Intelligent questioning and answering method and device for enhancing large model government affairs by combining speech recognition with database

Abstract

The invention provides an intelligent government affair question-answering method and an intelligent government affair question-answering device for enhancing a large model by combining speech recognition with a database, and relates to the technical field of intelligent government affair question-answering, comprising the following steps: S1, a speech recognition system is responsible for accurately converting a user speech instruction into a written text as a primary link of information conversion; and S2, the user intention analysis is used for accurately converting a user voice instruction into a written text, and then the text input is input into the large language model. According to the scheme, the convenience of user interaction is finally improved through the voice recognition technology, the user is allowed to rapidly interact with the system through the natural language, manual input is not needed, the information acquisition speed is increased, services are easier to access, and secondly, the natural language processing capacity of a large model is improved, and the user experience is improved. Therefore, the system can understand complex query intentions and provide accurate and humanized answers, and the naturalness and accuracy of query are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of government affairs intelligent question answering, and in particular to a method and device for government affairs intelligent question answering with a large model combining speech recognition and database enhancement. Background Art

[0002] Intelligent government question and answer refers to the use of artificial intelligence and big data technology to provide users with 24-hour uninterrupted government information query and service guide services. It uses natural language processing and machine learning technology to improve the intelligence level of government services, improve service efficiency and user satisfaction.

[0003] In the relevant technologies, in the current government interactive question-and-answer framework, the need to integrate a speech recognition mechanism to translate the user's oral inquiries into written form is still rare. This process involves converting the speech signal into a text sequence through a speech recognition algorithm, and then using large language models (LLMs) processing technology to perform semantic analysis on the text to clarify the user's query request. The traditional database query process requires users to clarify the query requirements, back-end engineers to write SQL statements, front-end engineers to create query interfaces, and operation and maintenance engineers to launch functions before users can execute queries.

[0004] In the reasoning process of large language models (LLMs), there are limitations in specific domain knowledge. These models mainly rely on text data absorbed in the pre-training stage to perform reasoning and generation tasks. Therefore, although the model can provide relevant information, the comprehensiveness and accuracy of the content provided may be limited. The reason is that the model fails to have an in-depth grasp of specific subject knowledge and cannot perform real-time knowledge updates. In addition, in the government intelligent question-and-answer system, there are few examples that can simultaneously integrate large language models (LLMs) with advanced speech recognition technology, which leads to the single input method and limits the convenience and applicability of the system in specific environments.

[0005] Therefore, it is necessary to provide a method and device for intelligent government question answering based on speech recognition combined with database enhancement and large-model to solve the above-mentioned technical problems. Summary of the invention

[0006] The present invention provides a method and device for intelligent government question answering by combining speech recognition with a database-enhanced large model, which solves the technical problems of the limitations of specific domain knowledge in the reasoning process of a large language model (LLM) in related technologies and the inconvenience of simultaneously integrating a large language model (LLM) and advanced speech recognition technology. In order to solve the above technical problems, the speech recognition combined with database enhanced large model government affairs intelligent question answering method provided by the present invention comprises the following steps: Step S1: Voice Recognition As the primary link in information conversion, the speech recognition system is responsible for accurately converting user voice commands into written text; Step S2: User Intention Analysis After accurately converting the user's voice command into written text, the text input is input into a large language model with semantic understanding capabilities. Through this analysis process, the model can accurately identify and interpret the user's query intent; Step S3: Database query code generation After the user's intention is clarified, the input information is passed to the SQL generation module, which, with the assistance of a large language model and combined with relevant data in the knowledge base, preliminarily constructs a query language (SQL) expression for a specific database; Step S4: SQL code verification The initially generated SQL statements are then optimized by the SQL correction mechanism, which ensures the grammatical correctness, query efficiency and accuracy of the statements, thereby generating the final, optimized database query statements; Step S5: Data query visualization and intelligent answering Using the final generated SQL statement, the system queries the database and retrieves the required data, which is then used to generate charts to display the query results in a visual way, enhancing the readability and intuitiveness of the data; Finally, large-scale language models in the government sector are further involved, using their generative capabilities to conduct in-depth analysis of data, provide insights and predictions, and assist in the decision-making process.

[0007] Preferably, the LLM (Large Language Models) is a large language model, and its application scenario is a general scenario. In the NL2SQL scenario, since LLM is required to focus more on semantic understanding and SQL code logical reasoning, some fine-tuning means are needed to enable LLM to enhance its domain SQL code generation capability, and Baichuan2-13b, which has strong Chinese comprehension and SQL generation capabilities, is used as the base model for fine-tuning; An instruction dataset is constructed by collecting natural language to SQL (nl2sql) datasets and information extraction datasets. Subsequently, the dataset is used to fine-tune the large language model. After the training is completed, the model is tested and the accuracy of the generated data is manually verified. For inaccurate data found during the verification, fine-tuning training is performed again to improve the performance and accuracy of the model.

[0008] Preferably, the speech recognition can be divided into four processes from the perspective of technology: input stage, encoding, and decoding output; Using the powerful semantic understanding ability of the fine-tuned LLM, the user's intention is judged from the user's question, and the entity information (domain data) and time information involved in generating SQL are generated; The SQL generation is the core stage of NL2SQL, and its purpose is to generate SQL that conforms to natural language by repeatedly using the reasoning ability of LLM for key information in natural language, returning from the library to the table and finally to the field; Since the LLM described in the generative formula has defects such as unstable output format and unstable generated content, it is necessary to verify after LLM generates SQL. This link includes knowledge correction, static verification and dynamic verification.

[0009] A speech recognition combined with a database enhanced large-model government affairs intelligent question-answering device, including a portable box, a partition plate, a lifting mechanism and a clamping mechanism; The lifting mechanism comprises guide plates fixedly arranged on both sides of the inner wall of the portable case, the inner sides of the two guide plates are vertically slidably connected with sliding blocks, the front sides of the two sliding blocks are fixedly provided with gear plates, the tops of the two gear plates are fixedly provided with sliding rods, the surfaces of the two sliding rods are slidably connected with the inner side of the portable case, the surfaces of the two sliding rods and located on the inner side of the portable case are sleeved with springs, and the surfaces of the two sliding rods and located on the top of the portable case are slidably connected with handles; The clamping mechanism includes a protective frame fixedly arranged on the inner sides of both sides of the partition plate, the inner sides of the two protective frames are horizontally rotatably connected with a first threaded screw, the two ends of the first threaded screw are rotatably connected to the two sides of the inner wall of the portable box, the surfaces of both ends of the first threaded screw are fixedly provided with a first gear, the two first gears are respectively meshed with the two gear plates, the surface of the first threaded screw is threadedly connected with two first screw blocks, the bottoms of the two first screw blocks are fixedly provided with a clamping frame, the opposite sides of the two clamping frames are fixedly provided with a clamping pad, the top of the partition plate and on the side opposite to the two clamping pads are provided with a server body, the opposite sides of the two protective frames are fixedly provided with a first guide frame, the surfaces of the two first screw blocks are slidably connected to the inner sides of the first guide frame.

[0010] Preferably, a sliding groove cooperating with the vertical sliding of the sliding block is provided on the inner side of the guide plate, an annular groove cooperating with the sliding rod is provided on the inner sides of the portable box and the handle, a limiting ring structure is provided on the top of the two sliding rods, and a threaded groove cooperating with the first threaded screw is provided on the inner side of the two first screw blocks, and the two first screw blocks can be driven to move to the opposite side by rotating the first threaded screw.

[0011] Preferably, the inner sides of the bottoms of the two protective frames are horizontally rotatably connected with a spreading mechanism, the spreading mechanism comprises a second threaded screw horizontally rotatably connected to the inner sides of the two protective frames, the two ends of the second threaded screw are rotatably connected to the two sides of the inner wall of the portable box, the surfaces of the two ends of the second threaded screw are fixed with second gears, and the two second gears are respectively meshed with the two gear plates; The surface of the second threaded screw is threadedly connected to two second screw blocks, and the bottom of the two second screw blocks is fixedly provided with a first rotating seat, the inner sides of both sides of the portable box are rotatably connected to a rotating plate, and the opposite side of the two rotating plates is fixedly provided with a second rotating seat, the two first rotating seats and the two second rotating seats are divided into two groups, and the inner sides of each group of the first rotating seats and the second rotating seats are rotatably connected to a connecting rod, and the bottom of the partition plate is fixedly provided with a second guide frame, and the surfaces of the two second screw blocks are slidably connected to the inner side of the second guide frame.

[0012] Preferably, the inner sides of the two second screw blocks are provided with threaded grooves used in conjunction with the second threaded screw rod, and the rotation of the second threaded screw rod can drive the two second screw blocks to move to the opposite side. The inner sides of both sides of the portable box are provided with through grooves used in conjunction with the rotating plate for heat dissipation.

[0013] Preferably, a heat dissipation mechanism is fixedly provided at the bottom of the inner wall of the portable box, and the heat dissipation mechanism includes a protective cover fixedly provided at the bottom of the inner wall of the portable box, a drive motor is fixedly provided at the bottom of the inner wall of the portable box and located on the inner side of the protective cover, a fan blade is fixedly provided on the output shaft of the drive motor, a heat dissipation protection plate is fixedly provided on the inner side of the protective cover and located on the top of the fan blade, and a heat dissipation net is provided on the inner side of both sides of the partition plate.

[0014] Preferably, sliding brackets are fixedly provided on the front sides of both sides of the inner wall of the portable box, and a pull-out plate is vertically slidably connected to the inner sides of the two sliding brackets, an observation window is provided on the inner side of the pull-out plate, a through-hole plate is provided on the surface of the pull-out plate and at the top of the observation window, a through groove is provided on the inner side of the top of the pull-out plate, and two supporting legs are respectively fixedly provided on both sides of the bottom of the portable box.

[0015] Compared with related technologies, the speech recognition combined with database enhanced large model government affairs intelligent question answering method and device provided by the present invention has the following beneficial effects: This case improves the convenience of user interaction through voice recognition technology, allowing users to quickly interact with the system through natural language without manual input, which not only improves the speed of information acquisition, but also makes the service more accessible. Secondly, the natural language processing capability of the big model enables the system to understand complex query intentions and provide accurate and humanized answers, enhancing the naturalness and accuracy of the query; by combining with the database, the breadth and depth of the knowledge base are expanded, and a large amount of structured data can be accessed and utilized to provide more comprehensive and in-depth information. This combination also improves the quality of decision support, enabling the big model to conduct in-depth analysis of government data and provide data support and predictions for decision makers. At the same time, the intelligent question-and-answer system reduces dependence on database engineers, helps optimize resource allocation, and reduces operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0017] Figure 1 The best framework diagram of the speech recognition combined with database enhanced large model government affairs intelligent question answering method provided by the present invention; Figure 2 A flowchart of the fine-tuning process of the speech recognition combined with database enhanced large-model government affairs intelligent question-answering method provided by the present invention; Figure 3 A data visualization flow chart of the speech recognition combined with database enhanced large-model government affairs intelligent question-answering method provided by the present invention; Figure 4 A flowchart of the SQL generation process of the speech recognition combined with database enhanced large-model government affairs intelligent question-answering method provided by the present invention; Figure 5 The optimal structural diagram of the speech recognition combined with database enhanced large-model government affairs intelligent question-answering device provided by the present invention; Figure 6 for Figure 5 A schematic diagram of the structure of the rear view of the portable box shown; Figure 7 for Figure 5 A schematic diagram of the structure inside the portable box shown; Figure 8 for Figure 5 A schematic structural diagram of a cross-sectional view of the portable box shown; Fig. 9 for Figure 8A schematic diagram of the structure of the guide plate shown; Fig.10 for Figure 8 A schematic diagram of the structure of the clamping frame shown; Fig.11 for Figure 8 The handle is pulled upwards, and the first threaded screw is driven to rotate through the two first gears under the action of the sliding rod and the gear plate, so that the two clamping frames and the clamping pad move to the opposite side; Fig.12 A schematic diagram of the structure of the opening mechanism provided by the present invention; Fig.13 for Fig.12 The enlarged structural diagram of A shown; Fig.14 A schematic diagram of a state in which two sliding rods provided by the present invention respectively drive two gear plates to move upward, and under the action of the second gear and the second threaded screw, the two second screw blocks move to the opposite side, so that the two rotating plates rotate and close; Fig.15 This is a schematic structural diagram of the heat dissipation mechanism provided by the present invention.

[0018] Description of Figure Numbers: 1. Portable box; 2. Divider; 3. lifting mechanism; 31. guide plate; 32. sliding block; 33. gear plate; 34. sliding rod; 35. spring; 36. handle; 4. Clamping mechanism; 41. Protective frame; 42. First threaded screw; 43. First gear; 44. First screw block; 45. Clamping frame; 46. Clamping pad; 47. First guide frame; 5. Server body; 6. Spreading mechanism; 61. Second threaded screw; 62. Second gear; 63. Second screw block; 64. First rotating seat; 65. Rotating plate; 66. Second rotating seat; 67. Connecting rod; 68. Second guide frame; 7. heat dissipation mechanism; 71. protective cover; 72. drive motor; 73. fan blade; 74. heat dissipation protection plate; 75. heat dissipation net; 8. Sliding bracket; 9. Pull-out plate; 10. Observation window; 11. Through-hole plate; 12. Support leg.

[0019] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] The present invention provides a method and device for intelligent government affairs question answering by combining speech recognition with database enhancement of a large model.

[0022] First embodiment: See also Figures 1 to 4 , a method for intelligent government affairs question answering based on speech recognition combined with database enhancement and large model, including the following steps: Step S1: Voice Recognition As the primary link in information conversion, the speech recognition system is responsible for accurately converting user voice commands into written text; Step S2: User Intention Analysis After accurately converting the user's voice command into written text, the text input is input into a large language model with semantic understanding capabilities. Through this analysis process, the model can accurately identify and interpret the user's query intent; Step S3: Database query code generation After the user's intention is clarified, the input information is passed to the SQL generation module, which, with the assistance of a large language model and combined with relevant data in the knowledge base, preliminarily constructs a query language (SQL) expression for a specific database; Preferably, the knowledge base mainly cleans and pre-calculates the data in the data table to achieve the purpose of data enhancement. In some more complex scenarios (such as month-on-month and year-on-year), the operation logic itself is more complex. LLM also needs to perform these complex data processing when performing SQL logical reasoning, which can easily cause the generated SQL to fail static verification. In order to improve the accuracy of SQL generated by LLM, it is necessary to process the original data and first estimate whether the business scenario contains complex data operations and data accuracy issues involving table structures; Furthermore, for multi-dimensional data calculation problems, such as average, percentage, standard deviation, partition, sort, month-on-month, year-on-year, and trend data, these data can be modeled regularly, pre-calculated, and persisted in the database table. If the problem involves these calculation scenarios, LLM only needs to select these data model tables in the table selection stage to avoid the stability degradation caused by complex data calculations. Furthermore, when it comes to the problem of fields being enumerated values, if the fields in the data table are enumerated values, data other than the enumerated values ​​should not appear in the query conditions. By analyzing the distribution of field data, the data range can be clearly known. In the knowledge correction link of the SQL verification stage, it can be analyzed whether the SQL generated by LLM contains data outside the range, thereby improving the accuracy of SQL. Step S4: SQL code verification The initially generated SQL statements are then optimized by the SQL correction mechanism, which ensures the grammatical correctness, query efficiency and accuracy of the statements, thereby generating the final, optimized database query statements; Step S5: Data query visualization and intelligent answering Using the final generated SQL statement, the system queries the database and retrieves the required data, which is then used to generate charts to display the query results in a visual way, enhancing the readability and intuitiveness of the data; Preferably, the data of the SQL query generated by LLM is used for drawing charts, data analysis and summary, which includes the following steps: S1: Query data Execute the SQL generated by LLM in the domain database to obtain the data to be displayed in the chart; S1: Chart type determination Through LLM, we extract semantic information from user questions and analyze data to find the most suitable chart type for displaying data. S1: Draw a chart LLM is used to assemble the data into the vage-lite data format and render it on the page. Supported graphics include pie charts, line charts, bar charts, etc. S1: Data Analysis Use data, prompt words and the language summarization ability of large models to present data from different angles and styles to help users understand the data more comprehensively; Finally, large-scale language models in the government sector are further involved, using their generative capabilities to conduct in-depth analysis of data, provide insights and predictions, and assist in the decision-making process.

[0023] The LLM (Large Language Models) is a large language model, and its application scenario is a general scenario. In the NL2SQL scenario, since LLM needs to focus more on semantic understanding and SQL code logical reasoning, some fine-tuning methods are needed to enable LLM to enhance its domain SQL code generation capability. Baichuan2-13b, which has strong Chinese comprehension and SQL generation capabilities, is used as the base model for fine-tuning. An instruction dataset is constructed by collecting natural language to SQL (nl2sql) datasets and information extraction datasets. Subsequently, the dataset is used to fine-tune the large language model. After the training is completed, the model is tested and the accuracy of the generated data is manually verified. For inaccurate data found during the verification, fine-tuning training is performed again to improve the performance and accuracy of the model.

[0024] The speech recognition can be divided into four processes from the perspective of technology: input stage, encoding, and decoding output. Further, input stage: First, the original audio signal needs to be preprocessed to remove noise and improve the quality of the speech signal. Then, the audio signal is divided into small time segments, usually 20-30 milliseconds per frame. These frames are further processed to convert into multi-dimensional feature vectors that reflect the characteristics of human auditory perception. Commonly used feature extraction methods include Mel-frequency cepstral coefficients (MFCCs), which simulates the human ear's ability to perceive sounds of different frequencies; Further, the encoding stage: In this stage, the extracted multidimensional feature vector is converted into a set of state sequences. The state is an abstract concept in the acoustic model, representing a short acoustic event in the speech signal, which is usually smaller than a phoneme and is the basic unit of the phoneme. The recognition of the state depends on the acoustic model. The present invention uses a hidden Markov model (HMM); Further, the decoding stage: the state sequence is then combined to form phonemes. The decoding process involves mapping the state sequence to the phoneme sequence. This process is implemented using the joint probability of the acoustic model and the language model. The language model provides probabilistic information about the word sequence to help improve the accuracy of decoding; Further, the output stage: Finally, the phoneme sequence is converted into a word sequence, and then these words are organized into sentences. This process involves word boundary detection and sentence structure construction. The output text is a written representation of the original speech signal, realizing the conversion from speech to text; Using the powerful semantic understanding ability of the fine-tuned LLM, the user's intention is judged from the user's question, and the entity information (domain data) and time information involved in generating SQL are generated; Further, intent classification: facing various user questions, the user input is diverse, so we need to identify the intent to determine whether the user's question is applicable to the business scope of NL2SQL. For those specific intents such as querying data and generating charts, we can directly use the subsequent system functions to process them, thereby realizing automated data query and chart generation. However, for questions that do not fall within the scope of these specific intents, we need to take general actions to meet user needs and provide corresponding support. This process aims to achieve more efficient user interaction while ensuring that the system can meet diverse user needs and provide comprehensive solutions. Furthermore, entity information: This information is a data element closely related to the domain business. The system needs to undergo intent analysis to determine whether this entity information can meet the requirements of SQL generation. In this process, the system will identify the user's intent and determine whether the entity information provided by the user is suitable for the construction of SQL queries. This analysis ensures that the entity information used matches the specific SQL query task; Furthermore, time information: Most questions asked by users are presented in the form of spoken texts. In these texts, due to the influence of oral expression and context, the time information is often irregular or missing. In this case, it is necessary to rely on the ability of large language models to extract time information from user questions and standardize it into the standard time format in the database; The SQL generation is the core stage of NL2SQL, and its purpose is to generate SQL that conforms to natural language by repeatedly using the reasoning ability of LLM for key information in natural language, returning from the library to the table and finally to the field; Preferably, the scope is gradually refined by selecting libraries, tables, and columns during the process, which can not only avoid the problem of prompt length limitation in LLM, but also improve the accuracy of SQL generated by LLM; Furthermore, in the database selection stage, the business database scope involved in the expected SQL is determined by the marked type, and the scope of subsequent LLM table selection is narrowed; Furthermore, in the table selection stage, the table name and table purpose of the selected business database are filled into the table selection prompt template and sent to the LLM together with the natural language question. The LLM infers the table range for implementing the natural language SQL based on the obtained data. Next, the column selection stage: In this stage, the selected table structure information, questions, and table selection prompt template are combined and filled to obtain the necessary fields for the table selection prompt to be handed over to LLM to judge the SQL; Further, construct SQL: first, pre-process the time in the question, extract the time information in the spoken language and context into a standard time format, then fill the query condition field with data based on the entity information extracted in the previous step, and finally give this information to LLM to construct SQL; Since the LLM described in the generative formula has defects such as unstable output format and unstable generated content, it is necessary to verify after LLM generates SQL, which includes knowledge correction, static verification and dynamic verification; Furthermore, knowledge correction can be performed by enhancing or analyzing data in the data table to correct incorrect data in SQL. For example, the gender field in the data table is an enumeration value of "male, female". If other data appears in this field in SQL, the SQL generated by LLM is incorrect. If the data is a synonym or expansion of "male, female", the data can be corrected; if the data is a text that is completely unrelated to the enumeration value, LLM needs to be notified and used for self-correction. Next, static verification, i.e. the syntax analysis phase, is used to determine whether the SQL meets the standards through syntax analysis. Tools such as ANTLR AST and SQLParser can be used for processing. Further, dynamic verification includes semantic analyzer, access control and permission check. Semantic Analyzer: After syntax verification, the semantic analyzer can be used to ensure that the tables, columns, relations and other elements in the SQL query statement are valid and conform to the database schema, which can help detect potential semantic errors; Access control and permission check: Database management systems usually verify the legitimacy of SQL queries based on the user's access rights. If the user does not have permission to execute a query, then the query will be considered irregular.

[0025] In this embodiment, by providing database access capabilities to large language models, external database knowledge is introduced to enhance the model's reasoning capabilities, especially in the government affairs field, to ensure that the model has sufficient knowledge reserves as support when reasoning about relevant domain knowledge. Through this knowledge enhancement, this case improves the model's deep understanding ability when dealing with specific issues in the government affairs field, thereby improving the accuracy and reliability of the reasoning results; Integrating a speech recognition module into the existing question-and-answer system to achieve a multimodal input mechanism not only enriches the input dimension of the system, but also significantly improves the convenience and user interaction experience of the government intelligent question-and-answer system in various work scenarios. By introducing speech recognition technology, the system can receive and process verbal language input, providing users with a more natural and direct way of communication, especially in situations where fast or remote access to government information is required; This case uses the advanced capabilities of large language models in natural language understanding (NLU) and structured query statement generation to respond to users' diverse query needs in natural language. Through natural language to SQL (NL2SQL) technology, the system can receive and process user queries entered in voice or text to achieve efficient interaction with the database. This technology converts users' natural language queries into precise database query language, retrieves information directly from structured data sources, and quickly generates accurate responses based on query results. This method not only lowers the professional skill threshold for statistical queries, but also greatly reduces the manpower burden and improves overall work efficiency.

[0026] Second embodiment: See also Figures 5 to 14 , a speech recognition combined with database enhanced large-model government affairs intelligent question-answering device, comprising a portable box 1, a partition plate 2, a lifting mechanism 3 and a clamping mechanism 4; The lifting mechanism 3 includes guide plates 31 fixedly arranged on both sides of the inner wall of the portable case 1, the inner sides of the two guide plates 31 are vertically slidably connected with sliding blocks 32, the front sides of the two sliding blocks 32 are fixedly arranged with gear plates 33, the tops of the two gear plates 33 are fixedly arranged with sliding rods 34, the surfaces of the two sliding rods 34 are slidably connected with the inner side of the portable case 1, the surfaces of the two sliding rods 34 and located on the inner side of the portable case 1 are sleeved with springs 35, and the surfaces of the two sliding rods 34 and located on the top of the portable case 1 are slidably connected with handles 36; Please combine Figure 8 and Fig.11 : Pull the handle 36 upward, and the movement of the handle 36 drives the two sliding rods 34 to move upward, so that the two springs 35 are contracted. At the same time, the two sliding rods 34 drive the two gear plates 33 at the bottom to move upward during the movement, and the movement of the two gear plates 33 drives the two sliding blocks 32 to move upward on the inner sides of the two guide plates 31 respectively. When the two sliding blocks 32 move to the top of the inner side of the guide plate 31, the portable case 1 can be lifted up through the handle 36; Furthermore, the portable case 1 is placed on the ground, and then the handle 36 is released. Under the action of the two springs 35, the two sliding rods 34 slide downward on the inner side of the portable case 1. The movement of the two sliding rods 34 drives the two gear plates 33 to move downward. During the downward movement of the two gear plates 33, the sliding blocks 32 behind them are respectively driven to move downward on the inner side of the guide plate 31, until the bottoms of the two sliding blocks 32 slide to the bottom of the inner sides of the two guide plates 31 respectively. The clamping mechanism 4 includes a protective frame 41 fixedly arranged on the inner sides of both sides of the partition plate 2, and the inner sides of the two protective frames 41 are horizontally rotatably connected with a first threaded screw 42, and the two ends of the first threaded screw 42 are rotatably connected to the two sides of the inner wall of the portable box 1, and the surfaces of both ends of the first threaded screw 42 are fixedly provided with a first gear 43, and the two first gears 43 are respectively meshed with the two gear plates 33, and the surface of the first threaded screw 42 is threadedly connected with two first screw blocks 44, and the bottoms of the two first screw blocks 44 are fixedly provided with a clamping frame 45, and the opposite sides of the two clamping frames 45 are fixedly provided with a clamping pad 46, and the top of the partition plate 2 and the side opposite to the two clamping pads 46 are provided with a server body 5, and the opposite sides of the two protective frames 41 are fixedly provided with a first guide frame 47, and the surfaces of the two first screw blocks 44 are slidably connected with the inner sides of the first guide frame 47; Please combine Figures 8 to 11 : When the handle 36 is pulled upward and the two gear plates 33 move upward, the movement of the two gear plates 33 drives the two first gears 43 to rotate, and the rotation of the two first gears 43 drives the first threaded screw 42 to rotate, and the rotation of the first threaded screw 42 drives the two first screw blocks 44 to move to the opposite side, and the two first screw blocks 44 move to the opposite side. When the handle 36 moves upward to the top (at this time, the top of the sliding block 32 is close to the top of the inner side of the guide plate 31), the two clamping pads 46 clamp the server body 5; Furthermore, when the handle 36 is released and the two gear plates 33 are moved downward by the elastic force of the two springs 35, the movement of the two gear plates 33 drives the two first gears 43 to rotate in the opposite direction, and the opposite rotation of the two first gears 43 drives the first threaded screw 42 to rotate in the opposite direction, so that the two first screw blocks 44 move away from each other, and the movement of the two first screw blocks 44 drives the two clamping frames 45 and the clamping pads 46 to move away from each other, thereby canceling the clamping of the server body 5.

[0027] The inner side of the guide plate 31 is provided with a sliding groove for cooperating with the vertical sliding of the sliding block 32, the inner sides of the portable box 1 and the handle 36 are provided with annular grooves for cooperating with the sliding rod 34, the top ends of the two sliding rods 34 are provided with limiting ring structures, and the inner sides of the two first screw blocks 44 are provided with threaded grooves for cooperating with the first threaded screw 42, and the two first screw blocks 44 can be driven to move to the opposite side by rotating the first threaded screw 42.

[0028] The inner side of the bottom of the two protective frames 41 is horizontally rotatably connected with a spreading mechanism 6, and the spreading mechanism 6 includes a second threaded screw 61 horizontally rotatably connected to the inner side of the two protective frames 41, and the two ends of the second threaded screw 61 are rotatably connected to the two sides of the inner wall of the portable box 1, and the surfaces of the two ends of the second threaded screw 61 are fixed with second gears 62, and the two second gears 62 are respectively meshed with the two gear plates 33; The surface of the second threaded screw 61 is threadedly connected with two second screw blocks 63, and the bottoms of the two second screw blocks 63 are fixedly provided with a first rotating seat 64. The inner sides of both sides of the portable box 1 are rotatably connected with a rotating plate 65, and the opposite sides of the two rotating plates 65 are fixedly provided with a second rotating seat 66. The two first rotating seats 64 and the two second rotating seats 66 are divided into two groups, and the inner sides of each group of the first rotating seats 64 and the second rotating seats 66 are rotatably connected with a connecting rod 67. The bottom of the partition plate 2 is fixedly provided with a second guide frame 68, and the surfaces of the two second screw blocks 63 are slidably connected with the inner sides of the second guide frames 68. Please combine Figure 7 , Figure 8 , Figure 12 to Figure 14 : When the handle 36 is pulled upward and the two gear plates 33 move upward, the movement of the two gear plates 33 drives the two second gears 62 to rotate, and the rotation of the two second gears 62 drives the second threaded screw 61 to rotate, and the rotation of the second threaded screw 61 drives the two second screw blocks 63 to move to the opposite side. During the movement, the two second screw blocks 63 drive the two first rotating seats 64 to move to the opposite side, and under the action of the two connecting rods 67 and the second rotating seat 66, the two rotating plates 65 are driven to rotate to the opposite side. When the top of the sliding block 32 is close to the top of the inner side of the guide plate 31, the two rotating plates 65 rotate to make the portable case 1 present a closed state; Furthermore, when the handle 36 is released and the two gear plates 33 are moved downward by the elastic force of the two springs 35, the movement of the two gear plates 33 drives the two second gears 62 to rotate in the opposite direction, and the rotation of the two second gears 62 drives the second threaded screw 61 to rotate in the opposite direction, and the rotation of the second threaded screw 61 drives the two second screw blocks 63 to move toward the separated side. During the movement, the two second screw blocks 63 drive the two first rotating seats 64 to move toward the separated side, and under the action of the two connecting rods 67 and the second rotating seats 66, the two rotating plates 65 are driven to rotate toward the separated side. When the bottom of the sliding block 32 is close to the bottom of the inner side of the guide plate 31, the two rotating plates 65 rotate to make the two sides of the portable case 1 appear in an open state.

[0029] The inner sides of the two second screw blocks 63 are provided with thread grooves used in conjunction with the second threaded screw 61. The rotation of the second threaded screw 61 can drive the two second screw blocks 63 to move to the opposite side. The inner sides of both sides of the portable box 1 are provided with through grooves used in conjunction with the rotating plate 65 for heat dissipation.

[0030] In this embodiment, by pulling the handle 36 upward, the gear plate 33 is driven to move upward, and the movement of the gear plate 33 drives the two first gears 43 and the second gear 62 to rotate. The rotation of the two first gears 43 drives the first threaded screw 42 to rotate. The rotation of the first threaded screw 42 drives the two first screw blocks 44 to move to the opposite side, thereby causing the two clamping frames 45 and the clamping pad 46 to move to the opposite side. When the two second gears 62 rotate, the second threaded screw 61 is driven to rotate, thereby driving the two second screw blocks 63 and the two first rotating seats 64 to move to the opposite side. 7 and the second rotating seat 66, the two rotating plates 65 are driven to rotate to the opposite side. When the top of the sliding block 32 is close to the top of the inner side of the guide plate 31, the two clamping pads 46 are used to clamp and fix the server body 5, and the portable case 1 is in a closed state, which is convenient for transporting the server body 5 without spending time to fix the server body 5 separately. When the server body 5 is placed on the ground, the two gear plates 33 are actively moved downward under the action of the spring 35, thereby canceling the clamping of the server body 5 and causing the two rotating plates 65 to rotate outward, which is convenient for the use and heat dissipation of the server body 5.

[0031] Third embodiment: See also Figures 5 to 8 and Fig.15 A heat dissipation mechanism 7 is fixedly arranged at the bottom of the inner wall of the portable case 1, and the heat dissipation mechanism 7 includes a protective cover 71 fixedly arranged at the bottom of the inner wall of the portable case 1, a driving motor 72 is fixedly arranged at the bottom of the inner wall of the portable case 1 and located inside the protective cover 71, a fan blade 73 is fixedly arranged on the output shaft of the driving motor 72, a heat dissipation protection plate 74 is fixedly arranged inside the protective cover 71 and located on the top of the fan blade 73, and a heat dissipation net 75 is arranged on the inner side of both sides of the partition plate 2; Please combine Figure 8 and Fig.15 : When the portable box 1 is placed on the ground and the server body 5 needs to be used, the drive motor 72 is started, and the fan blades 73 are driven to rotate by the rotation of the drive motor 72. The wind force generated by the rotation of the fan blades 73 will be exchanged with the gas on the top of the partition plate 2 through the heat dissipation net 75. After the exchange, the gas with heat will be dissipated from the gap formed by the two rotating plates 65 rotating outwards, thereby avoiding the problem of overheating of the server body 5 after long-term use.

[0032] The front sides of both sides of the inner wall of the portable case 1 are fixedly provided with sliding brackets 8, and the inner sides of the two sliding brackets 8 are vertically slidably connected with pull-out plates 9, and the inner sides of the pull-out plates 9 are provided with observation windows 10, and the surface of the pull-out plates 9 and the top of the observation windows 10 are provided with through-hole plates 11, and the inner side of the top of the pull-out plates 9 is provided with through grooves, and two supporting legs 12 are respectively fixedly provided on both sides of the bottom of the portable case 1; Preferably, when the device is in use, the pull-out plate 9 is pulled upward to start up and debug, and the pull-out plate 9 is closed after debugging. When the user uses it, the sound will be transmitted to the portable case 1 through the through-hole plate 11, and then the server body 5 is used to answer the user's questions. After the answer, the user can see the answer through the observation window 10, and the server body 5 can also be taken out for use separately. However, this method lacks the protection of the portable case 1, and the second method is not recommended.

[0033] In this embodiment, when the server body 5 is in use, the drive motor 72 is started, and the fan blades 73 are driven to rotate by the rotation of the drive motor 72. The wind force generated by the rotation of the fan blades 73 will be exchanged with the gas on the top of the partition plate 2 through the heat dissipation net 75. After the exchange, the gas with heat will be dissipated from the gap formed by the two rotating plates 65 rotating outward, thereby avoiding the problem of overheating of the server body 5 after long-term use.

[0034] Please refer to Figures 1 to 15 The working principle of the speech recognition combined with database enhanced large model government affairs intelligent question answering method and device provided by the present invention is as follows: Step S1, when transporting the device, pull the handle 36 upward, and the movement of the handle 36 drives the two sliding rods 34 to move upward, so that the two springs 35 are contracted. At the same time, the two sliding rods 34 drive the two gear plates 33 at the bottom to move upward during the movement, and the movement of the two gear plates 33 drives the two sliding blocks 32 to move upward on the inner sides of the two guide plates 31 respectively. When the two sliding blocks 32 move to the top of the inner side of the guide plate 31, the portable case 1 can be lifted up by the handle 36; Step S2, when the two gear plates 33 move upward, the two first gears 43 and the second gear 62 are driven to rotate at the same time, and the rotation of the two first gears 43 drives the first screw rod 42 to rotate, and the rotation of the first screw rod 42 drives the two first screw blocks 44 to move to the opposite side, and the two first screw blocks 44 respectively drive the two clamping frames 45 and the clamping pads 46 to move to the opposite side during the movement process, and when the handle 36 moves upward to the top (at this time, the top of the sliding block 32 is close to the top of the inner side of the guide plate 31), the two clamping pads 46 clamp the server body 5; The rotation of the two second gears 62 will drive the second screw rod 61 to rotate, and the rotation of the second screw rod 61 will drive the two second screw blocks 63 to move to the opposite side. The two second screw blocks 63 will drive the two first rotating seats 64 to move to the opposite side during the movement. Under the action of the two connecting rods 67 and the second rotating seat 66, the two rotating plates 65 will be driven to rotate to the opposite side. When the top of the sliding block 32 is close to the top of the inner side of the guide plate 31, the two rotating plates 65 will rotate to make the portable box 1 present a closed state; Step S3, when the portable case 1 is placed on the ground and the server body 5 needs to be used, after the handle 36 is released, the expansion of the spring 35 drives the two sliding rods 34 and the gear plate 33 to move downward, and the movement of the two gear plates 33 drives the two first gears 43 to rotate in the opposite direction, and the opposite rotation of the two first gears 43 drives the first threaded screw 42 to rotate in the opposite direction, so that the two first screw blocks 44 move away from each other, and the movement of the two first screw blocks 44 drives the two clamping frames 45 and the clamping pad 46 to move away from each other, thereby canceling the clamping of the server body 5; When the two gear plates 33 move downward, they will simultaneously drive the two second gears 62 to rotate in the opposite direction. The rotation of the two second gears 62 will drive the second screw rod 61 to rotate in the opposite direction. The rotation of the second screw rod 61 will drive the two second screw blocks 63 to move toward the separated side. During the movement, the two second screw blocks 63 will drive the two first rotating seats 64 to move toward the separated side. Under the action of the two connecting rods 67 and the second rotating seat 66, the two rotating plates 65 will be driven to rotate toward the separated side. When the bottom of the sliding block 32 is close to the bottom of the inner side of the guide plate 31 (that is, the original position of the handle 36), the two rotating plates 65 will rotate to open both sides of the portable case 1. Step S4, when the device is in use, the pull-out plate 9 is pulled upward to start and debug, and the pull-out plate 9 is closed after debugging. When the user uses it, the sound is transmitted to the portable box 1 through the through-hole plate 11, and then the server body 5 is used to answer the user's question. After the answer, the user can see the answer through the observation window 10, and the server body 5 can also be taken out separately for use; When the server body 5 is in use, the drive motor 72 is started, and the fan blades 73 are driven to rotate by the rotation of the drive motor 72. The wind force generated by the rotation of the fan blades 73 will be exchanged with the gas on the top of the partition plate 2 through the heat dissipation net 75. After the exchange, the gas with heat will be dissipated from the gap formed by the two rotating plates 65 rotating outwards, thereby avoiding the problem of overheating of the server body 5 after long-term use; Step S5, when the server body 5 is in use, the speech recognition system, as the primary link of information conversion, is responsible for accurately converting the user's voice command into written text. Subsequently, the text input is input into a large language model, which has semantic understanding capabilities. Through this analysis process, the model can accurately identify and parse the user's query intention; After the user's intention is clarified, the input information is passed to the SQL generation module, which, with the assistance of a large language model and combined with relevant data in the knowledge base, preliminarily constructs a query language (SQL) expression for a specific database. The preliminarily generated SQL statement is then optimized by the SQL correction mechanism, which ensures the grammatical correctness, query efficiency and accuracy of the statement, thereby generating the final, optimized database query statement; Using the final generated SQL statement, the system queries the database and retrieves the required data. This data is then used to generate charts to display the query results in a visual way, enhancing the readability and intuitiveness of the data. Finally, large-scale language models in the government field are further involved, using their generation capabilities to conduct in-depth analysis of the data, provide insights and predictions, and assist in the decision-making process.

[0035] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. All equivalent structural changes made by using the contents of the present invention specification and drawings under the concept of the present invention, or directly / indirectly applied in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. A method for intelligent government affairs question answering based on speech recognition combined with database enhancement and large model, characterized in that: The following steps are involved: Step S1: Voice Recognition As the primary link in information conversion, the speech recognition system is responsible for accurately converting user voice commands into written text; Step S2: User Intention Analysis After accurately converting the user's voice command into written text, the text input is input into a large language model with semantic understanding capabilities. Through this analysis process, the model can accurately identify and interpret the user's query intent; Step S3: Database query code generation After the user's intention is clarified, the input information is passed to the SQL generation module, which, with the assistance of a large language model and combined with relevant data in the knowledge base, preliminarily constructs a query language (SQL) expression for a specific database; Step S4: SQL code verification The initially generated SQL statements are then optimized by the SQL correction mechanism, which ensures the grammatical correctness, query efficiency and accuracy of the statements, thereby generating the final, optimized database query statements; Step S5: Data query visualization and intelligent answering Using the final generated SQL statement, the system queries the database and retrieves the required data, which is then used to generate charts to display the query results in a visual way, enhancing the readability and intuitiveness of the data; Finally, large-scale language models in the government sector are further involved, using their generative capabilities to conduct in-depth analysis of data, provide insights and predictions, and assist in the decision-making process.

2. The method for government affairs intelligent question answering based on speech recognition combined with database enhancement of large model according to claim 1 is characterized in that: The LLM (Large Language Models) is a large language model, and its application scenario is a general scenario. In the NL2SQL scenario, since LLM needs to focus more on semantic understanding and SQL code logical reasoning, some fine-tuning methods are needed to enable LLM to enhance its domain SQL code generation capability. Baichuan2-13b, which has strong Chinese comprehension and SQL generation capabilities, is used as the base model for fine-tuning. An instruction dataset is constructed by collecting natural language to SQL (nl2sql) datasets and information extraction datasets. Subsequently, the dataset is used to fine-tune the large language model. After the training is completed, the model is tested and the accuracy of the generated data is manually verified. For inaccurate data found during the verification, fine-tuning training is performed again to improve the performance and accuracy of the model.

3. The method for intelligent government affairs question answering based on speech recognition combined with database enhancement of large model according to claim 2 is characterized in that: The speech recognition can be divided into four processes from the perspective of technology: input stage, encoding, and decoding output. Using the powerful semantic understanding ability of the fine-tuned LLM, the user's intention is judged from the user's question, and the entity information (domain data) and time information involved in generating SQL are generated; The SQL generation is the core stage of NL2SQL, and its purpose is to generate SQL that conforms to natural language by repeatedly using the reasoning ability of LLM for key information in natural language, returning from the library to the table and finally to the field; Since the LLM described in the generative formula has defects such as unstable output format and unstable generated content, it is necessary to verify after LLM generates SQL. This link includes knowledge correction, static verification and dynamic verification.

4. A speech recognition combined with database enhanced large-model government affairs intelligent question-answering device, characterized in that: The intelligent question-answering device is used in the intelligent question-answering method according to any one of claims 1 to 3, comprising a portable case, a partition plate, a lifting mechanism and a clamping mechanism; The lifting mechanism comprises guide plates fixedly arranged on both sides of the inner wall of the portable case, the inner sides of the two guide plates are vertically slidably connected with sliding blocks, the front sides of the two sliding blocks are fixedly provided with gear plates, the tops of the two gear plates are fixedly provided with sliding rods, the surfaces of the two sliding rods are slidably connected with the inner side of the portable case, the surfaces of the two sliding rods and located on the inner side of the portable case are sleeved with springs, and the surfaces of the two sliding rods and located on the top of the portable case are slidably connected with handles; The clamping mechanism includes a protective frame fixedly arranged on the inner sides of both sides of the partition plate, the inner sides of the two protective frames are horizontally rotatably connected with a first threaded screw, the two ends of the first threaded screw are rotatably connected to the two sides of the inner wall of the portable box, the surfaces of both ends of the first threaded screw are fixedly provided with a first gear, the two first gears are respectively meshed with the two gear plates, the surface of the first threaded screw is threadedly connected with two first screw blocks, the bottoms of the two first screw blocks are fixedly provided with a clamping frame, the opposite sides of the two clamping frames are fixedly provided with a clamping pad, the top of the partition plate and on the side opposite to the two clamping pads are provided with a server body, the opposite sides of the two protective frames are fixedly provided with a first guide frame, the surfaces of the two first screw blocks are slidably connected to the inner sides of the first guide frame.

5. The speech recognition combined with database enhanced large model government affairs intelligent question and answer device according to claim 4 is characterized in that: The inner side of the guide plate is provided with a sliding groove for cooperating with the vertical sliding of the sliding block, the inner sides of the portable box and the handle are provided with annular grooves for cooperating with the sliding rod, the top ends of the two sliding rods are provided with limiting ring structures, and the inner sides of the two first screw blocks are provided with threaded grooves for cooperating with the first threaded screw rod, and the two first screw blocks can be driven to move to the opposite side by rotating the first threaded screw rod.

6. The speech recognition combined with database enhanced large model government affairs intelligent question and answer device according to claim 4 is characterized in that: The inner sides of the bottoms of the two protective frames are horizontally rotatably connected with a spreading mechanism, the spreading mechanism comprises a second threaded screw horizontally rotatably connected to the inner sides of the two protective frames, the two ends of the second threaded screw are rotatably connected to the two sides of the inner wall of the portable box, the surfaces of the two ends of the second threaded screw are fixed with second gears, and the two second gears are respectively meshed with the two gear plates; The surface of the second threaded screw is threadedly connected to two second screw blocks, and the bottom of the two second screw blocks is fixedly provided with a first rotating seat, the inner sides of both sides of the portable box are rotatably connected to a rotating plate, and the opposite side of the two rotating plates is fixedly provided with a second rotating seat, the two first rotating seats and the two second rotating seats are divided into two groups, and the inner sides of each group of the first rotating seats and the second rotating seats are rotatably connected to a connecting rod, and the bottom of the partition plate is fixedly provided with a second guide frame, and the surfaces of the two second screw blocks are slidably connected to the inner side of the second guide frame.

7. The speech recognition combined with database enhanced large model government affairs intelligent question and answer device according to claim 6 is characterized in that: The inner sides of the two second screw blocks are provided with thread grooves used in conjunction with the second threaded screw rod. The rotation of the second threaded screw rod can drive the two second screw blocks to move to the opposite side. The inner sides of both sides of the portable box are provided with through grooves used in conjunction with the rotating plate for heat dissipation.

8. The speech recognition combined with database enhanced large model government affairs intelligent question and answer device according to claim 4 is characterized in that: A heat dissipation mechanism is fixedly provided at the bottom of the inner wall of the portable box, and the heat dissipation mechanism includes a protective cover fixedly provided at the bottom of the inner wall of the portable box, a drive motor is fixedly provided at the bottom of the inner wall of the portable box and on the inner side of the protective cover, a fan blade is fixedly provided on the output shaft of the drive motor, a heat dissipation protection plate is fixedly provided on the inner side of the protective cover and on the top of the fan blade, and heat dissipation nets are provided on the inner sides of both sides of the partition plate.

9. The speech recognition combined with database enhanced large model government affairs intelligent question and answer device according to claim 4 is characterized in that: Sliding brackets are fixedly provided on the front sides of both sides of the inner wall of the portable box, and a pull-out plate is vertically slidably connected to the inner sides of the two sliding brackets. An observation window is provided on the inner side of the pull-out plate, and a through-hole plate is provided on the surface of the pull-out plate and at the top of the observation window. A through groove is provided on the inner side of the top of the pull-out plate, and two supporting legs are respectively fixedly provided on both sides of the bottom of the portable box.

Images