speech recognition dynamic dictionary update module uses user behavior data to drive dynamic dictionary updates, including adjusting the decoding weights of acoustic models and language models. Based on high-frequency terms and pronunciation patterns in user behavior data, the module dynamically updates the dictionary word table and decoding strategy. The update reduces the impact of dialects and spoken language variants, improving speech recognition accuracy.

[0084] The semantic mapping adaptive standardization module uses user behavior data to drive mapping adjustments, including modifying mapping rules and updating membership function parameters. By analyzing mapping feedback in user behavior data, the module optimizes the conversion logic from spoken language to standard policy. The adjustment improves mapping accuracy and reduces semantic bias.

[0085] The adaptive process forms a closed-loop optimization cycle, and the bidirectional data flow enables each module to respond in real time to changes and continuously improve performance. Through multi-level adaptive mechanisms, the system gradually optimizes policy language models and knowledge models, improving the accuracy of policy matching and resource utilization efficiency, and achieving continuous evolution capabilities.

[0086] In a second aspect, the present application provides a method for intelligent processing of government information based on speech recognition, which is applied to the intelligent processing system of government information based on speech recognition as described above, and comprises the following steps: Step 1: Collecting the speech signal of the entrepreneur and performing noise reduction and endpoint detection to output a pure speech frame sequence. Step 2, receive the sequence of pure speech frames, generate recognized text through a speech recognition model, and optimize the recognition process using a dynamic dictionary update mechanism; Step 3, receive the recognized text, perform intent classification and entity extraction, and query the policy knowledge graph to add policy association weights to entities, output a sequence of semantic units with weights; Step 4, store the hypergraph structure of policy provisions, receive the sequence of semantic units for logical reasoning, and derive a set of policy conditions; Step 5, receive the sequence of semantic units with weights, convert spoken language entities into standard policy indicators through multi-strategy mapping, and receive feedback confidence to adaptively adjust the mapping strategy, output a structured query vector; Step 6, receive the structured query vector, calculate the matching degree on the policy knowledge graph, and output a ranked policy list through reinforcement learning-driven multi-round query reconstruction; Step 7, receive the ranked policy list, generate guide questions based on list state uncertainty, collect user feedback data, and use user feedback data to update the dynamic dictionary and semantic mapping strategy of the speech recognition process.

[0087] The voice recognition-based government information intelligent processing method first performs step 1, collects the entrepreneur's voice signal and performs noise reduction and endpoint detection, and outputs a sequence of pure speech frames. The voice collection process uses a microphone array to capture multi-channel voice signals, suppresses environmental noise through beamforming technology, and forms a directional beam to enhance the target voice through adaptive algorithm to calculate the delay and weighting coefficient of each microphone channel. The endpoint detection algorithm sets double-threshold thresholds based on short-time energy and short-time zero-crossing rate features to identify the starting point and ending point of the voice segment, and extracts the effective voice segment. The effective voice segment is processed by framing, divided into fixed-length short-time frame sequences, and the frames are overlapped to maintain temporal continuity. Each frame is windowed using a Hamming window to reduce spectral leakage, and finally a sequence of pure speech frames is output.

[0088] Step 2 receives the sequence of pure speech frames, generates recognized text through a speech recognition model, and optimizes the recognition process using a dynamic dictionary update mechanism. The speech recognition model uses an end-to-end architecture, the encoder converts the sequence of speech frames into a high-dimensional feature representation, and the decoder maps the feature sequence to a text sequence based on an attention mechanism. The dynamic dictionary update mechanism receives high-frequency policy terms from the natural language understanding module, adjusts the decoding weights of the acoustic model and language model, and preferentially matches policy-related vocabulary. The optimization process introduces an adversarial domain adaptation method, which distinguishes between standard speech and dialect domains through a domain discriminator, and the encoder learns to generate domain-invariant features to reduce the influence of dialect variations.

[0089] Step 3 receives the identified text, performs intent classification and entity extraction, and queries the policy knowledge graph to add policy association weights to the entities, and outputs a weighted semantic unit sequence. The intent classification uses a Transformer architecture model to calculate word vector dependencies through a self-attention mechanism, maps to an intent category space through a fully connected layer, and outputs an intent probability distribution through a Softmax function. Entity extraction combines a policy domain dictionary and a conditional random field model. The policy domain dictionary provides entity type mapping rules, and the conditional random field model decodes the optimal entity label sequence through the Viterbi algorithm. The policy element pre-association mechanism sends the entity list to the policy knowledge graph module to query the association strength, calculates the policy association weight based on node degree and path weight, and generates a weighted semantic unit sequence.

[0090] Step 4 stores the hypergraph structure of the policy provisions, receives the semantic unit sequence for logical reasoning, and derives the policy condition set. The hypergraph structure represents policy indicators and declaration conditions with nodes, and logical relationships with hyperedges. The rule-based reasoning engine uses a Datalog rule base to traverse the hypergraph nodes and hyperedges through a forward chain reasoning algorithm to derive implicit policy requirements. The module matches the entity list with the hypergraph nodes using a graph embedding algorithm to calculate the cosine similarity and identify relevant policy nodes. The hypergraph traversal algorithm searches for connected node sets along the hyperedge direction from the matching nodes, collects directly and indirectly associated policy conditions, and forms the policy condition set.

[0091] Step 5 receives the weighted semantic unit sequence, converts the spoken entity into a standard policy indicator through multi-strategy mapping, and receives feedback confidence to adaptively adjust the mapping strategy, and outputs a structured query vector. Multi-strategy mapping performs rule mapping, fuzzy set membership degree calculation, and collaborative filtering prediction in parallel. Rule mapping converts spoken entities into standard policy indicators based on a pre-defined mapping rule base through pattern matching. Fuzzy set membership degree calculation constructs a membership function to evaluate the membership degree of numerical spoken descriptions belonging to the standard policy numerical interval. Collaborative filtering prediction uses historical mapping data to calculate the similarity between the current spoken description and historical cases, and aggregates the mapping results with weights. The module adjusts the weights of each strategy using gradient descent based on the feedback confidence, and generates a structured query vector by weighting and fusing the mapping results.

[0092] Step 6 receives the structured query vector, calculates the matching degree on the policy knowledge graph, and outputs the ranked policy list through reinforcement learning driven multi-round query reconstruction. The module embeds the structured query vector into the vector space of the policy knowledge graph, calculates the node similarity using the graph neural network, aggregates the neighbor node information through the message passing mechanism, updates the node feature representation, and measures the matching degree using the cosine similarity. The reinforcement learning framework takes the success rate of the declaration as the reward signal, optimizes the query reconstruction strategy using the Q-learning algorithm, and dynamically adjusts the query vector composition. The multi-round iteration process gradually narrows down the policy candidate set according to the matching degree score and the reward signal, and outputs the ranked policy list.

[0093] Step 7 receives the ranked policy list, generates guiding questions based on the list state uncertainty, collects user feedback data, and uses the user feedback data to update the dynamic dictionary and semantic mapping strategy of the speech recognition process. The partially observable Markov decision process model evaluates the list state uncertainty, measures the uncertainty level through information entropy, and generates a sequence of guiding questions through a policy function. The module collects user feedback data through a user interaction interface, pre-processes it, and distributes it to the speech recognition dynamic dictionary update module and the semantic mapping adaptive normalization module. The speech recognition module updates the acoustic model parameters, and the semantic mapping module adjusts the mapping weights, forming a closed-loop optimization path and improving the system's adaptive ability.

[0094] The application obtains the voice signal of an entrepreneur through a voice acquisition preprocessing module, and performs noise reduction and endpoint detection, and outputs a pure voice frame sequence to a speech recognition dynamic dictionary updating module. The module uses an end-to-end speech recognition model and introduces an adversarial domain adaptation method to reduce the influence of dialect variation, and simultaneously uses a dynamic dictionary updating mechanism to adjust the decoding weight according to high-frequency policy terms, thereby improving the speech recognition accuracy. A natural language understanding policy element association module receives the recognized text to perform intent classification and entity extraction, and queries a policy knowledge graph logical reasoning network module to add a policy association weight to the entity, and outputs a weighted semantic unit sequence. The policy knowledge graph logical reasoning network module stores the hypergraph structure of the policy provisions, and derives a policy condition set through logical reasoning. A semantic mapping adaptive standardization module receives the weighted semantic unit sequence, and converts the spoken language entity into a standard policy index through multiple strategies such as rule mapping, fuzzy set membership degree calculation and collaborative filtering prediction, and adaptively adjusts the mapping strategy based on the feedback confidence of a policy matching multi-round iteration optimization module, and outputs a structured query vector. The policy matching multi-round iteration optimization module embeds the structured query vector into the vector space of the policy knowledge graph, calculates the matching degree using a graph neural network, and outputs a ranked policy list through reinforcement learning driven multi-round query reconstruction. An interactive guidance system state feedback module generates a guidance question according to the state uncertainty of the ranked policy list, collects user feedback data and distributes it to the speech recognition dynamic dictionary updating module and the semantic mapping adaptive standardization module, forms a closed-loop optimization path, so that the system can continuously adapt to colloquial expressions and accurately map policy requirements, thereby solving the semantic gap problem between the standardized expression of policy text and the colloquial cognition of entrepreneurs, and achieving efficient matching of policy resources.

[0095] Based on the technical problem of the semantic gap between the standardized expression of policy text and the colloquial cognition of entrepreneurs in existing policy recommendation technologies, the present application provides a specific embodiment of a government information intelligent processing system and method based on speech recognition. The system realizes the complete process from voice input to policy matching through the cooperative work of multiple modules, effectively bridges the semantic gap, and improves the matching accuracy.

[0096] The voice collection preprocessing module adopts a microphone array to collect the voice signal of the entrepreneur. The microphone array is arranged in a linear or annular geometry, and the time and space characteristics of the voice signal are captured through multi-channel synchronous sampling. The collected analog voice signal is converted into a digital signal through an analog-to-digital converter, and then a beamforming technique is applied to suppress environmental noise. The beamforming technique calculates the delay and weighting coefficient of each microphone channel through an adaptive algorithm, forms a directional beam main lobe pointing to the user's sound source direction, and enhances the target voice signal. An endpoint detection algorithm dynamically sets a double-threshold threshold based on short-time energy and short-time zero-crossing rate characteristics, identifies the starting point and ending point of the voice segment, and extracts the effective voice segment. The effective voice segment is processed by frame division, divided into a fixed-length short-time frame sequence, and the frames are overlapped to maintain time sequence continuity. Each frame is windowed using a Hamming window to reduce spectral leakage, and a pure voice frame sequence is output to the voice recognition dynamic dictionary update module.

[0097] The voice recognition dynamic dictionary update module receives the pure voice frame sequence, processes it using an end-to-end speech recognition model, and the encoder usually adopts a convolutional neural network combined with a recurrent neural network structure to extract the acoustic features of the voice frame by frame. The decoder maps the feature sequence to a text sequence based on an attention mechanism to generate the recognition text. The module introduces an adversarial domain adaptation method to reduce the influence of dialect variation. A domain discriminator is added during the training process to distinguish between standard speech and dialect domains, and the encoder learns to generate domain-invariant features through adversarial training. The dynamic dictionary update mechanism receives high-frequency policy terms fed back by the natural language understanding policy element association module. The high-frequency policy terms are obtained by counting the frequency of entity occurrence, and the decoding weights of the acoustic model and the language model are dynamically adjusted to preferentially match policy-related vocabulary, forming a reverse optimization path from semantic understanding to speech recognition.

[0098] The natural language understanding policy element association module receives the recognition text, first performs word segmentation and vectorization processing, and converts the text into a word vector sequence. The module uses a Transformer architecture model to process the word vector sequence, calculates the word vector dependency relationship through a self-attention mechanism, and generates a context-dependent encoding representation. The intent recognition subtask maps the encoder output to an intent category space through a fully connected layer and uses a Softmax function to calculate the probability distribution of policy queries, eligibility evaluation, and other government affairs scenario intents. The entity recognition subtask combines a policy domain dictionary and a conditional random field model. The policy domain dictionary contains policy terms and entity type mapping rules, and the conditional random field model decodes the optimal entity label sequence through the Viterbi algorithm to label policy element entities such as personnel size and capital investment. The policy element pre-association mechanism sends the entity list to the policy knowledge graph logical reasoning network module to query the association strength, calculates the policy association weight based on node degree and path weight, and generates a weighted semantic unit sequence to the semantic mapping adaptive standardization module.

[0099] The policy knowledge graph logical reasoning network module stores policy clauses using a hypergraph structure. The hypergraph nodes represent policy indicators and declaration condition entities, and the hyperedges represent complex logical relationships between multiple nodes, such as AND, OR, and NOT relationships. The module integrates a rule-based reasoning engine, which is based on a Datalog logical rule base and uses a forward chain reasoning algorithm to traverse the hypergraph nodes and hyperedges to derive implicit policy requirements from explicit policy conditions. The module receives entity query requests from the natural language understanding policy element association module, calculates the cosine similarity between entity vectors and node vectors using a graph embedding algorithm, and identifies relevant policy nodes. The hypergraph traversal algorithm searches for connected node sets along the hyperedge direction from the matching nodes, collects all associated policy conditions, and forms a policy condition set that is output to the policy matching multi-round iteration optimization module. The entity policy association weight is fed back to the natural language understanding policy element association module.

[0100] The semantic mapping adaptive standardization module receives a weighted semantic unit sequence and performs rule mapping, fuzzy set membership calculation, and collaborative filtering prediction strategies in parallel. Rule mapping converts spoken entities into standard policy indicators based on a pre-defined policy term mapping rule base using pattern matching. Fuzzy set membership calculation constructs a membership function for numerical spoken descriptions to evaluate the membership of the spoken description to a standard policy numerical interval. Collaborative filtering prediction calculates the similarity between the current spoken description and historical cases using historical mapping data, and aggregates the mapping results with weights. The module performs weighted fusion on the outputs of the three strategies, with initial weights dynamically set based on historical accuracy. After receiving the mapping confidence score from the policy matching multi-round iteration optimization module, the strategy weights are adjusted using the gradient descent method, and the structured query vector is re-weighted and fused to output to the policy matching multi-round iteration optimization module.

[0101] The policy matching multi-round iteration optimization module receives the structured query vector, first embeds the vector into the vector space of the policy knowledge graph, and uses graph embedding algorithms such as Node2Vec or TransE to map nodes and relationships into low-dimensional vector representations. The module uses graph neural networks to calculate node similarity, and uses graph convolutional networks or graph attention network architectures to aggregate neighbor node information, update node feature representations, and measure matching degree using cosine similarity. The reinforcement learning framework uses the declaration success rate as a reward signal, and uses Q-learning or policy gradient algorithms to optimize query reconstruction strategies and dynamically adjust query vector composition, such as adding filter conditions or modifying weights. The multi-round iteration process gradually narrows down the policy candidate set based on the matching score and reward signal, outputs a ranked policy list to the interactive guidance system state feedback module, and feeds back the policy matching confidence to the semantic mapping adaptive standardization module.

[0102] The interactive guidance system state feedback module receives the ranking policy list, evaluates the list state uncertainty based on a partially observable Markov decision process model, and the uncertainty calculation is measured by information entropy or confidence interval. The model models the system state as a hidden Markov chain, and estimates the state probability distribution through Bayesian filtering. The policy function generates a guidance question sequence based on the state uncertainty, and the question sequence is designed to clarify the ambiguity in policy matching, such as asking for details of enterprise size or specific values of capital investment. The module collects user feedback data through the user interaction interface, including voice or text answers, which are preprocessed and distributed to the speech recognition dynamic dictionary update module and the semantic mapping adaptive standardization module. The speech recognition dynamic dictionary update module updates the acoustic model parameters using the feedback data, and adjusts the acoustic feature extraction network weights using an online learning algorithm. The semantic mapping adaptive standardization module adjusts the mapping weights according to the feedback data, optimizes the strategy weights with the mapping accuracy as the loss function, and forms a cross-module collaborative learning cycle.

[0103] The voice recognition-based government information intelligent processing method realizes the step-by-step process through the above modules. Step 1: Collects voice signals and outputs pure voice frame sequences, Step 2: generates recognition text and optimizes the recognition process, Step 3: outputs a weighted semantic unit sequence, Step 4: derives a policy condition set, Step 5: outputs a structured query vector, Step 6: outputs a ranked policy list, and Step 7: generates guidance questions and updates system parameters. The system realizes multi-level adaptation through bidirectional data flow between modules, voice recognition confidence triggers natural language understanding recalibration, matching results feedback optimizes semantic mapping strategies, user behavior data drives dynamic dictionary updates and mapping adjustments, forming a closed-loop optimization path. The present application effectively solves the semantic gap problem between policy text and spoken language expression, and improves the accuracy and efficiency of policy resource matching.

Claims

1. A government information intelligent processing system based on speech recognition, characterized in that, include: The voice acquisition and preprocessing module acquires the entrepreneur's voice signal and performs noise reduction and endpoint detection, outputting a clean voice frame sequence. The speech recognition dynamic dictionary update module receives the clean speech frame sequence from the speech acquisition and preprocessing module, generates the recognition text through the speech recognition model, and optimizes the recognition process using the dynamic dictionary update mechanism. The Natural Language Understanding Policy Element Association Module receives the recognized text from the Speech Recognition Dynamic Dictionary Update Module, performs intent classification and entity extraction, queries the Policy Knowledge Graph Logical Reasoning Network Module to add policy association weights to entities, and outputs a weighted semantic unit sequence. The policy knowledge graph logical reasoning network module stores the hypergraph structure of policy clauses, receives the semantic unit sequence from the natural language understanding policy element association module, performs logical reasoning, and derives the policy condition set; The semantic mapping adaptive standardization module receives the weighted semantic unit sequence from the natural language understanding policy element association module, converts spoken entities into standard policy indicators through multi-strategy mapping, receives feedback confidence from the policy matching multi-round iterative optimization module to adaptively adjust the mapping strategy, and outputs a structured query vector. The policy matching multi-round iterative optimization module receives the structured query vector from the semantic mapping adaptive standardization module, calculates the matching degree on the policy knowledge graph, and drives multi-round query reconstruction through reinforcement learning to output a ranked policy list; The interactive guidance system status feedback module receives the sorted policy list from the policy matching multi-round iterative optimization module, generates guidance questions based on the uncertainty of the list status, collects user feedback data, and distributes the user feedback data to the speech recognition dynamic dictionary update module and the semantic mapping adaptive standardization module.

2. The government information intelligent processing system based on speech recognition according to claim 1, characterized in that, The speech acquisition and preprocessing module uses a microphone array to acquire speech signals, uses beamforming technology to suppress environmental noise, and uses an endpoint detection algorithm to extract valid speech segments, outputting the clean speech frame sequence to the speech recognition dynamic dictionary update module.

3. The government information intelligent processing system based on speech recognition according to claim 1, characterized in that, The speech recognition dynamic dictionary update module adopts an end-to-end speech recognition model and introduces an adversarial domain adaptive method to reduce the impact of dialect variations. The dynamic dictionary update mechanism receives high-frequency policy terms output by the natural language understanding policy element association module, adjusts the decoding weights of the acoustic model and language model, and forms a reverse optimization path from semantic understanding to speech recognition.

4. The intelligent government information processing system based on speech recognition according to claim 1, characterized in that, The natural language understanding policy element association module uses the Transformer architecture model for intent recognition and entity recognition. Entity recognition combines a policy domain dictionary and a conditional random field model. The policy element pre-association mechanism queries the policy knowledge graph logic reasoning network module during entity extraction to assign policy association weights to entities and generate the weighted semantic unit sequence, which is then transmitted to the semantic mapping adaptive standardization module.

5. The government information intelligent processing system based on speech recognition according to claim 1, characterized in that, The policy knowledge graph logic reasoning network module uses a hypergraph to represent the logical relationship between policy indicators and application conditions, integrates a rule reasoning engine to deduce implicit policy requirements, receives entity query requests from the natural language understanding policy element association module, outputs a set of policy conditions to the policy matching multi-round iterative optimization module, and feeds back the entity policy association weights to the natural language understanding policy element association module.

6. The intelligent government information processing system based on speech recognition according to claim 1, characterized in that, The semantic mapping adaptive standardization module performs rule mapping, fuzzy set membership calculation, and collaborative filtering prediction in parallel to map spoken descriptions into standard policy fields; it receives the mapping confidence score from the policy matching multi-round iterative optimization module, adjusts the policy weights through gradient descent, and outputs the structured query vector to the policy matching multi-round iterative optimization module.

7. The government information intelligent processing system based on speech recognition according to claim 1, characterized in that, The policy matching multi-round iterative optimization module embeds the structured query vector into the vector space of the policy knowledge graph and uses a graph neural network to calculate node similarity. The reinforcement learning framework uses the application success rate as a reward signal to optimize the query reconstruction strategy, outputs the sorted policy list to the interactive guidance system status feedback module, and feeds back the matching confidence to the semantic mapping adaptive standardization module.

8. The intelligent government information processing system based on speech recognition according to claim 1, characterized in that, The interactive guidance system's state feedback module generates a sequence of guidance questions based on a partially observable Markov decision process model; user feedback data synchronously updates the acoustic model of the speech recognition dynamic dictionary update module, the entity recognition strategy of the natural language understanding policy element association module, and the mapping weights of the semantic mapping adaptive standardization module, forming a cross-module collaborative learning loop.

9. The intelligent government information processing system based on speech recognition according to claim 1, characterized in that, The system achieves multi-level adaptation through bidirectional data flow between modules: the recognition confidence of the speech recognition dynamic dictionary update module triggers the recalibration of the natural language understanding policy element association module; the matching results of the policy matching multi-round iterative optimization module optimize the strategy of the semantic mapping adaptive standardization module; and the user behavior data of the interactive guidance system status feedback module drives the dynamic dictionary update and mapping adjustment, enabling the system to have the ability to continuously evolve the policy language model and knowledge model.

10. A method for intelligent processing of government information based on speech recognition, applied to the intelligent processing system for government information based on speech recognition as described in any one of claims 1 to 9, characterized in that, include: Step 1: Collect the entrepreneur's voice signal and perform noise reduction and endpoint detection to output a clean voice frame sequence; Step 2: Receive the clean speech frame sequence, generate the recognized text through the speech recognition model, and optimize the recognition process using a dynamic dictionary update mechanism; Step 3: Receive the identified text, perform intent classification and entity extraction, query the policy knowledge graph to add policy association weights to the entities, and output a weighted semantic unit sequence; Step 4: Store the hypergraph structure of policy terms, receive the semantic unit sequence, perform logical reasoning, and derive the policy condition set; Step 5: Receive the weighted semantic unit sequence, convert spoken entities into standard policy indicators through multi-strategy mapping, receive feedback confidence to adaptively adjust the mapping strategy, and output a structured query vector. Step 6: Receive the structured query vector, calculate the matching degree on the policy knowledge graph, and drive multi-round query reconstruction through reinforcement learning to output a ranked policy list; Step 7: Receive the sorting policy list, generate guiding questions based on the uncertainty of the list status, collect user feedback data, and use the user feedback data to update the dynamic dictionary and semantic mapping strategy of the speech recognition process.

Images